Ubuntu Manpage: stress-ng - a tool to load and stress a computer system

Provided by: stress-ng_0.13.12-2ubuntu1_amd64

NAME

       stress-ng - a tool to load and stress a computer system

SYNOPSIS

       stress-ng [OPTION [ARG]] ...

DESCRIPTION

stress-ng will stress test a computer system in various selectable ways. It was designed to exercise
various physical subsystems of a computer as well as the various operating system kernel interfaces.
stress-ng also has a wide range of CPU specific stress tests that exercise floating point, integer, bit
manipulation and control flow.

stress-ng was originally intended to make a machine work hard and trip hardware issues such as thermal
overruns as well as operating system bugs that only occur when a system is being thrashed hard. Use
stress-ng with caution as some of the tests can make a system run hot on poorly designed hardware and
also can cause excessive system thrashing which may be difficult to stop.

stress-ng can also measure test throughput rates; this can be useful to observe performance changes
across different operating system releases or types of hardware. However, it has never been intended to
be used as a precise benchmark test suite, so do NOT use it in this manner.

Running stress-ng with root privileges will adjust out of memory settings on Linux systems to make the
stressors unkillable in low memory situations, so use this judiciously. With the appropriate privilege,
stress-ng can allow the ionice class and ionice levels to be adjusted, again, this should be used with
care.

One can specify the number of processes to invoke per type of stress test; specifying a zero value will
select the number of processors available as defined by sysconf(_SC_NPROCESSORS_CONF), if that can't be
determined then the number of online CPUs is used. If the value is less than zero then the number of
online CPUs is used.

OPTIONS

General stress-ng control options:

--abort
this option will force all running stressors to abort (terminate) if any other stressor terminates
prematurely because of a failure.

--aggressive
enables more file, cache and memory aggressive options. This may slow tests down, increase
latencies and reduce the number of bogo ops as well as changing the balance of user time vs system
time used depending on the type of stressor being used.

-a N, --all N, --parallel N
start N instances of all stressors in parallel. If N is less than zero, then the number of CPUs
online is used for the number of instances. If N is zero, then the number of configured CPUs in
the system is used.

-b N, --backoff N
wait N microseconds between the start of each stress worker process. This allows one to ramp up
the stress tests over time.

--class name
specify the class of stressors to run. Stressors are classified into one or more of the following
classes: cpu, cpu-cache, device, io, interrupt, filesystem, memory, network, os, pipe, scheduler
and vm. Some stressors fall into just one class. For example the 'get' stressor is just in the
'os' class. Other stressors fall into more than one class, for example, the 'lsearch' stressor
falls into the 'cpu', 'cpu-cache' and 'memory' classes as it exercises all these three. Selecting
a specific class will run all the stressors that fall into that class only when run with the
--sequential option.

Specifying a name followed by a question mark (for example --class vm?) will print out all the
stressors in that specific class.

-n, --dry-run
parse options, but do not run stress tests. A no-op.

--ftrace
enable kernel function call tracing (Linux only). This will use the kernel debugfs ftrace
mechanism to record all the kernel functions used on the system while stress-ng is running. This
is only as accurate as the kernel ftrace output, so there may be some variability on the data
reported.

-h, --help
show help.

--ignite-cpu
alter kernel controls to try and maximize the CPU. This requires root privilege to alter various
/sys interface controls. Currently this only works for Intel P-State enabled x86 systems on
Linux.

--ionice-class class
specify ionice class (only on Linux). Can be idle (default), besteffort, be, realtime, rt.

--ionice-level level
specify ionice level (only on Linux). For idle, 0 is the only possible option. For besteffort or
realtime values 0 (highest priority) to 7 (lowest priority). See ionice(1) for more details.

--iostat S
every S seconds show I/O statistics on the device that stores the stress-ng temporary files. This
is either the device of the current working directory or the --temp-path specified path. Currently
a Linux only option. The fields output are:

Column Heading Explanation
Inflight number of I/O requests that have been issued to the
device driver but have not yet completed
Rd K/s read rate in 1024 bytes per second
Wr K/s write rate in 1024 bytes per second
Dscd K/s discard rate in 1024 bytes per second
Rd/s reads per second
Wr/s writes per second
Dscd/s discards per second

--job jobfile
run stressors using a jobfile. The jobfile is essentially a file containing stress-ng options
(without the leading --) with one option per line. Lines may have comments with comment text
proceeded by the # character. A simple example is as follows:

run sequential # run stressors sequentially
verbose # verbose output
metrics-brief # show metrics at end of run
timeout 60s # stop each stressor after 60 seconds
#
# vm stressor options:
#
vm 2 # 2 vm stressors
vm-bytes 128M # 128MB available memory
vm-keep # keep vm mapping
vm-populate # populate memory
#
# memcpy stressor options:
#
memcpy 5 # 5 memcpy stressors

The job file introduces the run command that specifies how to run the stressors:

run sequential - run stressors sequentially
run parallel - run stressors together in parallel

Note that 'run parallel' is the default.

--keep-files
do not remove files and directories created by the stressors. This can be useful for debugging
purposes. Not generally recommended as it can fill up a file system.

-k, --keep-name
by default, stress-ng will attempt to change the name of the stress processes according to their
functionality; this option disables this and keeps the process names to be the name of the parent
process, that is, stress-ng.

--log-brief
by default stress-ng will report the name of the program, the message type and the process id as a
prefix to all output. The --log-brief option will output messages without these fields to produce
a less verbose output.

--log-file filename
write messages to the specified log file.

--maximize
overrides the default stressor settings and instead sets these to the maximum settings allowed.
These defaults can always be overridden by the per stressor settings options if required.

--max-fd N
set the maximum limit on file descriptors (value or a % of system allowed maximum). By default,
stress-ng can use all the available file descriptors; this option sets the limit in the range from
10 up to the maximum limit of RLIMIT_NOFILE. One can use a % setting too, e.g. 50% is half the
maximum allowed file descriptors. Note that stress-ng will use about 5 of the available file
descriptors so take this into consideration when using this setting.

--metrics
output number of bogo operations in total performed by the stress processes. Note that these are
not a reliable metric of performance or throughput and have not been designed to be used for
benchmarking whatsoever. The metrics are just a useful way to observe how a system behaves when
under various kinds of load.

The following columns of information are output:

Column Heading Explanation
bogo ops number of iterations of the stressor during the run.
This is metric of how much overall "work" has been
achieved in bogo operations.
real time (secs) average wall clock duration (in seconds) of the
stressor. This is the total wall clock time of all the
instances of that particular stressor divided by the
number of these stressors being run.
usr time (secs) total user time (in seconds) consumed running all the
instances of the stressor.
sys time (secs) total system time (in seconds) consumed running all the
instances of the stressor.
bogo ops/s (real time) total bogo operations per second based on wall clock run
time. The wall clock time reflects the apparent run
time. The more processors one has on a system the more
the work load can be distributed onto these and hence
the wall clock time will reduce and the bogo ops rate
will increase. This is essentially the "apparent" bogo
ops rate of the system.

bogo ops/s (usr+sys time) total bogo operations per second based on cumulative
user and system time. This is the real bogo ops rate of
the system taking into consideration the actual time
execution time of the stressor across all the
processors. Generally this will decrease as one adds
more concurrent stressors due to contention on cache,
memory, execution units, buses and I/O devices.
CPU used per instance (%) total percentage of CPU used divided by number of
stressor instances. 100% is 1 full CPU. Some stressors
run multiple threads so it is possible to have a figure
greater than 100%.

--metrics-brief
show shorter list of stressor metrics (no CPU used per instance).

--minimize
overrides the default stressor settings and instead sets these to the minimum settings allowed.
These defaults can always be overridden by the per stressor settings options if required.

--no-madvise
from version 0.02.26 stress-ng automatically calls madvise(2) with random advise options before
each mmap and munmap to stress the vm subsystem a little harder. The --no-advise option turns this
default off.

--no-oom-adjust
disable any form of out-of-memory score adjustments, keep the system defaults. Normally stress-ng
will adjust the out-of-memory scores on stressors to try to create more memory pressure. This
option disables the adjustments.

--no-rand-seed
Do not seed the stress-ng pseudo-random number generator with a quasi random start seed, but
instead seed it with constant values. This forces tests to run each time using the same start
conditions which can be useful when one requires reproducible stress tests.

--oomable
Do not respawn a stressor if it gets killed by the Out-of-Memory (OOM) killer. The default
behaviour is to restart a new instance of a stressor if the kernel OOM killer terminates the
process. This option disables this default behaviour.

--page-in
touch allocated pages that are not in core, forcing them to be paged back in. This is a useful
option to force all the allocated pages to be paged in when using the bigheap, mmap and vm
stressors. It will severely degrade performance when the memory in the system is less than the
allocated buffer sizes. This uses mincore(2) to determine the pages that are not in core and
hence need touching to page them back in.

--pathological
enable stressors that are known to hang systems. Some stressors can quickly consume resources in
such a way that they can rapidly hang a system before the kernel can OOM kill them. These
stressors are not enabled by default, this option enables them, but you probably don't want to do
this. You have been warned.

--perf measure processor and system activity using perf events. Linux only and caveat emptor, according
to perf_event_open(2): "Always double-check your results! Various generalized events have had
wrong values.". Note that with Linux 4.7 one needs to have CAP_SYS_ADMIN capabilities for this
option to work, or adjust /proc/sys/kernel/perf_event_paranoid to below 2 to use this without
CAP_SYS_ADMIN.

-q, --quiet
do not show any output.

-r N, --random N
start N random stress workers. If N is 0, then the number of configured processors is used for N.

--sched scheduler
select the named scheduler (only on Linux). To see the list of available schedulers use: stress-ng
--sched which

--sched-prio prio
select the scheduler priority level (only on Linux). If the scheduler does not support this then
the default priority level of 0 is chosen.

--sched-period period
select the period parameter for deadline scheduler (only on Linux). Default value is 0 (in
nanoseconds).

--sched-runtime runtime
select the runtime parameter for deadline scheduler (only on Linux). Default value is 99999 (in
nanoseconds).

--sched-deadline deadline
select the deadline parameter for deadline scheduler (only on Linux). Default value is 100000 (in
nanoseconds).

--sched-reclaim
use cpu bandwidth reclaim feature for deadline scheduler (only on Linux).

--seed N
set the random number generate seed with a 64 bit value. Allows stressors to use the same random
number generator sequences on each invocation.

--sequential N
sequentially run all the stressors one by one for a default of 60 seconds. The number of instances
of each of the individual stressors to be started is N. If N is less than zero, then the number
of CPUs online is used for the number of instances. If N is zero, then the number of CPUs in the
system is used. Use the --timeout option to specify the duration to run each stressor.

--skip-silent
silence messages that report that a stressor has been skipped because it requires features not
supported by the system, such as unimplemented system calls, missing resources or processor
specific features.

--smart
scan the block devices for changes S.M.A.R.T. statistics (Linux only). This requires root
privileges to read the Self-Monitoring, Analysis and Reporting Technology data from all block
devies and will report any changes in the statistics. One caveat is that device manufacturers
provide different sets of data, the exact meaning of the data can be vague and the data may be
inaccurate.

--stdout
all output goes to stdout. By default all output goes to stderr (which is a historical oversight
that will cause breakage to users if it is now changed). This option allows the output to be
written to stdout.

--stressors
output the names of the available stressors.

--syslog
log output (except for verbose -v messages) to the syslog.

--taskset list
set CPU affinity based on the list of CPUs provided; stress-ng is bound to just use these CPUs
(Linux only). The CPUs to be used are specified by a comma separated list of CPU (0 to N-1). One
can specify a range of CPUs using '-', for example: --taskset 0,2-3,6,7-11

--temp-path path
specify a path for stress-ng temporary directories and temporary files; the default path is the
current working directory. This path must have read and write access for the stress-ng stress
processes.

--thermalstat S
every S seconds show CPU and thermal load statistics. This option shows average CPU frequency in
GHz (average of online-CPUs), load averages (1 minute, 5 minute and 15 minutes) and available
thermal zone temperatures in degrees Centigrade.

--thrash
This can only be used when running on Linux and with root privilege. This option starts a
background thrasher process that works through all the processes on a system and tries to page as
many pages in the processes as possible. It also periodically drops the page cache, frees
reclaimable slab objects and pagecache. This will cause considerable amount of thrashing of swap
on an over-committed system.

-t N, --timeout T
run each stress test for at least T seconds. One can also specify the units of time in seconds,
minutes, hours, days or years with the suffix s, m, h, d or y. Each stressor will be sent a
SIGALRM signal at the timeout time, however if the stress test is swapped out, in a non-
interritable system call or performing clean up (such as removing hundreds of test file) it may
take a while to finally terminate. A 0 timeout will run stress-ng for ever with no timeout.

--timestamp
add a timestamp in hours, minutes, seconds and hundredths of a second to the log output.

--timer-slack N
adjust the per process timer slack to N nanoseconds (Linux only). Increasing the timer slack
allows the kernel to coalesce timer events by adding some fuzziness to timer expiration times and
hence reduce wakeups. Conversely, decreasing the timer slack will increase wakeups. A value of 0
for the timer-slack will set the system default of 50,000 nanoseconds.

--times
show the cumulative user and system times of all the child processes at the end of the stress run.
The percentage of utilisation of available CPU time is also calculated from the number of on-line
CPUs in the system.

--tz collect temperatures from the available thermal zones on the machine (Linux only). Some devices
may have one or more thermal zones, where as others may have none.

-v, --verbose
show all debug, warnings and normal information output.

--verify
verify results when a test is run. This is not available on all tests. This will sanity check the
computations or memory contents from a test run and report to stderr any unexpected failures.

--verifiable
print the names of stressors that can be verified with the --verify option.

-V, --version
show version of stress-ng, version of toolchain used to build stress-ng and system information.

--vmstat S
every S seconds show statistics about processes, memory, paging, block I/O, interrupts, context
switches, disks and cpu activity. The output is similar that to the output from the vmstat(8)
utility. Currently a Linux only option.

-x, --exclude list
specify a list of one or more stressors to exclude (that is, do not run them). This is useful to
exclude specific stressors when one selects many stressors to run using the --class option,
--sequential, --all and --random options. Example, run the cpu class stressors concurrently and
exclude the numa and search stressors:

stress-ng --class cpu --all 1 -x numa,bsearch,hsearch,lsearch

-Y, --yaml filename
output gathered statistics to a YAML formatted file named 'filename'.

Stressor specific options:

--access N
start N workers that work through various settings of file mode bits (read, write, execute) for
the file owner and checks if the user permissions of the file using access(2) and faccessat(2) are
sane.

--access-ops N
stop access workers after N bogo access sanity checks.

--affinity N
start N workers that run 16 processes that rapidly change CPU affinity (only on Linux). Rapidly
switching CPU affinity can contribute to poor cache behaviour and high context switch rate.

--affinity-ops N
stop affinity workers after N bogo affinity operations. Note that the counters across the 16
processes are not locked to improve affinity test rates so the final number of bogo-ops will be
equal or more than the specified ops stop threshold because of racy unlocked bogo-op counting.

--affinity-delay N
delay for N nanoseconds before changing affinity to the next CPU. The delay will spin on CPU
scheduling yield operations for N nanoseconds before the process is moved to another CPU. The
default is 0 nanosconds.

--affinity-pin
pin all the 16 per stressor processes to a CPU. All 16 processes follow the CPU chosen by the main
parent stressor, forcing heavy per CPU loading.

--affinity-rand
switch CPU affinity randomly rather than the default of sequentially.

--affinity-sleep N
sleep for N nanoseconds before changing affinity to the next CPU.

--af-alg N
start N workers that exercise the AF_ALG socket domain by hashing and encrypting various sized
random messages. This exercises the available hashes, ciphers, rng and aead crypto engines in the
Linux kernel.

--af-alg-ops N
stop af-alg workers after N AF_ALG messages are hashed.

--af-alg-dump
dump the internal list representing cryptographic algorithms parsed from the /proc/crypto file to
standard output (stdout).

--aio N
start N workers that issue multiple small asynchronous I/O writes and reads on a relatively small
temporary file using the POSIX aio interface. This will just hit the file system cache and soak
up a lot of user and kernel time in issuing and handling I/O requests. By default, each worker
process will handle 16 concurrent I/O requests.

--aio-ops N
stop POSIX asynchronous I/O workers after N bogo asynchronous I/O requests.

--aio-requests N
specify the number of POSIX asynchronous I/O requests each worker should issue, the default is 16;
1 to 4096 are allowed.

--aiol N
start N workers that issue multiple 4K random asynchronous I/O writes using the Linux aio system
calls io_setup(2), io_submit(2), io_getevents(2) and io_destroy(2). By default, each worker
process will handle 16 concurrent I/O requests.

--aiol-ops N
stop Linux asynchronous I/O workers after N bogo asynchronous I/O requests.

--aiol-requests N
specify the number of Linux asynchronous I/O requests each worker should issue, the default is 16;
1 to 4096 are allowed.

--alarm N
start N workers that exercise alarm(2) with MAXINT, 0 and random alarm and sleep delays that get
prematurely interrupted. Before each alarm is scheduled any previous pending alarms are cancelled
with zero second alarm calls.

--alarm-ops N
stop after N alarm bogo operations.

--apparmor N
start N workers that exercise various parts of the AppArmor interface. Currently one needs root
permission to run this particular test. Only available on Linux systems with AppArmor support and
requires the CAP_MAC_ADMIN capability.

--apparmor-ops
stop the AppArmor workers after N bogo operations.

--atomic N
start N workers that exercise various GCC __atomic_*() built in operations on 8, 16, 32 and 64 bit
integers that are shared among the N workers. This stressor is only available for builds using GCC
4.7.4 or higher. The stressor forces many front end cache stalls and cache references.

--atomic-ops N
stop the atomic workers after N bogo atomic operations.

--bad-altstack N
start N workers that create broken alternative signal stacks for SIGSEGV and SIGBUS handling that
in turn create secondary SIGSEGV/SIGBUS errors. A variety of randonly selected nefarious methods
are used to create the stacks:

• Unmapping the alternative signal stack, before triggering the signal handling.
• Changing the alternative signal stack to just being read only, write only, execute only.
• Using a NULL alternative signal stack.
• Using the signal handler object as the alternative signal stack.
• Unmapping the alternative signal stack during execution of the signal handler.
• Using a read-only text segment for the alternative signal stack.
• Using an undersized alternative signal stack.
• Using the VDSO as an alternative signal stack.
• Using an alternative stack mapped onto /dev/zero.
• Using an alternative stack mapped to a zero sized temporary file to generate a SIGBUS error.

--bad-altstack-ops N
stop the bad alternative stack stressors after N SIGSEGV bogo operations.

--bad-ioctl N
start N workers that perform a range of illegal bad read ioctls (using _IOR) across the device
drivers. This exercises page size, 64 bit, 32 bit, 16 bit and 8 bit reads as well as NULL
addresses, non-readable pages and PROT_NONE mapped pages. Currently only for Linux and requires
the --pathological option.

--bad-ioctl-ops N
stop the bad ioctl stressors after N bogo ioctl operations.

-B N, --bigheap N
start N workers that grow their heaps by reallocating memory. If the out of memory killer (OOM) on
Linux kills the worker or the allocation fails then the allocating process starts all over again.
Note that the OOM adjustment for the worker is set so that the OOM killer will treat these workers
as the first candidate processes to kill.

--bigheap-ops N
stop the big heap workers after N bogo allocation operations are completed.

--bigheap-growth N
specify amount of memory to grow heap by per iteration. Size can be from 4K to 64MB. Default is
64K.

--binderfs N
start N workers that mount, exercise and unmount binderfs. The binder control device is exercised
with 256 sequential BINDER_CTL_ADD ioctl calls per loop.

--binderfs-ops N
stop after N binderfs cycles.

--bind-mount N
start N workers that repeatedly bind mount / to / inside a user namespace. This can consume
resources rapidly, forcing out of memory situations. Do not use this stressor unless you want to
risk hanging your machine.

--bind-mount-ops N
stop after N bind mount bogo operations.

--branch N
start N workers that randomly branch to 1024 randomly selected locations and hence exercise the
CPU branch prediction logic.

--branch-ops N
stop the branch stressors after N × 1024 branches

--brk N
start N workers that grow the data segment by one page at a time using multiple brk(2) calls. Each
successfully allocated new page is touched to ensure it is resident in memory. If an out of
memory condition occurs then the test will reset the data segment to the point before it started
and repeat the data segment resizing over again. The process adjusts the out of memory setting so
that it may be killed by the out of memory (OOM) killer before other processes. If it is killed
by the OOM killer then it will be automatically re-started by a monitoring parent process.

--brk-ops N
stop the brk workers after N bogo brk operations.

--brk-mlock
attempt to mlock future brk pages into memory causing more memory pressure. If mlock(MCL_FUTURE)
is implemented then this will stop new brk pages from being swapped out.

--brk-notouch
do not touch each newly allocated data segment page. This disables the default of touching each
newly allocated page and hence avoids the kernel from necessarily backing the page with real
physical memory.

--bsearch N
start N workers that binary search a sorted array of 32 bit integers using bsearch(3). By default,
there are 65536 elements in the array. This is a useful method to exercise random access of
memory and processor cache.

--bsearch-ops N
stop the bsearch worker after N bogo bsearch operations are completed.

--bsearch-size N
specify the size (number of 32 bit integers) in the array to bsearch. Size can be from 1K to 4M.

-C N, --cache N
start N workers that perform random wide spread memory read and writes to thrash the CPU cache.
The code does not intelligently determine the CPU cache configuration and so it may be sub-optimal
in producing hit-miss read/write activity for some processors.

--cache-cldemote
cache line demote (x86 only). This is a no-op for non-x86 architectures and older x86 processors
that do not support this feature.

--cache-clflushopt
use optimized cache line flush (x86 only). This is a no-op for non-x86 architectures and older x86
processors that do not support this feature.

--cache-fence
force write serialization on each store operation (x86 only). This is a no-op for non-x86
architectures.

--cache-flush
force flush cache on each store operation (x86 only). This is a no-op for non-x86 architectures.

--cache-level N
specify level of cache to exercise (1=L1 cache, 2=L2 cache, 3=L3/LLC cache (the default)). If the
cache hierarchy cannot be determined, built-in defaults will apply.

--cache-no-affinity
do not change processor affinity when --cache is in effect.

--cache-sfence
force write serialization on each store operation using the sfence instruction (x86 only). This is
a no-op for non-x86 architectures.

--cache-ops N
stop cache thrash workers after N bogo cache thrash operations.

--cache-prefetch
force read prefetch on next read address on architectures that support prefetching.

--cache-ways N
specify the number of cache ways to exercise. This allows a subset of the overall cache size to be
exercised.

--cap N
start N workers that read per process capabilities via calls to capget(2) (Linux only).

--cap-ops N
stop after N cap bogo operations.

--chattr N
start N workers that attempt to exercise file attributes via the EXT2_IOC_SETFLAGS ioctl. This is
intended to be intentionally racy and exercise a range of chattr attributes by enabling and
disabling them on a file shared amongst the N chattr stressor processes. (Linux only).

--chattr-ops N
stop after N chattr bogo operations.

--chdir N
start N workers that change directory between directories using chdir(2).

--chdir-ops N
stop after N chdir bogo operations.

--chdir-dirs N
exercise chdir on N directories. The default is 8192 directories, this allows 64 to 65536
directories to be used instead.

--chmod N
start N workers that change the file mode bits via chmod(2) and fchmod(2) on the same file. The
greater the value for N then the more contention on the single file. The stressor will work
through all the combination of mode bits.

--chmod-ops N
stop after N chmod bogo operations.

--chown N
start N workers that exercise chown(2) on the same file. The greater the value for N then the more
contention on the single file.

--chown-ops N
stop the chown workers after N bogo chown(2) operations.

--chroot N
start N workers that exercise chroot(2) on various valid and invalid chroot paths. Only available
on Linux systems and requires the CAP_SYS_ADMIN capability.

--chroot-ops N
stop the chroot workers after N bogo chroot(2) operations.

--clock N
start N workers exercising clocks and POSIX timers. For all known clock types this will exercise
clock_getres(2), clock_gettime(2) and clock_nanosleep(2). For all known timers it will create a
50000ns timer and busy poll this until it expires. This stressor will cause frequent context
switching.

--clock-ops N
stop clock stress workers after N bogo operations.

--clone N
start N workers that create clones (via the clone(2) and clone3() system calls). This will
rapidly try to create a default of 8192 clones that immediately die and wait in a zombie state
until they are reaped. Once the maximum number of clones is reached (or clone fails because one
has reached the maximum allowed) the oldest clone thread is reaped and a new clone is then created
in a first-in first-out manner, and then repeated. A random clone flag is selected for each clone
to try to exercise different clone operations. The clone stressor is a Linux only option.

--clone-ops N
stop clone stress workers after N bogo clone operations.

--clone-max N
try to create as many as N clone threads. This may not be reached if the system limit is less than
N.

--close N
start N workers that try to force race conditions on closing opened file descriptors. These file
descriptors have been opened in various ways to try and exercise different kernel close handlers.

--close-ops N
stop close workers after N bogo close operations.

--context N
start N workers that run three threads that use swapcontext(3) to implement the thread-to-thread
context switching. This exercises rapid process context saving and restoring and is bandwidth
limited by register and memory save and restore rates.

--context-ops N
stop context workers after N bogo context switches. In this stressor, 1 bogo op is equivalent to
1000 swapcontext calls.

--copy-file N
start N stressors that copy a file using the Linux copy_file_range(2) system call. 2MB chunks of
data are copied from random locations from one file to random locations to a destination file. By
default, the files are 256 MB in size. Data is sync'd to the filesystem after each
copy_file_range(2) call.

--copy-file-ops N
stop after N copy_file_range() calls.

--copy-file-bytes N
copy file size, the default is 256 MB. One can specify the size as % of free space on the file
system or in units of Bytes, KBytes, MBytes and GBytes using the suffix b, k, m or g.

-c N, --cpu N
start N workers exercising the CPU by sequentially working through all the different CPU stress
methods. Instead of exercising all the CPU stress methods, one can specify a specific CPU stress
method with the --cpu-method option.

--cpu-ops N
stop cpu stress workers after N bogo operations.

-l P, --cpu-load P
load CPU with P percent loading for the CPU stress workers. 0 is effectively a sleep (no load) and
100 is full loading. The loading loop is broken into compute time (load%) and sleep time (100% -
load%). Accuracy depends on the overall load of the processor and the responsiveness of the
scheduler, so the actual load may be different from the desired load. Note that the number of
bogo CPU operations may not be linearly scaled with the load as some systems employ CPU frequency
scaling and so heavier loads produce an increased CPU frequency and greater CPU bogo operations.

Note: This option only applies to the --cpu stressor option and not to all of the cpu class of
stressors.

--cpu-load-slice S
note - this option is only useful when --cpu-load is less than 100%. The CPU load is broken into
multiple busy and idle cycles. Use this option to specify the duration of a busy time slice. A
negative value for S specifies the number of iterations to run before idling the CPU (e.g. -30
invokes 30 iterations of a CPU stress loop). A zero value selects a random busy time between 0
and 0.5 seconds. A positive value for S specifies the number of milliseconds to run before idling
the CPU (e.g. 100 keeps the CPU busy for 0.1 seconds). Specifying small values for S lends to
small time slices and smoother scheduling. Setting --cpu-load as a relatively low value and
--cpu-load-slice to be large will cycle the CPU between long idle and busy cycles and exercise
different CPU frequencies. The thermal range of the CPU is also cycled, so this is a good
mechanism to exercise the scheduler, frequency scaling and passive/active thermal cooling
mechanisms.

Note: This option only applies to the --cpu stressor option and not to all of the cpu class of
stressors.

--cpu-method method
specify a cpu stress method. By default, all the stress methods are exercised sequentially,
however one can specify just one method to be used if required. Available cpu stress methods are
described as follows:

Method Description
all iterate over all the below cpu stress methods
ackermann Ackermann function: compute A(3, 7), where:
A(m, n) = n + 1 if m = 0;
A(m - 1, 1) if m > 0 and n = 0;
A(m - 1, A(m, n - 1)) if m > 0 and n > 0
apery calculate Apery's constant ζ(3); the sum of 1/(n ↑ 3) to
a precision of 1.0x10↑14
bitops various bit operations from bithack, namely: reverse
bits, parity check, bit count, round to nearest power of
2
callfunc recursively call 8 argument C function to a depth of
1024 calls and unwind
cfloat 1000 iterations of a mix of floating point complex
operations
cdouble 1000 iterations of a mix of double floating point
complex operations
clongdouble 1000 iterations of a mix of long double floating point
complex operations
collatz compute the 1348 steps in the collatz sequence starting
from number 989345275647. Where f(n) = n / 2 (for even
n) and f(n) = 3n + 1 (for odd n).
correlate perform a 8192 × 512 correlation of random doubles
cpuid fetch cpu specific information using the cpuid
instruction (x86 only)
crc16 compute 1024 rounds of CCITT CRC16 on random data
decimal32 1000 iterations of a mix of 32 bit decimal floating
point operations (GCC only)
decimal64 1000 iterations of a mix of 64 bit decimal floating
point operations (GCC only)
decimal128 1000 iterations of a mix of 128 bit decimal floating
point operations (GCC only)
dither Floyd–Steinberg dithering of a 1024 × 768 random image
from 8 bits down to 1 bit of depth
div16 50,000 16 bit unsigned integer divisions
div32 50,000 32 bit unsigned integer divisions
div64 50,000 64 bit unsigned integer divisions
double 1000 iterations of a mix of double precision floating
point operations
euler compute e using n = (1 + (1 ÷ n)) ↑ n
explog iterate on n = exp(log(n) ÷ 1.00002)
factorial find factorials from 1..150 using Stirling's and
Ramanujan's approximations
fibonacci compute Fibonacci sequence of 0, 1, 1, 2, 5, 8...
fft 4096 sample Fast Fourier Transform
fletcher16 1024 rounds of a naive implementation of a 16 bit
Fletcher's checksum
float 1000 iterations of a mix of floating point operations
float16 1000 iterations of a mix of 16 bit floating point
operations
float32 1000 iterations of a mix of 32 bit floating point
operations
float64 1000 iterations of a mix of 64 bit floating point
operations
float80 1000 iterations of a mix of 80 bit floating point
operations
float128 1000 iterations of a mix of 128 bit floating point
operations

floatconversion perform 65536 iterations of floating point conversions
between float, double and long double floating point
variables.
gamma calculate the Euler-Mascheroni constant γ using the
limiting difference between the harmonic series (1 + 1/2
+ 1/3 + 1/4 + 1/5 ... + 1/n) and the natural logarithm
ln(n), for n = 80000.
gcd compute GCD of integers
gray calculate binary to gray code and gray code back to
binary for integers from 0 to 65535
hamming compute Hamming H(8,4) codes on 262144 lots of 4 bit
data. This turns 4 bit data into 8 bit Hamming code
containing 4 parity bits. For data bits d1..d4, parity
bits are computed as:
p1 = d2 + d3 + d4
p2 = d1 + d3 + d4
p3 = d1 + d2 + d4
p4 = d1 + d2 + d3
hanoi solve a 21 disc Towers of Hanoi stack using the
recursive solution
hyperbolic compute sinh(θ) × cosh(θ) + sinh(2θ) + cosh(3θ) for
float, double and long double hyperbolic sine and cosine
functions where θ = 0 to 2π in 1500 steps
idct 8 × 8 IDCT (Inverse Discrete Cosine Transform).
int8 1000 iterations of a mix of 8 bit integer operations.
int16 1000 iterations of a mix of 16 bit integer operations.
int32 1000 iterations of a mix of 32 bit integer operations.
int64 1000 iterations of a mix of 64 bit integer operations.
int128 1000 iterations of a mix of 128 bit integer operations
(GCC only).
int32float 1000 iterations of a mix of 32 bit integer and floating
point operations.
int32double 1000 iterations of a mix of 32 bit integer and double
precision floating point operations.
int32longdouble 1000 iterations of a mix of 32 bit integer and long
double precision floating point operations.
int64float 1000 iterations of a mix of 64 bit integer and floating
point operations.
int64double 1000 iterations of a mix of 64 bit integer and double
precision floating point operations.
int64longdouble 1000 iterations of a mix of 64 bit integer and long
double precision floating point operations.
int128float 1000 iterations of a mix of 128 bit integer and floating
point operations (GCC only).
int128double 1000 iterations of a mix of 128 bit integer and double
precision floating point operations (GCC only).
int128longdouble 1000 iterations of a mix of 128 bit integer and long
double precision floating point operations (GCC only).
int128decimal32 1000 iterations of a mix of 128 bit integer and 32 bit
decimal floating point operations (GCC only).
int128decimal64 1000 iterations of a mix of 128 bit integer and 64 bit
decimal floating point operations (GCC only).
int128decimal128 1000 iterations of a mix of 128 bit integer and 128 bit
decimal floating point operations (GCC only).
intconversion perform 65536 iterations of integer conversions between
int16, int32 and int64 variables.
ipv4checksum compute 1024 rounds of the 16 bit ones' complement IPv4
checksum.
jmp Simple unoptimised compare >, <, == and jmp branching.
lfsr32 16384 iterations of a 32 bit Galois linear feedback
shift register using the polynomial x↑32 + x↑31 + x↑29 +
x + 1. This generates a ring of 2↑32 - 1 unique values
(all 32 bit values except for 0).
ln2 compute ln(2) based on series:
1 - 1/2 + 1/3 - 1/4 + 1/5 - 1/6 ...
logmap 16384 iterations computing chaotic double precision
values using the logistic map Χn+1 = r × Χn × (1 - Χn)
where r > ≈ 3.56994567
longdouble 1000 iterations of a mix of long double precision
floating point operations.
loop simple empty loop.

matrixprod matrix product of two 128 × 128 matrices of double
floats. Testing on 64 bit x86 hardware shows that this
is provides a good mix of memory, cache and floating
point operations and is probably the best CPU method to
use to make a CPU run hot.
nsqrt compute sqrt() of long doubles using Newton-Raphson.
omega compute the omega constant defined by Ωe↑Ω = 1 using
efficient iteration of Ωn+1 = (1 + Ωn) / (1 + e↑Ωn).
parity compute parity using various methods from the Standford
Bit Twiddling Hacks. Methods employed are: the naïve
way, the naïve way with the Brian Kernigan bit counting
optimisation, the multiply way, the parallel way, the
lookup table ways (2 variations) and using the
__builtin_parity function.
phi compute the Golden Ratio ϕ using series.
pi compute π using the Srinivasa Ramanujan fast convergence
algorithm.
prime find the first 10000 prime numbers using a slightly
optimised brute force naïve trial division search.
psi compute ψ (the reciprocal Fibonacci constant) using the
sum of the reciprocals of the Fibonacci numbers.
queens compute all the solutions of the classic 8 queens
problem for board sizes 1..11.
rand 16384 iterations of rand(), where rand is the MWC pseudo
random number generator. The MWC random function
concatenates two 16 bit multiply-with-carry generators:
x(n) = 36969 × x(n - 1) + carry,
y(n) = 18000 × y(n - 1) + carry mod 2 ↑ 16

and has period of around 2 ↑ 60.
rand48 16384 iterations of drand48(3) and lrand48(3).
rgb convert RGB to YUV and back to RGB (CCIR 601).
sieve find the first 10000 prime numbers using the sieve of
Eratosthenes.
stats calculate minimum, maximum, arithmetic mean, geometric
mean, harmoninc mean and standard deviation on 250
randomly generated positive double precision values.
sqrt compute sqrt(rand()), where rand is the MWC pseudo
random number generator.
trig compute sin(θ) × cos(θ) + sin(2θ) + cos(3θ) for float,
double and long double sine and cosine functions where θ
= 0 to 2π in 1500 steps.
union perform integer arithmetic on a mix of bit fields in a C
union. This exercises how well the compiler and CPU can
perform integer bit field loads and stores.
zeta compute the Riemann Zeta function ζ(s) for s = 2.0..10.0

Note that some of these methods try to exercise the CPU with computations found in some real world
use cases. However, the code has not been optimised on a per-architecture basis, so may be a sub-
optimal compared to hand-optimised code used in some applications. They do try to represent the
typical instruction mixes found in these use cases.

--cpu-online N
start N workers that put randomly selected CPUs offline and online. This Linux only stressor
requires root privilege to perform this action. By default the first CPU (CPU 0) is never offlined
as this has been found to be problematic on some systems and can result in a shutdown.

--cpu-online-all
The default is to never offline the first CPU. This option will offline and online all the CPUs
include CPU 0. This may cause some systems to shutdown.

--cpu-online-ops N
stop after offline/online operations.

--crypt N
start N workers that encrypt a 16 character random password using crypt(3). The password is
encrypted using MD5, SHA-256 and SHA-512 encryption methods.

--crypt-ops N
stop after N bogo encryption operations.

--cyclic N
start N workers that exercise the real time FIFO or Round Robin schedulers with cyclic nanosecond
sleeps. Normally one would just use 1 worker instance with this stressor to get reliable
statistics. This stressor measures the first 10 thousand latencies and calculates the mean, mode,
minimum, maximum latencies along with various latency percentiles for the just the first cyclic
stressor instance. One has to run this stressor with CAP_SYS_NICE capability to enable the real
time scheduling policies. The FIFO scheduling policy is the default.

--cyclic-ops N
stop after N sleeps.

--cyclic-dist N
calculate and print a latency distribution with the interval of N nanoseconds. This is helpful to
see where the latencies are clustering.

Method Description
clock_ns sleep for the specified time using the
clock_nanosleep(2) high resolution nanosleep and the
CLOCK_REALTIME real time clock.
itimer wakeup a paused process with a CLOCK_REALTIME itimer
signal.
poll delay for the specified time using a poll delay loop
that checks for time changes using clock_gettime(2) on
the CLOCK_REALTIME clock.
posix_ns sleep for the specified time using the POSIX
nanosleep(2) high resolution nanosleep.
pselect sleep for the specified time using pselect(2) with null
file descriptors.
usleep sleep to the nearest microsecond using usleep(2).

--cyclic-policy [ fifo | rr ]
specify the desired real time scheduling policy, ff (first-in, first-out) or rr (round robin).

--cyclic-prio P
specify the scheduling priority P. Range from 1 (lowest) to 100 (highest).

--cyclic-sleep N
sleep for N nanoseconds per test cycle using clock_nanosleep(2) with the CLOCK_REALTIME timer.
Range from 1 to 1000000000 nanoseconds.

--daemon N
start N workers that each create a daemon that dies immediately after creating another daemon and
so on. This effectively works through the process table with short lived processes that do not
have a parent and are waited for by init. This puts pressure on init to do rapid child reaping.
The daemon processes perform the usual mix of calls to turn into typical UNIX daemons, so this
artificially mimics very heavy daemon system stress.

--daemon-ops N
stop daemon workers after N daemons have been created.

--dccp N
start N workers that send and receive data using the Datagram Congestion Control Protocol (DCCP)
(RFC4340). This involves a pair of client/server processes performing rapid connect, send and
receives and disconnects on the local host.

--dccp-domain D
specify the domain to use, the default is ipv4. Currently ipv4 and ipv6 are supported.

--dccp-port P
start DCCP at port P. For N dccp worker processes, ports P to P - 1 are used.

--dccp-ops N
stop dccp stress workers after N bogo operations.

--dccp-opts [ send | sendmsg | sendmmsg ]
by default, messages are sent using send(2). This option allows one to specify the sending method
using send(2), sendmsg(2) or sendmmsg(2). Note that sendmmsg is only available for Linux systems
that support this system call.

-D N, --dentry N
start N workers that create and remove directory entries. This should create file system meta
data activity. The directory entry names are suffixed by a gray-code encoded number to try to mix
up the hashing of the namespace.

--dentry-ops N
stop denty thrash workers after N bogo dentry operations.

--dentry-order [ forward | reverse | stride | random ]
specify unlink order of dentries, can be one of forward, reverse, stride or random. By default,
dentries are unlinked in random order. The forward order will unlink them from first to last,
reverse order will unlink them from last to first, stride order will unlink them by stepping
around order in a quasi-random pattern and random order will randomly select one of forward,
reverse or stride orders.

--dentries N
create N dentries per dentry thrashing loop, default is 2048.

--dev N
start N workers that exercise the /dev devices. Each worker runs 5 concurrent threads that perform
open(2), fstat(2), lseek(2), poll(2), fcntl(2), mmap(2), munmap(2), fsync(2) and close(2) on each
device. Note that watchdog devices are not exercised.

--dev-ops N
stop dev workers after N bogo device exercising operations.

--dev-file filename
specify the device file to exercise, for example, /dev/null. By default the stressor will work
through all the device files it can fine, however, this option allows a single device file to be
exercised.

--dev-shm N
start N workers that fallocate large files in /dev/shm and then mmap these into memory and touch
all the pages. This exercises pages being moved to/from the buffer cache. Linux only.

--dev-shm-ops N
stop after N bogo allocation and mmap /dev/shm operations.

--dir N
start N workers that create and remove directories using mkdir and rmdir.

--dir-ops N
stop directory thrash workers after N bogo directory operations.

--dir-dirs N
exercise dir on N directories. The default is 8192 directories, this allows 64 to 65536
directories to be used instead.

--dirdeep N
start N workers that create a depth-first tree of directories to a maximum depth as limited by
PATH_MAX or ENAMETOOLONG (which ever occurs first). By default, each level of the tree contains
one directory, but this can be increased to a maximum of 10 sub-trees using the --dirdeep-dir
option. To stress inode creation, a symlink and a hardlink to a file at the root of the tree is
created in each level.

--dirdeep-ops N
stop directory depth workers after N bogo directory operations.

--dirdeep-dirs N
create N directories at each tree level. The default is just 1 but can be increased to a maximum
of 36 per level.

--dirdeep-inodes N
consume up to N inodes per dirdeep stressor while creating directories and links. The value N can
be the number of inodes or a percentage of the total available free inodes on the filesystem being
used.

--dirmany N
start N stressors that create as many empty files in a directory as possible and then remove them.
The file creation phase stops when an error occurs (for example, out of inodes, too many files,
quota reached, etc.) and then the files are removed. This cycles until the the run time is reached
or the file creation count bogo-ops metric is reached. This is a much faster and light weight
directory exercising stressor compared to the dentry stressor.

--dirmany-ops N
stop dirmany stressors after N empty files have been created.

--dnotify N
start N workers performing file system activities such as making/deleting files/directories,
renaming files, etc. to stress exercise the various dnotify events (Linux only).

--dnotify-ops N
stop inotify stress workers after N dnotify bogo operations.

--dup N
start N workers that perform dup(2) and then close(2) operations on /dev/zero. The maximum opens
at one time is system defined, so the test will run up to this maximum, or 65536 open file
descriptors, which ever comes first.

--dup-ops N
stop the dup stress workers after N bogo open operations.

--dynlib N
start N workers that dynamically load and unload various shared libraries. This exercises memory
mapping and dynamic code loading and symbol lookups. See dlopen(3) for more details of this
mechanism.

--dynlib-ops N
stop workers after N bogo load/unload cycles.

--efivar N
start N works that exercise the Linux /sys/firmware/efi/vars interface by reading the EFI
variables. This is a Linux only stress test for platforms that support the EFI vars interface and
requires the CAP_SYS_ADMIN capability.

--efivar-ops N
stop the efivar stressors after N EFI variable read operations.

--enosys N
start N workers that exercise non-functional system call numbers. This calls a wide range of
system call numbers to see if it can break a system where these are not wired up correctly. It
also keeps track of system calls that exist (ones that don't return ENOSYS) so that it can focus
on purely finding and exercising non-functional system calls. This stressor exercises system calls
from 0 to __NR_syscalls + 1024, random system calls within constrained in the ranges of 0 to 2^8,
2^16, 2^24, 2^32, 2^40, 2^48, 2^56 and 2^64 bits, high system call numbers and various other bit
patterns to try to get wide coverage. To keep the environment clean, each system call being tested
runs in a child process with reduced capabilities.

--enosys-ops N
stop after N bogo enosys system call attempts

--env N
start N workers that creates numerous large environment variables to try to trigger out of memory
conditions using setenv(3). If ENOMEM occurs then the environment is emptied and another memory
filling retry occurs. The process is restarted if it is killed by the Out Of Memory (OOM) killer.

--env-ops N
stop after N bogo setenv/unsetenv attempts.

--epoll N
start N workers that perform various related socket stress activity using epoll_wait(2) to monitor
and handle new connections. This involves client/server processes performing rapid connect,
send/receives and disconnects on the local host. Using epoll allows a large number of connections
to be efficiently handled, however, this can lead to the connection table filling up and blocking
further socket connections, hence impacting on the epoll bogo op stats. For ipv4 and ipv6
domains, multiple servers are spawned on multiple ports. The epoll stressor is for Linux only.

--epoll-domain D
specify the domain to use, the default is unix (aka local). Currently ipv4, ipv6 and unix are
supported.

--epoll-port P
start at socket port P. For N epoll worker processes, ports P to (P * 4) - 1 are used for ipv4,
ipv6 domains and ports P to P - 1 are used for the unix domain.

--epoll-ops N
stop epoll workers after N bogo operations.

--eventfd N
start N parent and child worker processes that read and write 8 byte event messages between them
via the eventfd mechanism (Linux only).

--eventfd-ops N
stop eventfd workers after N bogo operations.

--eventfd-nonblock N
enable EFD_NONBLOCK to allow non-blocking on the event file descriptor. This will cause reads and
writes to return with EAGAIN rather the blocking and hence causing a high rate of polling I/O.

--exec N
start N workers continually forking children that exec stress-ng and then exit almost immediately.
If a system has pthread support then 1 in 4 of the exec's will be from inside a pthread to
exercise exec'ing from inside a pthread context.

--exec-ops N
stop exec stress workers after N bogo operations.

--exec-max P
create P child processes that exec stress-ng and then wait for them to exit per iteration. The
default is just 1; higher values will create many temporary zombie processes that are waiting to
be reaped. One can potentially fill up the process table using high values for --exec-max and
--exec.

--exit-group N
start N workers that create 16 pthreads and terminate the pthreads and the controlling child
process using exit_group(2). (Linux only stressor).

--exit-group-ops N
stop after N iterations of pthread creation and deletion loops.

-F N, --fallocate N
start N workers continually fallocating (preallocating file space) and ftruncating (file
truncating) temporary files. If the file is larger than the free space, fallocate will produce an
ENOSPC error which is ignored by this stressor.

--fallocate-bytes N
allocated file size, the default is 1 GB. One can specify the size as % of free space on the file
system or in units of Bytes, KBytes, MBytes and GBytes using the suffix b, k, m or g.

--fallocate-ops N
stop fallocate stress workers after N bogo fallocate operations.

--fanotify N
start N workers performing file system activities such as creating, opening, writing, reading and
unlinking files to exercise the fanotify event monitoring interface (Linux only). Each stressor
runs a child process to generate file events and a parent process to read file events using
fanotify. Has to be run with CAP_SYS_ADMIN capability.

--fanotify-ops N
stop fanotify stress workers after N bogo fanotify events.

--fault N
start N workers that generates minor and major page faults.

--fault-ops N
stop the page fault workers after N bogo page fault operations.

--fcntl N
start N workers that perform fcntl(2) calls with various commands. The exercised commands (if
available) are: F_DUPFD, F_DUPFD_CLOEXEC, F_GETFD, F_SETFD, F_GETFL, F_SETFL, F_GETOWN, F_SETOWN,
F_GETOWN_EX, F_SETOWN_EX, F_GETSIG, F_SETSIG, F_GETLK, F_SETLK, F_SETLKW, F_OFD_GETLK, F_OFD_SETLK
and F_OFD_SETLKW.

--fcntl-ops N
stop the fcntl workers after N bogo fcntl operations.

--fiemap N
start N workers that each create a file with many randomly changing extents and has 4 child
processes per worker that gather the extent information using the FS_IOC_FIEMAP ioctl(2).

--fiemap-ops N
stop after N fiemap bogo operations.

--fiemap-bytes N
specify the size of the fiemap'd file in bytes. One can specify the size as % of free space on
the file system or in units of Bytes, KBytes, MBytes and GBytes using the suffix b, k, m or g.
Larger files will contain more extents, causing more stress when gathering extent information.

--fifo N
start N workers that exercise a named pipe by transmitting 64 bit integers.

--fifo-ops N
stop fifo workers after N bogo pipe write operations.

--fifo-readers N
for each worker, create N fifo reader workers that read the named pipe using simple blocking
reads.

--file-ioctl N
start N workers that exercise various file specific ioctl(2) calls. This will attempt to use the
FIONBIO, FIOQSIZE, FIGETBSZ, FIOCLEX, FIONCLEX, FIONBIO, FIOASYNC, FIOQSIZE, FIFREEZE, FITHAW,
FICLONE, FICLONERANGE, FIONREAD, FIONWRITE and FS_IOC_RESVSP ioctls if these are defined.

--file-ioctl-ops N
stop file-ioctl workers after N file ioctl bogo operations.

--filename N
start N workers that exercise file creation using various length filenames containing a range of
allowed filename characters. This will try to see if it can exceed the file system allowed
filename length was well as test various filename lengths between 1 and the maximum allowed by the
file system.

--filename-ops N
stop filename workers after N bogo filename tests.

--filename-opts opt
use characters in the filename based on option 'opt'. Valid options are:

Option Description
probe default option, probe the file system for valid allowed characters
in a file name and use these
posix use characters as specified by The Open Group Base Specifications
Issue 7, POSIX.1-2008, 3.278 Portable Filename Character Set
ext use characters allowed by the ext2, ext3, ext4 file systems, namely
any 8 bit character apart from NUL and /

--flock N
start N workers locking on a single file.

--flock-ops N
stop flock stress workers after N bogo flock operations.

-f N, --fork N
start N workers continually forking children that immediately exit.

--fork-ops N
stop fork stress workers after N bogo operations.

--fork-max P
create P child processes and then wait for them to exit per iteration. The default is just 1;
higher values will create many temporary zombie processes that are waiting to be reaped. One can
potentially fill up the process table using high values for --fork-max and --fork.

--fork-vm
enable detrimental performance virtual memory advice using madvise on all pages of the forked
process. Where possible this will try to set every page in the new process with using madvise
MADV_MERGEABLE, MADV_WILLNEED, MADV_HUGEPAGE and MADV_RANDOM flags. Linux only.

--fp-error N
start N workers that generate floating point exceptions. Computations are performed to force and
check for the FE_DIVBYZERO, FE_INEXACT, FE_INVALID, FE_OVERFLOW and FE_UNDERFLOW exceptions. EDOM
and ERANGE errors are also checked.

--fp-error-ops N
stop after N bogo floating point exceptions.

--fpunch N
start N workers that punch and fill holes in a 16 MB file using five concurrent processes per
stressor exercising on the same file. Where available, this uses fallocate(2) FALLOC_FL_KEEP_SIZE,
FALLOC_FL_PUNCH_HOLE, FALLOC_FL_ZERO_RANGE, FALLOC_FL_COLLAPSE_RANGE and FALLOC_FL_INSERT_RANGE to
make and fill holes across the file and breaks it into multiple extents.

--fpunch-ops N
stop fpunch workers after N punch and fill bogo operations.

--fstat N
start N workers fstat'ing files in a directory (default is /dev).

--fstat-ops N
stop fstat stress workers after N bogo fstat operations.

--fstat-dir directory
specify the directory to fstat to override the default of /dev. All the files in the directory
will be fstat'd repeatedly.

--full N
start N workers that exercise /dev/full. This attempts to write to the device (which should
always get error ENOSPC), to read from the device (which should always return a buffer of zeros)
and to seek randomly on the device (which should always succeed). (Linux only).

--full-ops N
stop the stress full workers after N bogo I/O operations.

--funccall N
start N workers that call functions of 1 through to 9 arguments. By default functions with
uint64_t arguments are called, however, this can be changed using the --funccall-method option.

--funccall-ops N
stop the funccall workers after N bogo function call operations. Each bogo operation is 1000 calls
of functions of 1 through to 9 arguments of the chosen argument type.

--funccall-method method
specify the method of funccall argument type to be used. The default is uint64_t but can be one of
bool, uint8, uint16, uint32, uint64, uint128, float, double, longdouble, cfloat (complex float),
cdouble (complex double), clongdouble (complex long double), float16, float32, float64, float80,
float128, decimal32, decimal64 and decimal128. Note that some of these types are only available
with specific architectures and compiler versions.

--funcret N
start N workers that pass and return by value various small to large data types.

--funcret-ops N
stop the funcret workers after N bogo function call operations.

--funcret-method method
specify the method of funcret argument type to be used. The default is uint64_t but can be one of
uint8 uint16 uint32 uint64 uint128 float double longdouble float80 float128 decimal32 decimal64
decimal128 uint8x32 uint8x128 uint64x128.

--futex N
start N workers that rapidly exercise the futex system call. Each worker has two processes, a
futex waiter and a futex waker. The waiter waits with a very small timeout to stress the timeout
and rapid polled futex waiting. This is a Linux specific stress option.

--futex-ops N
stop futex workers after N bogo successful futex wait operations.

--get N
start N workers that call system calls that fetch data from the kernel, currently these are:
getpid, getppid, getcwd, getgid, getegid, getuid, getgroups, getpgrp, getpgid, getpriority,
getresgid, getresuid, getrlimit, prlimit, getrusage, getsid, gettid, getcpu, gettimeofday, uname,
adjtimex, sysfs. Some of these system calls are OS specific.

--get-ops N
stop get workers after N bogo get operations.

--getdent N
start N workers that recursively read directories /proc, /dev/, /tmp, /sys and /run using getdents
and getdents64 (Linux only).

--getdent-ops N
stop getdent workers after N bogo getdent bogo operations.

--getrandom N
start N workers that get 8192 random bytes from the /dev/urandom pool using the getrandom(2)
system call (Linux) or getentropy(2) (OpenBSD).

--getrandom-ops N
stop getrandom workers after N bogo get operations.

--goto N
start N workers that perform 1024 forward branches (to next instruction) or backward branches (to
previous instruction) for each bogo operation loop. By default, every 1024 branches the direction
is randomly chosen to be forward or backward. This stressor exercises suboptimal pipelined
execution and branch prediction logic.

--goto-ops N
stop goto workers after N bogo loops of 1024 branch instructions.

--goto-direction [ forward | backward | random ]
select the branching direction in the stressor loop, forward for forward only branching, backward
for a backward only branching, random for a random choice of forward or random branching every
1024 branches.

--handle N
start N workers that exercise the name_to_handle_at(2) and open_by_handle_at(2) system calls.
(Linux only).

--handle-ops N
stop after N handle bogo operations.

--hash N
start N workers that exercise various hashing functions. Random strings from 1 to 128 bytes are
hashed and the hashing rate and chi squared is calculated from the number of hashes performed over
a period of time. The chi squared value is the goodness-of-fit measure, it is the actual
distribution of items in hash buckets versus the expected distribution of items. Typically a chi
squared value of 0.95..1.05 indicates a good hash distribution.

--hash-ops N
stop after N hashing rounds

--hash-method M
specify the hashing method to use, by default all the hashing methods are cycled through. Methods
available are:

Method Description
all cycle through all the hashing methods
adler32 Mark Adler checksum, a modification of the Fletcher
checksum
coffin xor and 5 bit rotate left hash
coffin32 xor and 5 bit rotate left hash with 32 bit fetch
optimization
crc32c compute CRC32C (Castagnoli CRC32) integer hash
djb2a Dan Bernstein hash using the xor variant
fnv1a FNV-1a Fowler-Noll-Vo hash using the xor then multiply
variant
jenkin Jenkin's integer hash
kandr Kernighan and Richie's multiply by 31 and add hash from
"The C Programming Language", 2nd Edition
knuth Donald E. Knuth's hash from "The Art Of Computer
Programming", Volume 3, chapter 6.4
loselose Kernighan and Richie's simple hash from "The C
Programming Language", 1st Edition
mid5 xor shift hash of the middle 5 characters of the string.
Designed by Colin Ian King
muladd32 simple multiply and add hash using 32 bit math and xor
folding of overflow
muladd64 simple multiply and add hash using 64 bit math and xor
folding of overflow
mulxror64 64 bit multiply, xor and rotate right. Mangles 64 bits
where possible. Designed by Colin Ian King
murmur3_32 murmur3_32 hash, Austin Appleby's Murmur3 hash, 32 bit
variant
nhash exim's nhash.
pjw a non-cryptographic hash function created by Peter J.
Weinberger of AT&T Bell Labs, used in UNIX ELF object
files
sdbm sdbm hash as used in the SDBM database and GNU awk
x17 multiply by 17 and add. The multiplication can be
optimized down to a fast right shift by 4 and add on
some architectures
xor simple rotate shift and xor of values
xxhash the "Extremely fast" hash in non-streaming mode

-d N, --hdd N
start N workers continually writing, reading and removing temporary files. The default mode is to
stress test sequential writes and reads. With the --aggressive option enabled without any
--hdd-opts options the hdd stressor will work through all the --hdd-opt options one by one to
cover a range of I/O options.

--hdd-bytes N
write N bytes for each hdd process, the default is 1 GB. One can specify the size as % of free
space on the file system or in units of Bytes, KBytes, MBytes and GBytes using the suffix b, k, m
or g.

--hdd-opts list
specify various stress test options as a comma separated list. Options are as follows:

Option Description
direct try to minimize cache effects of the I/O. File I/O writes are
performed directly from user space buffers and synchronous transfer
is also attempted. To guarantee synchronous I/O, also use the sync
option.
dsync ensure output has been transferred to underlying hardware and file
metadata has been updated (using the O_DSYNC open flag). This is
equivalent to each write(2) being followed by a call to
fdatasync(2). See also the fdatasync option.
fadv-dontneed advise kernel to expect the data will not be accessed in the near
future.
fadv-noreuse advise kernel to expect the data to be accessed only once.
fadv-normal advise kernel there are no explicit access pattern for the data.
This is the default advice assumption.
fadv-rnd advise kernel to expect random access patterns for the data.

fadv-seq advise kernel to expect sequential access patterns for the data.
fadv-willneed advise kernel to expect the data to be accessed in the near future.
fsync flush all modified in-core data after each write to the output
device using an explicit fsync(2) call.
fdatasync similar to fsync, but do not flush the modified metadata unless
metadata is required for later data reads to be handled correctly.
This uses an explicit fdatasync(2) call.
iovec use readv/writev multiple buffer I/Os rather than read/write.
Instead of 1 read/write operation, the buffer is broken into an
iovec of 16 buffers.
noatime do not update the file last access timestamp, this can reduce
metadata writes.
sync ensure output has been transferred to underlying hardware (using
the O_SYNC open flag). This is equivalent to a each write(2) being
followed by a call to fsync(2). See also the fsync option.
rd-rnd read data randomly. By default, written data is not read back,
however, this option will force it to be read back randomly.
rd-seq read data sequentially. By default, written data is not read back,
however, this option will force it to be read back sequentially.
syncfs write all buffered modifications of file metadata and data on the
filesystem that contains the hdd worker files.
utimes force update of file timestamp which may increase metadata writes.
wr-rnd write data randomly. The wr-seq option cannot be used at the same
time.
wr-seq write data sequentially. This is the default if no write modes are
specified.

Note that some of these options are mutually exclusive, for example, there can be only one method of
writing or reading. Also, fadvise flags may be mutually exclusive, for example fadv-willneed cannot be
used with fadv-dontneed.

--hdd-ops N
stop hdd stress workers after N bogo operations.

--hdd-write-size N
specify size of each write in bytes. Size can be from 1 byte to 4MB.

--heapsort N
start N workers that sort 32 bit integers using the BSD heapsort.

--heapsort-ops N
stop heapsort stress workers after N bogo heapsorts.

--heapsort-size N
specify number of 32 bit integers to sort, default is 262144 (256 × 1024).

--hrtimers N
start N workers that exercise high resolution times at a high frequency. Each stressor starts 32
processes that run with random timer intervals of 0..499999 nanoseconds. Running this stressor
with appropriate privilege will run these with the SCHED_RR policy.

--hrtimers-ops N
stop hrtimers stressors after N timer event bogo operations

--hrtimers-adjust
enable automatic timer rate adjustment to try to maximize the hrtimer frequency. The signal rate
is measured every 0.1 seconds and the hrtimer delay is adjusted to try and set the optimal hrtimer
delay to generate the highest hrtimer rates.

--hsearch N
start N workers that search a 80% full hash table using hsearch(3). By default, there are 8192
elements inserted into the hash table. This is a useful method to exercise access of memory and
processor cache.

--hsearch-ops N
stop the hsearch workers after N bogo hsearch operations are completed.

--hsearch-size N
specify the number of hash entries to be inserted into the hash table. Size can be from 1K to 4M.

--icache N
start N workers that stress the instruction cache by forcing instruction cache reloads. This is
achieved by modifying an instruction cache line, causing the processor to reload it when we call
a function in inside it. Currently only verified and enabled for Intel x86 CPUs.

--icache-ops N
stop the icache workers after N bogo icache operations are completed.

--icmp-flood N
start N workers that flood localhost with randonly sized ICMP ping packets. This stressor
requires the CAP_NET_RAW capbility.

--icmp-flood-ops N
stop icmp flood workers after N ICMP ping packets have been sent.

--idle-scan N
start N workers that scan the idle page bitmap across a range of physical pages. This sets and
checks for idle pages via the idle page tracking interface /sys/kernel/mm/page_idle/bitmap. This
is for Linux only.

--idle-scan-ops N
stop after N bogo page scan operations. Currently one bogo page scan operation is equivalent to
setting and checking 64 physical pages.

--idle-page N
start N workers that walks through every page exercising the Linux /sys/kernel/mm/page_idle/bitmap
interface. Requires CAP_SYS_RESOURCE capability.

--idle-page-ops N
stop after N bogo idle page operations.

--inode-flags N
start N workers that exercise inode flags using the FS_IOC_GETFLAGS and FS_IOC_SETFLAGS ioctl(2).
This attempts to apply all the available inode flags onto a directory and file even if the
underlying file system may not support these flags (errors are just ignored). Each worker runs 4
threads that exercise the flags on the same directory and file to try to force races. This is a
Linux only stressor, see ioctl_iflags(2) for more details.

--inode-flags-ops N
stop the inode-flags workers after N ioctl flag setting attempts.

--inotify N
start N workers performing file system activities such as making/deleting files/directories,
moving files, etc. to stress exercise the various inotify events (Linux only).

--inotify-ops N
stop inotify stress workers after N inotify bogo operations.

-i N, --io N
start N workers continuously calling sync(2) to commit buffer cache to disk. This can be used in
conjunction with the --hdd options.

--io-ops N
stop io stress workers after N bogo operations.

--iomix N
start N workers that perform a mix of sequential, random and memory mapped read/write operations
as well as random copy file read/writes, forced sync'ing and (if run as root) cache dropping.
Multiple child processes are spawned to all share a single file and perform different I/O
operations on the same file.

--iomix-bytes N
write N bytes for each iomix worker process, the default is 1 GB. One can specify the size as % of
free space on the file system or in units of Bytes, KBytes, MBytes and GBytes using the suffix b,
k, m or g.

--iomix-ops N
stop iomix stress workers after N bogo iomix I/O operations.

--ioport N
start N workers than perform bursts of 16 reads and 16 writes of ioport 0x80 (x86 Linux systems
only). I/O performed on x86 platforms on port 0x80 will cause delays on the CPU performing the
I/O.

--ioport-ops N
stop the ioport stressors after N bogo I/O operations

--ioport-opts [ in | out | inout ]
specify if port reads in, port read writes out or reads and writes are to be performed. The
default is both in and out.

--ioprio N
start N workers that exercise the ioprio_get(2) and ioprio_set(2) system calls (Linux only).

--ioprio-ops N
stop after N io priority bogo operations.

--io-uring N
start N workers that perform iovec write and read I/O operations using the Linux io-uring
interface. On each bogo-loop 1024 × 512 byte writes and 1024 × reads are performed on a temporary
file.

--io-uring-ops
stop after N rounds of write and reads.

--ipsec-mb N
start N workers that perform cryptographic processing using the highly optimized Intel Multi-
Buffer Crypto for IPsec library. Depending on the features available, SSE, AVX, AVX and AVX512 CPU
features will be used on data encrypted by SHA, DES, CMAC, CTR, HMAC MD5, HMAC SHA1 and HMAC
SHA512 cryptographic routines. This is only available for x86-64 modern Intel CPUs.

--ipsec-mb-ops N
stop after N rounds of processing of data using the cryptographic routines.

--ipsec-mb-feature [ sse | avx | avx2 | avx512 ]
Just use the specified processor CPU feature. By default, all the available features for the CPU
are exercised.

--itimer N
start N workers that exercise the system interval timers. This sets up an ITIMER_PROF itimer that
generates a SIGPROF signal. The default frequency for the itimer is 1 MHz, however, the Linux
kernel will set this to be no more that the jiffy setting, hence high frequency SIGPROF signals
are not normally possible. A busy loop spins on getitimer(2) calls to consume CPU and hence
decrement the itimer based on amount of time spent in CPU and system time.

--itimer-ops N
stop itimer stress workers after N bogo itimer SIGPROF signals.

--itimer-freq F
run itimer at F Hz; range from 1 to 1000000 Hz. Normally the highest frequency is limited by the
number of jiffy ticks per second, so running above 1000 Hz is difficult to attain in practice.

--itimer-rand
select an interval timer frequency based around the interval timer frequency +/- 12.5% random
jitter. This tries to force more variability in the timer interval to make the scheduling less
predictable.

--judy N
start N workers that insert, search and delete 32 bit integers in a Judy array using a predictable
yet sparse array index. By default, there are 131072 integers used in the Judy array. This is a
useful method to exercise random access of memory and processor cache.

--judy-ops N
stop the judy workers after N bogo judy operations are completed.

--judy-size N
specify the size (number of 32 bit integers) in the Judy array to exercise. Size can be from 1K
to 4M 32 bit integers.

--kcmp N
start N workers that use kcmp(2) to compare parent and child processes to determine if they share
kernel resources. Supported only for Linux and requires CAP_SYS_PTRACE capability.

--kcmp-ops N
stop kcmp workers after N bogo kcmp operations.

--key N
start N workers that create and manipulate keys using add_key(2) and ketctl(2). As many keys are
created as the per user limit allows and then the following keyctl commands are exercised on each
key: KEYCTL_SET_TIMEOUT, KEYCTL_DESCRIBE, KEYCTL_UPDATE, KEYCTL_READ, KEYCTL_CLEAR and
KEYCTL_INVALIDATE.

--key-ops N
stop key workers after N bogo key operations.

--kill N
start N workers sending SIGUSR1 kill signals to a SIG_IGN signal handler in the stressor and
SIGUSR1 kill signal to a child stressor with a SIGUSR1 handler. Most of the process time will end
up in kernel space.

--kill-ops N
stop kill workers after N bogo kill operations.

--klog N
start N workers exercising the kernel syslog(2) system call. This will attempt to read the kernel
log with various sized read buffers. Linux only.

--klog-ops N
stop klog workers after N syslog operations.

--kvm N
start N workers that create, run and destroy a minimal virtual machine. The virtual machine reads,
increments and writes to port 0x80 in a spin loop and the stressor handles the I/O transactions.
Currently for x86 and Linux only.

--kvm-ops N
stop kvm stressors after N virtual machines have been created, run and destroyed.

--l1cache N
start N workers that exercise the CPU level 1 cache with reads and writes. A cache aligned buffer
that is twice the level 1 cache size is read and then written in level 1 cache set sized steps
over each level 1 cache set. This is designed to exercise cache block evictions. The bogo-op count
measures the number of million cache lines touched. Where possible, the level 1 cache geometry is
determined from the kernel, however, this is not possible on some architectures or kernels, so one
may need to specify these manually. One can specify 3 out of the 4 cache geometric parameters,
these are as follows:

--l1cache-line-size N
specify the level 1 cache line size (in bytes)

--l1cache-sets N
specify the number of level 1 cache sets

--l1cache-size N
specify the level 1 cache size (in bytes)

--l1cache-ways N
specify the number of level 1 cache ways

--landlock N
start N workers that exercise Linux 5.13 landlocking. A range of landlock_create_ruleset flags are
exercised with a read only file rule to see if a directory can be accessed and a read-write file
create can be blocked. Each ruleset attempt is exercised in a new child context and this is the
limiting factor on the speed of the stressor.

--landlock-ops N
stop the landlock stressors after N landlock ruleset bogo operations.

--lease N
start N workers locking, unlocking and breaking leases via the fcntl(2) F_SETLEASE operation. The
parent processes continually lock and unlock a lease on a file while a user selectable number of
child processes open the file with a non-blocking open to generate SIGIO lease breaking
notifications to the parent. This stressor is only available if F_SETLEASE, F_WRLCK and F_UNLCK
support is provided by fcntl(2).

--lease-ops N
stop lease workers after N bogo operations.

--lease-breakers N
start N lease breaker child processes per lease worker. Normally one child is plenty to force
many SIGIO lease breaking notification signals to the parent, however, this option allows one to
specify more child processes if required.

--link N
start N workers creating and removing hardlinks.

--link-ops N
stop link stress workers after N bogo operations.

--list N
start N workers that exercise list data structures. The default is to add, find and remove 5,000
64 bit integers into circleq (doubly linked circle queue), list (doubly linked list), slist
(singly linked list), slistt (singly linked list using tail), stailq (singly linked tail queue)
and tailq (doubly linked tail queue) lists. The intention of this stressor is to exercise memory
and cache with the various list operations.

--list-ops N
stop list stressors after N bogo ops. A bogo op covers the addition, finding and removing all the
items into the list(s).

--list-size N
specify the size of the list, where N is the number of 64 bit integers to be added into the list.

--loadavg N
start N workers that attempt to create thousands of pthreads that run at the lowest nice priority
to force very high load averages. Linux systems will also perform some I/O writes as pending I/O
is also factored into system load accounting.

--loadavg-ops N
stop loadavg workers after N bogo scheduling yields by the pthreads have been reached.

--lockbus N
start N workers that rapidly lock and increment 64 bytes of randomly chosen memory from a 16MB
mmap'd region (Intel x86 and ARM CPUs only). This will cause cacheline misses and stalling of
CPUs.

--lockbus-ops N
stop lockbus workers after N bogo operations.

--locka N
start N workers that randomly lock and unlock regions of a file using the POSIX advisory locking
mechanism (see fcntl(2), F_SETLK, F_GETLK). Each worker creates a 1024 KB file and attempts to
hold a maximum of 1024 concurrent locks with a child process that also tries to hold 1024
concurrent locks. Old locks are unlocked in a first-in, first-out basis.

--locka-ops N
stop locka workers after N bogo locka operations.

--lockf N
start N workers that randomly lock and unlock regions of a file using the POSIX lockf(3) locking
mechanism. Each worker creates a 64 KB file and attempts to hold a maximum of 1024 concurrent
locks with a child process that also tries to hold 1024 concurrent locks. Old locks are unlocked
in a first-in, first-out basis.

--lockf-ops N
stop lockf workers after N bogo lockf operations.

--lockf-nonblock
instead of using blocking F_LOCK lockf(3) commands, use non-blocking F_TLOCK commands and re-try
if the lock failed. This creates extra system call overhead and CPU utilisation as the number of
lockf workers increases and should increase locking contention.

--lockofd N
start N workers that randomly lock and unlock regions of a file using the Linux open file
description locks (see fcntl(2), F_OFD_SETLK, F_OFD_GETLK). Each worker creates a 1024 KB file
and attempts to hold a maximum of 1024 concurrent locks with a child process that also tries to
hold 1024 concurrent locks. Old locks are unlocked in a first-in, first-out basis.

--lockofd-ops N
stop lockofd workers after N bogo lockofd operations.

--longjmp N
start N workers that exercise setjmp(3)/longjmp(3) by rapid looping on longjmp calls.

--longjmp-ops N
stop longjmp stress workers after N bogo longjmp operations (1 bogo op is 1000 longjmp calls).

--loop N
start N workers that exercise the loopback control device. This creates 2MB loopback devices,
expands them to 4MB, performs some loopback status information get and set operations and then
destoys them. Linux only and requires CAP_SYS_ADMIN capability.

--loop-ops N
stop after N bogo loopback creation/deletion operations.

--lsearch N
start N workers that linear search a unsorted array of 32 bit integers using lsearch(3). By
default, there are 8192 elements in the array. This is a useful method to exercise sequential
access of memory and processor cache.

--lsearch-ops N
stop the lsearch workers after N bogo lsearch operations are completed.

--lsearch-size N
specify the size (number of 32 bit integers) in the array to lsearch. Size can be from 1K to 4M.

--madvise N
start N workers that apply random madvise(2) advise settings on pages of a 4MB file backed shared
memory mapping.

--madvise-ops N
stop madvise stressors after N bogo madvise operations.

--malloc N
start N workers continuously calling malloc(3), calloc(3), realloc(3) and free(3). By default, up
to 65536 allocations can be active at any point, but this can be altered with the --malloc-max
option. Allocation, reallocation and freeing are chosen at random; 50% of the time memory is
allocation (via malloc, calloc or realloc) and 50% of the time allocations are free'd. Allocation
sizes are also random, with the maximum allocation size controlled by the --malloc-bytes option,
the default size being 64K. The worker is re-started if it is killed by the out of memory (OOM)
killer.

--malloc-bytes N
maximum per allocation/reallocation size. Allocations are randomly selected from 1 to N bytes. One
can specify the size as % of total available memory or in units of Bytes, KBytes, MBytes and
GBytes using the suffix b, k, m or g. Large allocation sizes cause the memory allocator to use
mmap(2) rather than expanding the heap using brk(2).

--malloc-max N
maximum number of active allocations allowed. Allocations are chosen at random and placed in an
allocation slot. Because about 50%/50% split between allocation and freeing, typically half of the
allocation slots are in use at any one time.

--malloc-ops N
stop after N malloc bogo operations. One bogo operations relates to a successful malloc(3),
calloc(3) or realloc(3).

--malloc-pthreads N
specify number of malloc stressing concurrent pthreads to run. The default is 0 (just one main
process, no pthreads). This option will do nothing if pthreads are not supported.

--malloc-thresh N
specify the threshold where malloc uses mmap(2) instead of sbrk(2) to allocate more memory. This
is only available on systems that provide the GNU C mallopt(3) tuning function.

--malloc-touch
touch every allocated page to force pages to be populated in memory. This will increase the memory
pressure and exercise the virtual memory harder. By default the malloc stressor will madvise pages
into memory or use mincore to check for non-resident memory pages and try to force them into
memory; this option aggressively forces pages to be memory resident.

--matrix N
start N workers that perform various matrix operations on floating point values. Testing on 64 bit
x86 hardware shows that this provides a good mix of memory, cache and floating point operations
and is an excellent way to make a CPU run hot.

By default, this will exercise all the matrix stress methods one by one. One can specify a
specific matrix stress method with the --matrix-method option.

--matrix-ops N
stop matrix stress workers after N bogo operations.

--matrix-method method
specify a matrix stress method. Available matrix stress methods are described as follows:

Method Description
all iterate over all the below matrix stress methods
add add two N × N matrices
copy copy one N × N matrix to another
div divide an N × N matrix by a scalar
frobenius Frobenius product of two N × N matrices
hadamard Hadamard product of two N × N matrices
identity create an N × N identity matrix
mean arithmetic mean of two N × N matrices
mult multiply an N × N matrix by a scalar
negate negate an N × N matrix
prod product of two N × N matrices
sub subtract one N × N matrix from another N × N matrix
square multiply an N × N matrix by itself
trans transpose an N × N matrix
zero zero an N × N matrix

--matrix-size N
specify the N × N size of the matrices. Smaller values result in a floating point compute
throughput bound stressor, where as large values result in a cache and/or memory bandwidth bound
stressor.

--matrix-yx
perform matrix operations in order y by x rather than the default x by y. This is suboptimal
ordering compared to the default and will perform more data cache stalls.

--matrix-3d N
start N workers that perform various 3D matrix operations on floating point values. Testing on 64
bit x86 hardware shows that this provides a good mix of memory, cache and floating point
operations and is an excellent way to make a CPU run hot.

By default, this will exercise all the 3D matrix stress methods one by one. One can specify a
specific 3D matrix stress method with the --matrix-3d-method option.

--matrix-3d-ops N
stop the 3D matrix stress workers after N bogo operations.

--matrix-3d-method method
specify a 3D matrix stress method. Available 3D matrix stress methods are described as follows:

Method Description
all iterate over all the below matrix stress methods
add add two N × N × N matrices
copy copy one N × N × N matrix to another
div divide an N × N × N matrix by a scalar
frobenius Frobenius product of two N × N × N matrices
hadamard Hadamard product of two N × N × N matrices
identity create an N × N × N identity matrix
mean arithmetic mean of two N × N × N matrices
mult multiply an N × N × N matrix by a scalar
negate negate an N × N × N matrix
sub subtract one N × N × N matrix from another N × N × N matrix
trans transpose an N × N × N matrix
zero zero an N × N × N matrix

--matrix-3d-size N
specify the N × N × N size of the matrices. Smaller values result in a floating point compute
throughput bound stressor, where as large values result in a cache and/or memory bandwidth bound
stressor.

--matrix-3d-zyx
perform matrix operations in order z by y by x rather than the default x by y by z. This is
suboptimal ordering compared to the default and will perform more data cache stalls.

--mcontend N
start N workers that produce memory contention read/write patterns. Each stressor runs with 5
threads that read and write to two different mappings of the same underlying physical page.
Various caching operations are also exercised to cause sub-optimal memory access patterns. The
threads also randomly change CPU affinity to exercise CPU and memory migration stress.

--mcontend-ops N
stop mcontend stressors after N bogo read/write operations.

--membarrier N
start N workers that exercise the membarrier system call (Linux only).

--membarrier-ops N
stop membarrier stress workers after N bogo membarrier operations.

--memcpy N
start N workers that copy 2MB of data from a shared region to a buffer using memcpy(3) and then
move the data in the buffer with memmove(3) with 3 different alignments. This will exercise
processor cache and system memory.

--memcpy-ops N
stop memcpy stress workers after N bogo memcpy operations.

--memcpy-method [ all | libc | builtin | naive ]
specify a memcpy copying method. Available memcpy methods are described as follows:

Method Description
all use libc, builtin and naive methods
libc use libc memcpy and memmove functions, this is the default
builtin use the compiler built in optimized memcpy and memmove functions
naive use naive byte by byte copying and memory moving build with default
compiler optimization flags
naive_o0 use unoptimized naive byte by byte copying and memory moving
naive_o3 use optimized naive byte by byte copying and memory moving build
with -O3 optimization and where possible use CPU specific
optimizations

--memfd N
start N workers that create allocations of 1024 pages using memfd_create(2) and ftruncate(2) for
allocation and mmap(2) to map the allocation into the process address space. (Linux only).

--memfd-bytes N
allocate N bytes per memfd stress worker, the default is 256MB. One can specify the size in as %
of total available memory or in units of Bytes, KBytes, MBytes and GBytes using the suffix b, k, m
or g.

--memfd-fds N
create N memfd file descriptors, the default is 256. One can select 8 to 4096 memfd file
descriptions with this option.

--memfd-ops N
stop after N memfd-create(2) bogo operations.

--memhotplug N
start N workers that offline and online memory hotplug regions. Linux only and requires
CAP_SYS_ADMIN capabilities.

--memhotplug-ops N
stop memhotplug stressors after N memory offline and online bogo operations.

--memrate N
start N workers that exercise a buffer with 1024, 512, 256, 128, 64, 32, 16 and 8 bit reads and
writes. 1024, 512 and 256 reads and writes are available with compilers that support integer
vectors. x86-64 cpus that support uncached (non-temporal "nt") writes also exercise 128, 64 and
32 writes providing higher write rates than the normal cached writes. CPUs that support
prefetching reads also exercise 64 prefetched "pf" reads. This memory stressor allows one to also
specify the maximum read and write rates. The stressors will run at maximum speed if no read or
write rates are specified.

--memrate-ops N
stop after N bogo memrate operations.

--memrate-bytes N
specify the size of the memory buffer being exercised. The default size is 256MB. One can specify
the size in units of Bytes, KBytes, MBytes and GBytes using the suffix b, k, m or g.

--memrate-rd-mbs N
specify the maximum allowed read rate in MB/sec. The actual read rate is dependent on scheduling
jitter and memory accesses from other running processes.

--memrate-wr-mbs N
specify the maximum allowed read rate in MB/sec. The actual write rate is dependent on scheduling
jitter and memory accesses from other running processes.

--memthrash N
start N workers that thrash and exercise a 16MB buffer in various ways to try and trip thermal
overrun. Each stressor will start 1 or more threads. The number of threads is chosen so that
there will be at least 1 thread per CPU. Note that the optimal choice for N is a value that
divides into the number of CPUs.

--memthrash-ops N
stop after N memthrash bogo operations.

--memthrash-method method
specify a memthrash stress method. Available memthrash stress methods are described as follows:

Method Description
all iterate over all the below memthrash methods
chunk1 memset 1 byte chunks of random data into random locations
chunk8 memset 8 byte chunks of random data into random locations
chunk64 memset 64 byte chunks of random data into random locations
chunk256 memset 256 byte chunks of random data into random locations
chunkpage memset page size chunks of random data into random locations
flip flip (invert) all bits in random locations
flush flush cache line in random locations
lock lock randomly choosing locations (Intel x86 and ARM CPUs only)
matrix treat memory as a 2 × 2 matrix and swap random elements
memmove copy all the data in buffer to the next memory location
memset memset the memory with random data
memset64 memset the memory with a random 64 bit value in 64 byte chunks
using non-temporal stores if possible or normal stores as a
fallback
mfence stores with write serialization
prefetch prefetch data at random memory locations
random randomly run any of the memthrash methods except for 'random' and
'all'
spinread spin loop read the same random location 2^19 times
spinwrite spin loop write the same random location 2^19 times
swap step through memory swapping bytes in steps of 65 and 129 byte
strides

--mergesort N
start N workers that sort 32 bit integers using the BSD mergesort.

--mergesort-ops N
stop mergesort stress workers after N bogo mergesorts.

--mergesort-size N
specify number of 32 bit integers to sort, default is 262144 (256 × 1024).

--mincore N
start N workers that walk through all of memory 1 page at a time checking if the page mapped and
also is resident in memory using mincore(2). It also maps and unmaps a page to check if the page
is mapped or not using mincore(2).

--mincore-ops N
stop after N mincore bogo operations. One mincore bogo op is equivalent to a 300 mincore(2) calls.
--mincore-random instead of walking through pages sequentially, select pages at random. The chosen
address is iterated over by shifting it right one place and checked by mincore until the address
is less or equal to the page size.

--misaligned N
start N workers that perform misaligned read and writes. By default, this will exercise 128 bit
misaligned read and writes in 8 x 16 bits, 4 x 32 bits, 2 x 64 bits and 1 x 128 bits at the start
of a page boundary, at the end of a page boundary and over a cache boundary. Misaligned read and
writes operate at 1 byte offset from the natural alignment of the data type. On some architectures
this can cause SIGBUS, SIGILL or SIGSEGV, these are handled and the misaligned stressor method
causing the error is disabled.

--misaligned-ops N
stop after N misaligned bogo operation. A misaligned bogo op is equivalent to 65536 x 128 bit
reads or writes.

--misaligned-method M
Available misaligned stress methods are described as follows:

Method Description
all iterate over all the following misaligned methods
int16rd 8 x 16 bit integer reads
int16wr 8 x 16 bit integer writes
int16inc 8 x 16 bit integer increments
int16atomic 8 x 16 bit atomic integer increments
int32rd 4 x 32 bit integer reads
int32wr 4 x 32 bit integer writes
int32wtnt 4 x 32 bit non-termporal stores (x86 only)
int32inc 4 x 32 bit integer increments
int32atomic 4 x 32 bit atomic integer increments
int64rd 2 x 64 bit integer reads
int64wr 2 x 64 bit integer writes
int64wtnt 4 x 64 bit non-termporal stores (x86 only)
int64inc 2 x 64 bit integer increments
int64atomic 2 x 64 bit atomic integer increments
int128rd 1 x 128 bit integer reads
int128wr 1 x 128 bit integer writes
int128inc 1 x 128 bit integer increments
int128atomic 1 x 128 bit atomic integer increments

Note that some of these options (128 bit integer and/or atomic operations) may not be available on some
systems.

--mknod N
start N workers that create and remove fifos, empty files and named sockets using mknod and
unlink.

--mknod-ops N
stop directory thrash workers after N bogo mknod operations.

--mlock N
start N workers that lock and unlock memory mapped pages using mlock(2), munlock(2), mlockall(2)
and munlockall(2). This is achieved by the mapping of three contiguous pages and then locking the
second page, hence ensuring non-contiguous pages are locked . This is then repeated until the
maximum allowed mlocks or a maximum of 262144 mappings are made. Next, all future mappings are
mlocked and the worker attempts to map 262144 pages, then all pages are munlocked and the pages
are unmapped.

--mlock-ops N
stop after N mlock bogo operations.

--mlockmany N
start N workers that fork off a default of 1024 child processes in total; each child will attempt
to anonymously mmap and mlock the maximum allowed mlockable memory size. The stress test attempts
to avoid swapping by tracking low memory and swap allocations (but some swapping may occur). Once
either the maximum number of child process is reached or all mlockable in-core memory is locked
then child processes are killed and the stress test is repeated.

--mlockmany-ops N
stop after N mlockmany (mmap and mlock) operations.

--mlockmany-procs N
set the number of child processes to create per stressor. The default is to start a maximum of
1024 child processes in total across all the stressors. This option allows the setting of N child
processes per stressor.

--mmap N
start N workers continuously calling mmap(2)/munmap(2). The initial mapping is a large chunk
(size specified by --mmap-bytes) followed by pseudo-random 4K unmappings, then pseudo-random 4K
mappings, and then linear 4K unmappings. Note that this can cause systems to trip the kernel OOM
killer on Linux systems if not enough physical memory and swap is not available. The MAP_POPULATE
option is used to populate pages into memory on systems that support this. By default, anonymous
mappings are used, however, the --mmap-file and --mmap-async options allow one to perform file
based mappings if desired.

--mmap-ops N
stop mmap stress workers after N bogo operations.

--mmap-async
enable file based memory mapping and use asynchronous msync'ing on each page, see --mmap-file.

--mmap-bytes N
allocate N bytes per mmap stress worker, the default is 256MB. One can specify the size as % of
total available memory or in units of Bytes, KBytes, MBytes and GBytes using the suffix b, k, m or
g.

--mmap-file
enable file based memory mapping and by default use synchronous msync'ing on each page.

--mmap-mmap2
use mmap2 for 4K page aligned offsets if mmap2 is available, otherwise fall back to mmap.

--mmap-mprotect
change protection settings on each page of memory. Each time a page or a group of pages are
mapped or remapped then this option will make the pages read-only, write-only, exec-only, and
read-write.

--mmap-odirect
enable file based memory mapping and use O_DIRECT direct I/O.

--mmap-osync
enable file based memory mapping and used O_SYNC synchronous I/O integrity completion.

--mmapaddr N
start N workers that memory map pages at a random memory location that is not already mapped. On
64 bit machines the random address is randomly chosen 32 bit or 64 bit address. If the mapping
works a second page is memory mapped from the first mapped address. The stressor exercises
mmap/munmap, mincore and segfault handling.

--mmapaddr-ops N
stop after N random address mmap bogo operations.

--mmapfork N
start N workers that each fork off 32 child processes, each of which tries to allocate some of the
free memory left in the system (and trying to avoid any swapping). The child processes then hint
that the allocation will be needed with madvise(2) and then memset it to zero and hint that it is
no longer needed with madvise before exiting. This produces significant amounts of VM activity, a
lot of cache misses and with minimal swapping.

--mmapfork-ops N
stop after N mmapfork bogo operations.

--mmapfixed N
start N workers that perform fixed address allocations from the top virtual address down to 128K.
The allocated sizes are from 1 page to 8 pages and various random mmap flags are used
MAP_SHARED/MAP_PRIVATE, MAP_LOCKED, MAP_NORESERVE, MAP_POPULATE. If successfully map'd then the
allocation is remap'd to an address that is several pages higher in memory. Mappings and
remappings are madvised with random madvise options to further exercise the mappings.

--mmapfixed-ops N
stop after N mmapfixed memory mapping bogo operations.

--mmaphuge N
start N workers that attempt to mmap a set of huge pages and large huge page sized mappings.
Successful mappings are madvised with MADV_NOHUGEPAGE and MADV_HUGEPAGE settings and then 1/64th
of the normal small page size pages are touched. Finally, an attempt to unmap a small page size
page at the end of the mapping is made (these may fail on huge pages) before the set of pages are
unmapped. By default 8192 mappings are attempted per round of mappings or until swapping is
detected.

--mmaphuge-ops N
stop after N mmaphuge bogo operations

--mmaphuge-mmaps N
set the number of huge page mappings to attempt in each round of mappings. The default is 8192
mappings.

--mmapmany N
start N workers that attempt to create the maximum allowed per-process memory mappings. This is
achieved by mapping 3 contiguous pages and then unmapping the middle page hence splitting the
mapping into two. This is then repeated until the maximum allowed mappings or a maximum of 262144
mappings are made.

--mmapmany-ops N
stop after N mmapmany bogo operations

--mq N start N sender and receiver processes that continually send and receive messages using POSIX
message queues. (Linux only).

--mq-ops N
stop after N bogo POSIX message send operations completed.

--mq-size N
specify size of POSIX message queue. The default size is 10 messages and most Linux systems this
is the maximum allowed size for normal users. If the given size is greater than the allowed
message queue size then a warning is issued and the maximum allowed size is used instead.

--mremap N
start N workers continuously calling mmap(2), mremap(2) and munmap(2). The initial anonymous
mapping is a large chunk (size specified by --mremap-bytes) and then iteratively halved in size by
remapping all the way down to a page size and then back up to the original size. This worker is
only available for Linux.

--mremap-ops N
stop mremap stress workers after N bogo operations.

--mremap-bytes N
initially allocate N bytes per remap stress worker, the default is 256MB. One can specify the size
in units of Bytes, KBytes, MBytes and GBytes using the suffix b, k, m or g.

--mremap-mlock
attempt to mlock remapped pages into memory prohibiting them from being paged out. This is a no-
op if mlock(2) is not available.

--msg N
start N sender and receiver processes that continually send and receive messages using System V
message IPC.

--msg-ops N
stop after N bogo message send operations completed.

--msg-types N
select the quality of message types (mtype) to use. By default, msgsnd sends messages with a mtype
of 1, this option allows one to send messages types in the range 1..N to exercise the message
queue receive ordering. This will also impact throughput performance.

--msync N
start N stressors that msync data from a file backed memory mapping from memory back to the file
and msync modified data from the file back to the mapped memory. This exercises the msync(2)
MS_SYNC and MS_INVALIDATE sync operations.

--msync-ops N
stop after N msync bogo operations completed.

--msync-bytes N
allocate N bytes for the memory mapped file, the default is 256MB. One can specify the size as %
of total available memory or in units of Bytes, KBytes, MBytes and GBytes using the suffix b, k, m
or g.

--munmap N
start N stressors that exercise unmapping of shared non-executable mapped regions of child
processes (Linux only). The unmappings map shared memory regions page by page with a prime sized
stride that creates many temporary mapping holes. One the unmappings are complete the child will
exit and a new one is started. Note that this may trigger segmentation faults in the child
process, these are handled where possible by forcing the child process to call _exit(2).

--munmap-ops N
stop after N page unmappings.

--nanosleep N
start N workers that each run 256 pthreads that call nanosleep with random delays from 1 to 2^18
nanoseconds. This should exercise the high resolution timers and scheduler.

--nanosleep-ops N
stop the nanosleep stressor after N bogo nanosleep operations.

--netdev N
start N workers that exercise various netdevice ioctl commands across all the available network
devices. The ioctls exercised by this stressor are as follows: SIOCGIFCONF, SIOCGIFINDEX,
SIOCGIFNAME, SIOCGIFFLAGS, SIOCGIFADDR, SIOCGIFNETMASK, SIOCGIFMETRIC, SIOCGIFMTU, SIOCGIFHWADDR,
SIOCGIFMAP and SIOCGIFTXQLEN. See netdevice(7) for more details of these ioctl commands.

--netdev-ops N
stop after N netdev bogo operations completed.

--netlink-proc N
start N workers that spawn child processes and monitor fork/exec/exit process events via the proc
netlink connector. Each event received is counted as a bogo op. This stressor can only be run on
Linux and requires CAP_NET_ADMIN capability.

--netlink-proc-ops N
stop the proc netlink connector stressors after N bogo ops.

--netlink-task N
start N workers that collect task statistics via the netlink taskstats interface. This stressor
can only be run on Linux and requires CAP_NET_ADMIN capability.

--netlink-task-ops N
stop the taskstats netlink connector stressors after N bogo ops.

--nice N
start N cpu consuming workers that exercise the available nice levels. Each iteration forks off a
child process that runs through the all the nice levels running a busy loop for 0.1 seconds per
level and then exits.

--nice-ops N
stop after N nice bogo nice loops

--nop N
start N workers that consume cpu cycles issuing no-op instructions. This stressor is available if
the assembler supports the "nop" instruction.

--nop-ops N
stop nop workers after N no-op bogo operations. Each bogo-operation is equivalent to 256 loops of
256 no-op instructions.

--nop-instr INSTR
use alternative nop instruction INSTR. For x86 CPUs INSTR can be one of nop, pause, nop2 (2 byte
nop) through to nop11 (11 byte nop). For ARM CPUs, INSTR can be one of nop or yield. For PPC64
CPUs, INSTR can be one of nop, mdoio, mdoom or yield. For S390 CPUs, INSTR can be one of nop or
nopr. For other processors, INSTR is only nop. The random INSTR option selects a randon mix of the
available nop instructions. If the chosen INSTR generates an SIGILL signal, then the stressor
falls back to the vanilla nop instruction.

--null N
start N workers writing to /dev/null.

--null-ops N
stop null stress workers after N /dev/null bogo write operations.

--numa N
start N workers that migrate stressors and a 4MB memory mapped buffer around all the available
NUMA nodes. This uses migrate_pages(2) to move the stressors and mbind(2) and move_pages(2) to
move the pages of the mapped buffer. After each move, the buffer is written to force activity over
the bus which results cache misses. This test will only run on hardware with NUMA enabled and
more than 1 NUMA node.

--numa-ops N
stop NUMA stress workers after N bogo NUMA operations.

--oom-pipe N
start N workers that create as many pipes as allowed and exercise expanding and shrinking the
pipes from the largest pipe size down to a page size. Data is written into the pipes and read out
again to fill the pipe buffers. With the --aggressive mode enabled the data is not read out when
the pipes are shrunk, causing the kernel to OOM processes aggressively. Running many instances of
this stressor will force kernel to OOM processes due to the many large pipe buffer allocations.

--oom-pipe-ops N
stop after N bogo pipe expand/shrink operations.

--opcode N
start N workers that fork off children that execute randomly generated executable code. This will
generate issues such as illegal instructions, bus errors, segmentation faults, traps, floating
point errors that are handled gracefully by the stressor.

--opcode-ops N
stop after N attempts to execute illegal code.

--opcode-method [ inc | mixed | random | text ]
select the opcode generation method. By default, random bytes are used to generate the executable
code. This option allows one to select one of the three methods:

Method Description
inc use incrementing 32 bit opcode patterns from 0x00000000
to 0xfffffff inclusive.
mixed use a mix of incrementing 32 bit opcode patterns and
random 32 bit opcode patterns that are also inverted,
encoded with gray encoding and bit reversed.
random generate opcodes using random bytes from a mwc random
generator.
text copies random chunks of code from the stress-ng text
segment and randomly flips single bits in a random
choice of 1/8th of the code.

-o N, --open N
start N workers that perform open(2) and then close(2) operations on /dev/zero. The maximum opens
at one time is system defined, so the test will run up to this maximum, or 65536 open file
descriptors, which ever comes first.

--open-ops N
stop the open stress workers after N bogo open operations.

--open-fd
run a child process that scans /proc/$PID/fd and attempts to open the files that the stressor has
opened. This exercises racing open/close operations on the proc interface.

--pageswap N
start N workers that exercise page swap in and swap out. Pages are allocated and paged out using
madvise MADV_PAGEOUT. One the maximum per process number of mmaps are reached or 65536 pages are
allocated the pages are read to page them back in and unmapped in reverse mapping order.

--pageswap-ops N
stop after N page allocation bogo operations.

--pci N
exercise PCI sysfs by running N workers that read data (and mmap/unmap PCI config or PCI resource
files). Linux only. Running as root will allow config and resource mmappings to be read and
exercises PCI I/O mapping.

--pci-ops N
stop pci stress workers after N PCI subdirectory exercising operations.

--personality N
start N workers that attempt to set personality and get all the available personality types
(process execution domain types) via the personality(2) system call. (Linux only).

--personality-ops N
stop personality stress workers after N bogo personality operations.

--physpage N
start N workers that use /proc/self/pagemap and /proc/kpagecount to determine the physical page
and page count of a virtual mapped page and a page that is shared among all the stressors. Linux
only and requires the CAP_SYS_ADMIN capabilities.

--physpage-ops N
stop physpage stress workers after N bogo physical address lookups.

--pidfd N
start N workers that exercise signal sending via the pidfd_send_signal system call. This stressor
creates child processes and checks if they exist and can be stopped, restarted and killed using
the pidfd_send_signal system call.

--pidfd-ops N
stop pidfd stress workers after N child processes have been created, tested and killed with
pidfd_send_signal.

--ping-sock N
start N workers that send small randomized ICMP messages to the localhost across a range of ports
(1024..65535) using a "ping" socket with an AF_INET domain, a SOCK_DGRAM socket type and an
IPPROTO_ICMP protocol.

--ping-sock-ops N
stop the ping-sock stress workers after N ICMP messages are sent.

-p N, --pipe N
start N workers that perform large pipe writes and reads to exercise pipe I/O. This exercises
memory write and reads as well as context switching. Each worker has two processes, a reader and
a writer.

--pipe-ops N
stop pipe stress workers after N bogo pipe write operations.

--pipe-data-size N
specifies the size in bytes of each write to the pipe (range from 4 bytes to 4096 bytes). Setting
a small data size will cause more writes to be buffered in the pipe, hence reducing the context
switch rate between the pipe writer and pipe reader processes. Default size is the page size.

--pipe-size N
specifies the size of the pipe in bytes (for systems that support the F_SETPIPE_SZ fcntl()
command). Setting a small pipe size will cause the pipe to fill and block more frequently, hence
increasing the context switch rate between the pipe writer and the pipe reader processes. Default
size is 512 bytes.

--pipeherd N
start N workers that pass a 64 bit token counter to/from 100 child processes over a shared pipe.
This forces a high context switch rate and can trigger a "thundering herd" of wakeups on processes
that are blocked on pipe waits.

--pipeherd-ops N
stop pipe stress workers after N bogo pipe write operations.

--pipeherd-yield
force a scheduling yield after each write, this increases the context switch rate.

--pkey N
start N workers that change memory protection using a protection key (pkey) and the pkey_mprotect
call (Linux only). This will try to allocate a pkey and use this for the page protection, however,
if this fails then the special pkey -1 will be used (and the kernel will use the normal mprotect
mechanism instead). Various page protection mixes of read/write/exec/none will be cycled through
on randomly chosen pre-allocated pages.

--pkey-ops N
stop after N pkey_mprotect page protection cycles.

-P N, --poll N
start N workers that perform zero timeout polling via the poll(2), ppoll(2), select(2), pselect(2)
and sleep(3) calls. This wastes system and user time doing nothing.

--poll-ops N
stop poll stress workers after N bogo poll operations.

--poll-fds N
specify the number of file descriptors to poll/ppoll/select/pselect on. The maximum number for
select/pselect is limited by FD_SETSIZE and the upper maximum is also limited by the maximum
number of pipe open descriptors allowed.

--prctl N
start N workers that exercise the majority of the prctl(2) system call options. Each batch of
prctl calls is performed inside a new child process to ensure the limit of prctl is contained
inside a new process every time. Some prctl options are architecture specific, however, this
stressor will exercise these even if they are not implemented.

--prctl-ops N
stop prctl workers after N batches of prctl calls

--prefetch N
start N workers that benchmark prefetch and non-prefetch reads of a L3 cache sized buffer. The
buffer is read with loops of 8 × 64 bit reads per iteration. In the prefetch cases, data is
prefetched ahead of the current read position by various sized offsets, from 64 bytes to 8K to
find the best memory read throughput. The stressor reports the non-prefetch read rate and the best
prefetched read rate. It also reports the prefetch offset and an estimate of the amount of time
between the prefetch issue and the actual memory read operation. These statistics will vary from
run-to-run due to system noise and CPU frequency scaling.

--prefetch-ops N
stop prefetch stressors after N benchmark operations

--prefetch-l3-size N
specify the size of the l3 cache

--procfs N
start N workers that read files from /proc and recursively read files from /proc/self (Linux
only).

--procfs-ops N
stop procfs reading after N bogo read operations. Note, since the number of entries may vary
between kernels, this bogo ops metric is probably very misleading.

--pthread N
start N workers that iteratively creates and terminates multiple pthreads (the default is 1024
pthreads per worker). In each iteration, each newly created pthread waits until the worker has
created all the pthreads and then they all terminate together.

--pthread-ops N
stop pthread workers after N bogo pthread create operations.

--pthread-max N
create N pthreads per worker. If the product of the number of pthreads by the number of workers is
greater than the soft limit of allowed pthreads then the maximum is re-adjusted down to the
maximum allowed.

--ptrace N
start N workers that fork and trace system calls of a child process using ptrace(2).

--ptrace-ops N
stop ptracer workers after N bogo system calls are traced.

--pty N
start N workers that repeatedly attempt to open pseudoterminals and perform various pty ioctls
upon the ptys before closing them.

--pty-ops N
stop pty workers after N pty bogo operations.

--pty-max N
try to open a maximum of N pseudoterminals, the default is 65536. The allowed range of this
setting is 8..65536.

-Q, --qsort N
start N workers that sort 32 bit integers using qsort.

--qsort-ops N
stop qsort stress workers after N bogo qsorts.

--qsort-size N
specify number of 32 bit integers to sort, default is 262144 (256 × 1024).

--quota N
start N workers that exercise the Q_GETQUOTA, Q_GETFMT, Q_GETINFO, Q_GETSTATS and Q_SYNC
quotactl(2) commands on all the available mounted block based file systems. Requires CAP_SYS_ADMIN
capability to run.

--quota-ops N
stop quota stress workers after N bogo quotactl operations.

--radixsort N
start N workers that sort random 8 byte strings using radixsort.

--radixsort-ops N
stop radixsort stress workers after N bogo radixsorts.

--radixsort-size N
specify number of strings to sort, default is 262144 (256 × 1024).

--ramfs N
start N workers mounting a memory based file system using ramfs and tmpfs (Linux only). This
alternates between mounting and umounting a ramfs or tmpfs file system using the traditional
mount(2) and umount(2) system call as well as the newer Linux 5.2 fsopen(2), fsmount(2),
fsconfig(2) and move_mount(2) system calls if they are available. The default ram file system size
is 2MB.

--ramfs-ops N
stop after N ramfs mount operations.

--ramfs-size N
set the ramfs size (must be multiples of the page size).

--rawdev N
start N workers that read the underlying raw drive device using direct IO reads. The device (with
minor number 0) that stores the current working directory is the raw device to be read by the
stressor. The read size is exactly the size of the underlying device block size. By default,
this stressor will exercise all the of the rawdev methods (see the --rawdev-method option). This
is a Linux only stressor and requires root privilege to be able to read the raw device.

--rawdev-ops N
stop the rawdev stress workers after N raw device read bogo operations.

--rawdev-method M
Available rawdev stress methods are described as follows:

Method Description
all iterate over all the rawdev stress methods as listed below:
sweep repeatedly read across the raw device from the 0th block to the end
block in steps of the number of blocks on the device / 128 and back
to the start again.
wiggle repeatedly read across the raw device in 128 evenly steps with each
step reading 1024 blocks backwards from each step.
ends repeatedly read the first and last 128 start and end blocks of the
raw device alternating from start of the device to the end of the
device.
random repeatedly read 256 random blocks
burst repeatedly read 256 sequential blocks starting from a random block
on the raw device.

--randlist N
start N workers that creates a list of objects in randomized memory order and traverses the list
setting and reading the objects. This is designed to exerise memory and cache thrashing. Normally
the objects are allocated on the heap, however for objects of page size or larger there is a 1 in
16 chance of objects being allocated using shared anonymous memory mapping to mix up the address
spaces of the allocations to create more TLB thrashing.

--randlist-ops N
stop randlist workers after N list traversals

--randist-compact
Allocate all the list objects using one large heap allocation and divide this up for all the list
objects. This removes the overhead of the heap keeping track of each list object, hence uses less
memory.

--randlist-items N
Allocate N items on the list. By default, 100,000 items are allocated.

--randlist-size N
Allocate each item to be N bytes in size. By default, the size is 64 bytes of data payload plus
the list handling pointer overhead.

--rawsock N
start N workers that send and receive packet data using raw sockets on the localhost. Requires
CAP_NET_RAW to run.

--rawsock-ops N
stop rawsock workers after N packets are received.

--rawpkt N
start N workers that sends and receives ethernet packets using raw packets on the localhost via
the loopback device. Requires CAP_NET_RAW to run.

--rawpkt-ops N
stop rawpkt workers after N packets from the sender process are received.

--rawpkt-port N
start at port P. For N rawpkt worker processes, ports P to (P * 4) - 1 are used. The default
starting port is port 14000.

--rawudp N
start N workers that send and receive UDP packets using raw sockets on the localhost. Requires
CAP_NET_RAW to run.

--rawudp-ops N
stop rawudp workers after N packets are received.

--rawudp-port N
start at port P. For N rawudp worker processes, ports P to (P * 4) - 1 are used. The default
starting port is port 13000.

--rdrand N
start N workers that read a random number from an on-chip random number generator This uses the
rdrand instruction on Intel x86 processors or the darn instruction on Power9 processors.

--rdrand-ops N
stop rdrand stress workers after N bogo rdrand operations (1 bogo op = 2048 random bits
successfully read).

--rdrand-seed
use rdseed instead of rdrand (x86 only).

--readahead N
start N workers that randomly seek and perform 4096 byte read/write I/O operations on a file with
readahead. The default file size is 64 MB. Readaheads and reads are batched into 16 readaheads
and then 16 reads.

--readahead-bytes N
set the size of readahead file, the default is 1 GB. One can specify the size as % of free space
on the file system or in units of Bytes, KBytes, MBytes and GBytes using the suffix b, k, m or g.

--readahead-ops N
stop readahead stress workers after N bogo read operations.

--reboot N
start N workers that exercise the reboot(2) system call. When possible, it will create a process
in a PID namespace and perform a reboot power off command that should shutdown the process. Also,
the stressor exercises invalid reboot magic values and invalid reboots when there are insufficient
privileges that will not actually reboot the system.

--reboot-ops N
stop the reboot stress workers after N bogo reboot cycles.

--remap N
start N workers that map 512 pages and re-order these pages using the deprecated system call
remap_file_pages(2). Several page re-orderings are exercised: forward, reverse, random and many
pages to 1 page.

--remap-ops N
stop after N remapping bogo operations.

-R N, --rename N
start N workers that each create a file and then repeatedly rename it.

--rename-ops N
stop rename stress workers after N bogo rename operations.

--resched N
start N workers that exercise process rescheduling. Each stressor spawns a child process for each
of the positive nice levels and iterates over the nice levels from 0 to the lowest priority level
(highest nice value). For each of the nice levels 1024 iterations over 3 non-real time scheduling
polices SCHED_OTHER, SCHED_BATCH and SCHED_IDLE are set and a sched_yield occurs to force heavy
rescheduling activity. When the -v verbose option is used the distribution of the number of
yields across the nice levels is printed for the first stressor out of the N stressors.

--resched-ops N
stop after N rescheduling sched_yield calls.

--resources N
start N workers that consume various system resources. Each worker will spawn 1024 child processes
that iterate 1024 times consuming shared memory, heap, stack, temporary files and various file
descriptors (eventfds, memoryfds, userfaultfds, pipes and sockets).

--resources-ops N
stop after N resource child forks.

--revio N
start N workers continually writing in reverse position order to temporary files. The default mode
is to stress test reverse position ordered writes with randomly sized sparse holes between each
write. With the --aggressive option enabled without any --revio-opts options the revio stressor
will work through all the --revio-opt options one by one to cover a range of I/O options.

--revio-bytes N
write N bytes for each revio process, the default is 1 GB. One can specify the size as % of free
space on the file system or in units of Bytes, KBytes, MBytes and GBytes using the suffix b, k, m
or g.

--revio-opts list
specify various stress test options as a comma separated list. Options are the same as --hdd-opts
but without the iovec option.

--revio-ops N
stop revio stress workers after N bogo operations.

--revio-write-size N
specify size of each write in bytes. Size can be from 1 byte to 4MB.

--rlimit N
start N workers that exceed CPU and file size resource imits, generating SIGXCPU and SIGXFSZ
signals.

--rlimit-ops N
stop after N bogo resource limited SIGXCPU and SIGXFSZ signals have been caught.

--rmap N
start N workers that exercise the VM reverse-mapping. This creates 16 processes per worker that
write/read multiple file-backed memory mappings. There are 64 lots of 4 page mappings made onto
the file, with each mapping overlapping the previous by 3 pages and at least 1 page of non-mapped
memory between each of the mappings. Data is synchronously msync'd to the file 1 in every 256
iterations in a random manner.

--rmap-ops N
stop after N bogo rmap memory writes/reads.

--rseq N
start N workers that exercise restartable sequences via the rseq(2) system call. This loops over
a long duration critical section that is likely to be interrupted. A rseq abort handler keeps
count of the number of interruptions and a SIGSEV handler also tracks any failed rseq aborts that
can occur if there is a mistmatch in a rseq check signature. Linux only.

--rseq-ops N
stop after N bogo rseq operations. Each bogo rseq operation is equivalent to 10000 iterations over
a long duration rseq handled critical section.

--rtc N
start N workers that exercise the real time clock (RTC) interfaces via /dev/rtc and
/sys/class/rtc/rtc0. No destructive writes (modifications) are performed on the RTC. This is a
Linux only stressor.

--rtc-ops N
stop after N bogo RTC interface accesses.

--schedpolicy N
start N workers that work set the worker to various available scheduling policies out of
SCHED_OTHER, SCHED_BATCH, SCHED_IDLE, SCHED_FIFO, SCHED_RR and SCHED_DEADLINE. For the real time
scheduling policies a random sched priority is selected between the minimum and maximum scheduling
priority settings.

--schedpolicy-ops N
stop after N bogo scheduling policy changes.

--sctp N
start N workers that perform network sctp stress activity using the Stream Control Transmission
Protocol (SCTP). This involves client/server processes performing rapid connect, send/receives
and disconnects on the local host.

--sctp-domain D
specify the domain to use, the default is ipv4. Currently ipv4 and ipv6 are supported.

--sctp-ops N
stop sctp workers after N bogo operations.

--sctp-port P
start at sctp port P. For N sctp worker processes, ports P to (P * 4) - 1 are used for ipv4, ipv6
domains and ports P to P - 1 are used for the unix domain.

--seal N
start N workers that exercise the fcntl(2) SEAL commands on a small anonymous file created using
memfd_create(2). After each SEAL command is issued the stressor also sanity checks if the seal
operation has sealed the file correctly. (Linux only).

--seal-ops N
stop after N bogo seal operations.

--seccomp N
start N workers that exercise Secure Computing system call filtering. Each worker creates child
processes that write a short message to /dev/null and then exits. 2% of the child processes have a
seccomp filter that disallows the write system call and hence it is killed by seccomp with a
SIGSYS. Note that this stressor can generate many audit log messages each time the child is
killed. Requires CAP_SYS_ADMIN to run.

--seccomp-ops N
stop seccomp stress workers after N seccomp filter tests.

--secretmem N
start N workers that mmap pages using file mapping off a memfd_secret file descriptor. Each stress
loop iteration will expand the mappable region by 3 pages using ftruncate and mmap and touches the
pages. The pages are then fragmented by unmapping the middle page and then umapping the first and
last pages. This tries to force page fragmentation and also trigger out of memory (OOM) kills of
the stressor when the secret memory is exhausted. Note this is a Linux 5.11+ only stressor and
the kernel needs to be booted with "secretmem=" option to allocate a secret memory reservation.

--secretmem-ops N
stop secretmem stress workers after N stress loop iterations.

--seek N
start N workers that randomly seeks and performs 512 byte read/write I/O operations on a file. The
default file size is 16 GB.

--seek-ops N
stop seek stress workers after N bogo seek operations.

--seek-punch
punch randomly located 8K holes into the file to cause more extents to force a more demanding seek
stressor, (Linux only).

--seek-size N
specify the size of the file in bytes. Small file sizes allow the I/O to occur in the cache,
causing greater CPU load. Large file sizes force more I/O operations to drive causing more wait
time and more I/O on the drive. One can specify the size in units of Bytes, KBytes, MBytes and
GBytes using the suffix b, k, m or g.

--sem N
start N workers that perform POSIX semaphore wait and post operations. By default, a parent and 4
children are started per worker to provide some contention on the semaphore. This stresses fast
semaphore operations and produces rapid context switching.

--sem-ops N
stop semaphore stress workers after N bogo semaphore operations.

--sem-procs N
start N child workers per worker to provide contention on the semaphore, the default is 4 and a
maximum of 64 are allowed.

--sem-sysv N
start N workers that perform System V semaphore wait and post operations. By default, a parent and
4 children are started per worker to provide some contention on the semaphore. This stresses fast
semaphore operations and produces rapid context switching.

--sem-sysv-ops N
stop semaphore stress workers after N bogo System V semaphore operations.

--sem-sysv-procs N
start N child processes per worker to provide contention on the System V semaphore, the default is
4 and a maximum of 64 are allowed.

--sendfile N
start N workers that send an empty file to /dev/null. This operation spends nearly all the time in
the kernel. The default sendfile size is 4MB. The sendfile options are for Linux only.

--sendfile-ops N
stop sendfile workers after N sendfile bogo operations.

--sendfile-size S
specify the size to be copied with each sendfile call. The default size is 4MB. One can specify
the size in units of Bytes, KBytes, MBytes and GBytes using the suffix b, k, m or g.

--session N
start N workers that create child and grandchild processes that set and get their session ids. 25%
of the grandchild processes are not waited for by the child to create orphaned sessions that need
to be reaped by init.

--session-ops N
stop session workers after N child processes are spawned and reaped.

--set N
start N workers that call system calls that try to set data in the kernel, currently these are:
setgid, sethostname, setpgid, setpgrp, setuid, setgroups, setreuid, setregid, setresuid, setresgid
and setrlimit. Some of these system calls are OS specific.

--set-ops N
stop set workers after N bogo set operations.

--shellsort N
start N workers that sort 32 bit integers using shellsort.

--shellsort-ops N
stop shellsort stress workers after N bogo shellsorts.

--shellsort-size N
specify number of 32 bit integers to sort, default is 262144 (256 × 1024).

--shm N
start N workers that open and allocate shared memory objects using the POSIX shared memory
interfaces. By default, the test will repeatedly create and destroy 32 shared memory objects,
each of which is 8MB in size.

--shm-ops N
stop after N POSIX shared memory create and destroy bogo operations are complete.

--shm-bytes N
specify the size of the POSIX shared memory objects to be created. One can specify the size as %
of total available memory or in units of Bytes, KBytes, MBytes and GBytes using the suffix b, k, m
or g.

--shm-objs N
specify the number of shared memory objects to be created.

--shm-sysv N
start N workers that allocate shared memory using the System V shared memory interface. By
default, the test will repeatedly create and destroy 8 shared memory segments, each of which is
8MB in size.

--shm-sysv-ops N
stop after N shared memory create and destroy bogo operations are complete.

--shm-sysv-bytes N
specify the size of the shared memory segment to be created. One can specify the size as % of
total available memory or in units of Bytes, KBytes, MBytes and GBytes using the suffix b, k, m or
g.

--shm-sysv-segs N
specify the number of shared memory segments to be created. The default is 8 segments.

--sigabrt N
start N workers that create children that are killed by SIGABRT signals or by calling abort(3).

--sigabrt-ops N
stop the sigabrt workers after N SIGABRT signals are successfully handled.

--sigchld N
start N workers that create children to generate SIGCHLD signals. This exercises children that
exit (CLD_EXITED), get killed (CLD_KILLED), get stopped (CLD_STOPPED) or continued
(CLD_CONTINUED).

--sigchld-ops N
stop the sigchld workers after N SIGCHLD signals are successfully handled.

--sigfd N
start N workers that generate SIGRT signals and are handled by reads by a child process using a
file descriptor set up using signalfd(2). (Linux only). This will generate a heavy context switch
load when all CPUs are fully loaded.

--sigfd-ops
stop sigfd workers after N bogo SIGUSR1 signals are sent.

--sigfpe N
start N workers that rapidly cause division by zero SIGFPE faults.

--sigfpe-ops N
stop sigfpe stress workers after N bogo SIGFPE faults.

--sigio N
start N workers that read data from a child process via a pipe and generate SIGIO signals. This
exercises asynchronous I/O via SIGIO.

--sigio-ops N
stop sigio stress workers after handling N SIGIO signals.

--signal N
start N workers that exercise the signal system call three different signal handlers, SIG_IGN
(ignore), a SIGCHLD handler and SIG_DFL (default action). For the SIGCHLD handler, the stressor
sends itself a SIGCHLD signal and checks if it has been handled. For other handlers, the stressor
checks that the SIGCHLD handler has not been called. This stress test calls the signal system
call directly when possible and will try to avoid the C library attempt to replace signal with the
more modern sigaction system call.

--signal-ops N
stop signal stress workers after N rounds of signal handler setting.

--signest N
start N workers that exercise nested signal handling. A signal is raised and inside the signal
handler a different signal is raised, working through a list of signals to exercise. An
alternative signal stack is used that is large enough to handle all the nested signal calls. The
-v option will log the approximate size of the stack required and the average stack size per
nested call.

--signest-ops N
stop after handling N nested signals.

--sigpending N
start N workers that check if SIGUSR1 signals are pending. This stressor masks SIGUSR1, generates
a SIGUSR1 signal and uses sigpending(2) to see if the signal is pending. Then it unmasks the
signal and checks if the signal is no longer pending.

--sigpending-ops N
stop sigpending stress workers after N bogo sigpending pending/unpending checks.

--sigpipe N
start N workers that repeatedly spawn off child process that exits before a parent can complete a
pipe write, causing a SIGPIPE signal. The child process is either spawned using clone(2) if it is
available or use the slower fork(2) instead.

--sigpipe-ops N
stop N workers after N SIGPIPE signals have been caught and handled.

--sigq N
start N workers that rapidly send SIGUSR1 signals using sigqueue(3) to child processes that wait
for the signal via sigwaitinfo(2).

--sigq-ops N
stop sigq stress workers after N bogo signal send operations.

--sigrt N
start N workers that each create child processes to handle SIGRTMIN to SIGRMAX real time signals.
The parent sends each child process a RT signal via siqueue(2) and the child process waits for
this via sigwaitinfo(2). When the child receives the signal it then sends a RT signal to one of
the other child processes also via sigqueue(2).

--sigrt-ops N
stop sigrt stress workers after N bogo sigqueue signal send operations.

--sigsegv N
start N workers that rapidly create and catch segmentation faults.

--sigsegv-ops N
stop sigsegv stress workers after N bogo segmentation faults.

--sigsuspend N
start N workers that each spawn off 4 child processes that wait for a SIGUSR1 signal from the
parent using sigsuspend(2). The parent sends SIGUSR1 signals to each child in rapid succession.
Each sigsuspend wakeup is counted as one bogo operation.

--sigsuspend-ops N
stop sigsuspend stress workers after N bogo sigsuspend wakeups.

--sigtrap N
start N workers that exercise the SIGTRAP signal. For systems that support SIGTRAP, the signal is
generated using raise(SIGTRAP). Only x86 Linux systems the SIGTRAP is also generated by an int 3
instruction.

--sigtrap-ops N
stop sigtrap stress workers after N SIGTRAPs have been handled.

--skiplist N
start N workers that store and then search for integers using a skiplist. By default, 65536
integers are added and searched. This is a useful method to exercise random access of memory and
processor cache.

--skiplist-ops N
stop the skiplist worker after N skiplist store and search cycles are completed.

--skiplist-size N
specify the size (number of integers) to store and search in the skiplist. Size can be from 1K to
4M.

--sleep N
start N workers that spawn off multiple threads that each perform multiple sleeps of ranges 1us to
0.1s. This creates multiple context switches and timer interrupts.

--sleep-ops N
stop after N sleep bogo operations.

--sleep-max P
start P threads per worker. The default is 1024, the maximum allowed is 30000.

--smi N
start N workers that attempt to generate system management interrupts (SMIs) into the x86 ring -2
system management mode (SMM) by exercising the advanced power management (APM) port 0xb2. This
requires the --pathological option and root privilege and is only implemented on x86 Linux
platforms. This probably does not work in a virtualized environment. The stressor will attempt to
determine the time stolen by SMIs with some naive benchmarking.

--smi-ops N
stop after N attempts to trigger the SMI.

-S N, --sock N
start N workers that perform various socket stress activity. This involves a pair of client/server
processes performing rapid connect, send and receives and disconnects on the local host.

--sock-domain D
specify the domain to use, the default is ipv4. Currently ipv4, ipv6 and unix are supported.

--sock-nodelay
This disables the TCP Nagle algorithm, so data segments are always sent as soon as possible. This
stops data from being buffered before being transmitted, hence resulting in poorer network
utilisation and more context switches between the sender and receiver.

--sock-port P
start at socket port P. For N socket worker processes, ports P to P - 1 are used.

--sock-protocol P
Use the specified protocol P, default is tcp. Options are tcp and mptcp (if supported by the
operating system).

--sock-ops N
stop socket stress workers after N bogo operations.

--sock-opts [ random | send | sendmsg | sendmmsg ]
by default, messages are sent using send(2). This option allows one to specify the sending method
using send(2), sendmsg(2), sendmmsg(2) or a random selection of one of thse 3 on each iteration.
Note that sendmmsg is only available for Linux systems that support this system call.

--sock-type [ stream | seqpacket ]
specify the socket type to use. The default type is stream. seqpacket currently only works for the
unix socket domain.

--sock-zerocopy
enable zerocopy for send and recv calls if the MSG_ZEROCOPY is supported.

--sockabuse N
start N workers that abuse a socket file descriptor with various file based system that don't
normally act on sockets. The kernel should handle these illegal and unexpected calls gracefully.

--sockabuse-ops N
stop after N iterations of the socket abusing stressor loop.

--sockdiag N
start N workers that exercise the Linux sock_diag netlink socket diagnostics (Linux only). This
currently requests diagnostics using UDIAG_SHOW_NAME, UDIAG_SHOW_VFS, UDIAG_SHOW_PEER,
UDIAG_SHOW_ICONS, UDIAG_SHOW_RQLEN and UDIAG_SHOW_MEMINFO for the AF_UNIX family of socket
connections.

--sockdiag-ops N
stop after receiving N sock_diag diagnostic messages.

--sockfd N
start N workers that pass file descriptors over a UNIX domain socket using the CMSG(3) ancillary
data mechanism. For each worker, pair of client/server processes are created, the server opens as
many file descriptors on /dev/null as possible and passing these over the socket to a client that
reads these from the CMSG data and immediately closes the files.

--sockfd-ops N
stop sockfd stress workers after N bogo operations.

--sockfd-port P
start at socket port P. For N socket worker processes, ports P to P - 1 are used.

--sockmany N
start N workers that use a client process to attempt to open as many as 100000 TCP/IP socket
connections to a server on port 10000.

--sockmany-ops N
stop after N connections.

--sockpair N
start N workers that perform socket pair I/O read/writes. This involves a pair of client/server
processes performing randomly sized socket I/O operations.

--sockpair-ops N
stop socket pair stress workers after N bogo operations.

--softlockup N
start N workers that flip between with the "real-time" SCHED_FIO and SCHED_RR scheduling policies
at the highest priority to force softlockups. This can only be run with CAP_SYS_NICE capability
and for best results the number of stressors should be at least the number of online CPUs. Once
running, this is practically impossible to stop and it will force softlockup issues and may
trigger watchdog timeout reboots.

--softlockup-ops N
stop softlockup stress workers after N bogo scheduler policy changes.

--sparsematrix N
start N workers that exercise 3 different sparse matrix implementations based on hashing, Judy
array (for 64 bit systems), 2-d circular linked-lists, memory mapped 2-d matrix (non-sparse),
quick hashing (on preallocated nodes) and red-black tree. The sparse matrix is populated with
values, random values potentially non-existing values are read, known existing values are read and
known existing values are marked as zero. This default 500 x 500 sparse matrix is used and 5000
items are put into the sparse matrix making it 2% utilized.

--sparsematrix-ops N
stop after N sparsematrix test iterations.

--sparsematrix-items N
populate the sparse matrix with N items. If N is greater than the number of elements in the sparse
matrix than N will be capped to create at 100% full sparse matrix.

--sparsematrix-size N
use a N × N sized sparse matrix

Method Description
all exercise with all the sparsematrix stressor methods (see
below):
hash use a hash table and allocate nodes on the heap each
unique value at a (x, y) matrix position.
judy use a Judy array with a unique 1-to-1 mapping of (x, y)
matrix position into the array.
list use a circular linked-list for sparse y positions each
with circular linked-lists for sparse x positions for
the (x, y) matrix coordinates.
mmap a non-sparse mmap the entire 2-d matrix space. Only (x,
y) matrix positions that are referenced will get
physically mapped. Note that large sparse matrices
cannot be mmap'd due to lack of virtual address
limitations, and too many referenced pages can trigger
the out of memory killer on Linux.
qhash use a hash table with pre-allocated nodes for each
unique value. This is a quick hash table implementation,
nodes are not allocated each time with calloc and are
allocated from a pre-allocated pool leading to quicker
hash table performance than the hash method.

--spawn N
start N workers continually spawn children using posix_spawn(3) that exec stress-ng and then exit
almost immediately. Currently Linux only.

--spawn-ops N
stop spawn stress workers after N bogo spawns.

--splice N
move data from /dev/zero to /dev/null through a pipe without any copying between kernel address
space and user address space using splice(2). This is only available for Linux.

--splice-ops N
stop after N bogo splice operations.

--splice-bytes N
transfer N bytes per splice call, the default is 64K. One can specify the size as % of total
available memory or in units of Bytes, KBytes, MBytes and GBytes using the suffix b, k, m or g.

--stack N
start N workers that rapidly cause and catch stack overflows by use of large recursive stack
allocations. Much like the brk stressor, this can eat up pages rapidly and may trigger the kernel
OOM killer on the process, however, the killed stressor is respawned again by a monitoring parent
process.

--stack-fill
the default action is to touch the lowest page on each stack allocation. This option touches all
the pages by filling the new stack allocation with zeros which forces physical pages to be
allocated and hence is more aggressive.

--stack-mlock
attempt to mlock stack pages into memory prohibiting them from being paged out. This is a no-op
if mlock(2) is not available.

--stack-ops N
stop stack stress workers after N bogo stack overflows.

--stackmmap N
start N workers that use a 2MB stack that is memory mapped onto a temporary file. A recursive
function works down the stack and flushes dirty stack pages back to the memory mapped file using
msync(2) until the end of the stack is reached (stack overflow). This exercises dirty page and
stack exception handling.

--stackmmap-ops N
stop workers after N stack overflows have occurred.

--str N
start N workers that exercise various libc string functions on random strings.

--str-method strfunc
select a specific libc string function to stress. Available string functions to stress are: all,
index, rindex, strcasecmp, strcat, strchr, strcoll, strcmp, strcpy, strlen, strncasecmp, strncat,
strncmp, strrchr and strxfrm. See string(3) for more information on these string functions. The
'all' method is the default and will exercise all the string methods.

--str-ops N
stop after N bogo string operations.

--stream N
start N workers exercising a memory bandwidth stressor loosely based on the STREAM "Sustainable
Memory Bandwidth in High Performance Computers" benchmarking tool by John D. McCalpin, Ph.D. This
stressor allocates buffers that are at least 4 times the size of the CPU L2 cache and continually
performs rounds of following computations on large arrays of double precision floating point
numbers:

Operation Description
copy c[i] = a[i]
scale b[i] = scalar * c[i]
add c[i] = a[i] + b[i]
triad a[i] = b[i] + (c[i] * scalar)

Since this is loosely based on a variant of the STREAM benchmark code, DO NOT submit results based
on this as it is intended to in stress-ng just to stress memory and compute and NOT intended for
STREAM accurate tuned or non-tuned benchmarking whatsoever. Use the official STREAM benchmarking
tool if you desire accurate and standardised STREAM benchmarks.

--stream-ops N
stop after N stream bogo operations, where a bogo operation is one round of copy, scale, add and
triad operations.

--stream-index N
specify number of stream indices used to index into the data arrays a, b and c. This adds
indirection into the data lookup by using randomly shuffled indexing into the three data arrays.
Level 0 (no indexing) is the default, and 3 is where all 3 arrays are indexed via 3 different
randomly shuffled indexes. The higher the index setting the more impact this has on L1, L2 and L3
caching and hence forces higher memory read/write latencies.

--stream-l3-size N
Specify the CPU Level 3 cache size in bytes. One can specify the size in units of Bytes, KBytes,
MBytes and GBytes using the suffix b, k, m or g. If the L3 cache size is not provided, then
stress-ng will attempt to determine the cache size, and failing this, will default the size to
4MB.

--stream-madvise [ hugepage | nohugepage | normal ]
Specify the madvise options used on the memory mapped buffer used in the stream stressor. Non-
linux systems will only have the 'normal' madvise advice. The default is 'normal'.

--swap N
start N workers that add and remove small randomly sizes swap partitions (Linux only). Note that
if too many swap partitions are added then the stressors may exit with exit code 3 (not enough
resources). Requires CAP_SYS_ADMIN to run.

--swap-ops N
stop the swap workers after N swapon/swapoff iterations.

-s N, --switch N
start N workers that force context switching between two mutually blocking/unblocking tied
processes. By default message passing over a pipe is used, but different methods are available.

--switch-ops N
stop context switching workers after N bogo operations.

--switch-freq F
run the context switching at the frequency of F context switches per second. Note that the
specified switch rate may not be achieved because of CPU speed and memory bandwidth limitations.

--switch-method [ mq | pipe | sem-sysv ]
select the preferred context switch block/run synchronization method, these are as follows:

Method Description
mq use posix message queue with a 1 item size. Messages are passed
between a sender and receiver process.
pipe single character messages are passed down a single character sized
pipe between a sender and receiver process.
sem-sysv a SYSV semaphore is used to block/run two processes.

--symlink N
start N workers creating and removing symbolic links.

--symlink-ops N
stop symlink stress workers after N bogo operations.

--sync-file N
start N workers that perform a range of data syncs across a file using sync_file_range(2). Three
mixes of syncs are performed, from start to the end of the file, from end of the file to the
start, and a random mix. A random selection of valid sync types are used, covering the
SYNC_FILE_RANGE_WAIT_BEFORE, SYNC_FILE_RANGE_WRITE and SYNC_FILE_RANGE_WAIT_AFTER flag bits.

--sync-file-ops N
stop sync-file workers after N bogo sync operations.

--sync-file-bytes N
specify the size of the file to be sync'd. One can specify the size as % of free space on the file
system in units of Bytes, KBytes, MBytes and GBytes using the suffix b, k, m or g.

--syncload N
start N workers that produce sporadic short lived loads synchronized across N stressor processes.
By default repeated cycles of 125ms busy load followed by 62.5ms sleep occur across all the
workers in step to create bursts of load to exercise C state transitions and CPU frequency
scaling. The busy load and sleeps have +/-10% jitter added to try exercising scheduling patterns.

--syncload-ops N
stop syncload workers after N load/sleep cycles.

--syncload-msbusy M
specify the busy load duration in milliseconds.

--syncload-mssleep M
specify the sleep duration in milliseconds.

--sysbadaddr N
start N workers that pass bad addresses to system calls to exercise bad address and fault
handling. The addresses used are null pointers, read only pages, write only pages, unmapped
addresses, text only pages, unaligned addresses and top of memory addresses.

--sysbadaddr-ops N
stop the sysbadaddr stressors after N bogo system calls.

--sysinfo N
start N workers that continually read system and process specific information. This reads the
process user and system times using the times(2) system call. For Linux systems, it also reads
overall system statistics using the sysinfo(2) system call and also the file system statistics for
all mounted file systems using statfs(2).

--sysinfo-ops N
stop the sysinfo workers after N bogo operations.

--sysinval N
start N workers that exercise system calls in random order with permutations of invalid arguments
to force kernel error handling checks. The stress test autodetects system calls that cause
processes to crash or exit prematurely and will blocklist these after several repeated breakages.
System call arguments that cause system calls to work successfully are also detected an
blocklisted too. Linux only.

--sysinval-ops N
stop sysinval workers after N system call attempts.

--sysfs N
start N workers that recursively read files from /sys (Linux only). This may cause specific
kernel drivers to emit messages into the kernel log.

--sys-ops N
stop sysfs reading after N bogo read operations. Note, since the number of entries may vary
between kernels, this bogo ops metric is probably very misleading.

--tee N
move data from a writer process to a reader process through pipes and to /dev/null without any
copying between kernel address space and user address space using tee(2). This is only available
for Linux.

--tee-ops N
stop after N bogo tee operations.

-T N, --timer N
start N workers creating timer events at a default rate of 1 MHz (Linux only); this can create a
many thousands of timer clock interrupts. Each timer event is caught by a signal handler and
counted as a bogo timer op.

--timer-ops N
stop timer stress workers after N bogo timer events (Linux only).

--timer-freq F
run timers at F Hz; range from 1 to 1000000000 Hz (Linux only). By selecting an appropriate
frequency stress-ng can generate hundreds of thousands of interrupts per second. Note: it is also
worth using --timer-slack 0 for high frequencies to stop the kernel from coalescing timer events.

--timer-rand
select a timer frequency based around the timer frequency +/- 12.5% random jitter. This tries to
force more variability in the timer interval to make the scheduling less predictable.

--timerfd N
start N workers creating timerfd events at a default rate of 1 MHz (Linux only); this can create a
many thousands of timer clock events. Timer events are waited for on the timer file descriptor
using select(2) and then read and counted as a bogo timerfd op.

--timerfd-ops N
stop timerfd stress workers after N bogo timerfd events (Linux only).

--timerfs-fds N
try to use a maximum of N timerfd file descriptors per stressor.

--timerfd-freq F
run timers at F Hz; range from 1 to 1000000000 Hz (Linux only). By selecting an appropriate
frequency stress-ng can generate hundreds of thousands of interrupts per second.

--timerfd-rand
select a timerfd frequency based around the timer frequency +/- 12.5% random jitter. This tries to
force more variability in the timer interval to make the scheduling less predictable.

--tlb-shootdown N
start N workers that force Translation Lookaside Buffer (TLB) shootdowns. This is achieved by
creating up to 16 child processes that all share a region of memory and these processes are shared
amongst the available CPUs. The processes adjust the page mapping settings causing TLBs to be
force flushed on the other processors, causing the TLB shootdowns.

--tlb-shootdown-ops N
stop after N bogo TLB shootdown operations are completed.

--tmpfs N
start N workers that create a temporary file on an available tmpfs file system and perform various
file based mmap operations upon it.

--tmpfs-ops N
stop tmpfs stressors after N bogo mmap operations.

--tmpfs-mmap-async
enable file based memory mapping and use asynchronous msync'ing on each page, see
--tmpfs-mmap-file.

--tmpfs-mmap-file
enable tmpfs file based memory mapping and by default use synchronous msync'ing on each page.

--tree N
start N workers that exercise tree data structures. The default is to add, find and remove 250,000
64 bit integers into AVL (avl), Red-Black (rb), Splay (splay) and binary trees. The intention of
this stressor is to exercise memory and cache with the various tree operations.

--tree-ops N
stop tree stressors after N bogo ops. A bogo op covers the addition, finding and removing all the
items into the tree(s).

--tree-size N
specify the size of the tree, where N is the number of 64 bit integers to be added into the tree.

--tree-method [ all | avl | binary | rb | splay ]
specify the tree to be used. By default, both the rb ad splay trees are used (the 'all' option).

--tsc N
start N workers that read the Time Stamp Counter (TSC) 256 times per loop iteration (bogo
operation). This exercises the tsc instruction for x86, the mftb instruction for ppc64 and the
rdcycle instruction for RISC-V.

--tsc-ops N
stop the tsc workers after N bogo operations are completed.

--tsearch N
start N workers that insert, search and delete 32 bit integers on a binary tree using tsearch(3),
tfind(3) and tdelete(3). By default, there are 65536 randomized integers used in the tree. This
is a useful method to exercise random access of memory and processor cache.

--tsearch-ops N
stop the tsearch workers after N bogo tree operations are completed.

--tsearch-size N
specify the size (number of 32 bit integers) in the array to tsearch. Size can be from 1K to 4M.

--tun N
start N workers that create a network tunnel device and sends and receives packets over the tunnel
using UDP and then destroys it. A new random 192.168.*.* IPv4 address is used each time a tunnel
is created.

--tun-ops N
stop after N iterations of creating/sending/receiving/destroying a tunnel.

--tun-tap
use network tap device using level 2 frames (bridging) rather than a tun device for level 3 raw
packets (tunnelling).

--udp N
start N workers that transmit data using UDP. This involves a pair of client/server processes
performing rapid connect, send and receives and disconnects on the local host.

--udp-domain D
specify the domain to use, the default is ipv4. Currently ipv4, ipv6 and unix are supported.

--udp-lite
use the UDP-Lite (RFC 3828) protocol (only for ipv4 and ipv6 domains).

--udp-ops N
stop udp stress workers after N bogo operations.

--udp-port P
start at port P. For N udp worker processes, ports P to P - 1 are used. By default, ports 7000
upwards are used.

--udp-flood N
start N workers that attempt to flood the host with UDP packets to random ports. The IP address of
the packets are currently not spoofed. This is only available on systems that support AF_PACKET.

--udp-flood-domain D
specify the domain to use, the default is ipv4. Currently ipv4 and ipv6 are supported.

--udp-flood-ops N
stop udp-flood stress workers after N bogo operations.

--unshare N
start N workers that each fork off 32 child processes, each of which exercises the unshare(2)
system call by disassociating parts of the process execution context. (Linux only).

--unshare-ops N
stop after N bogo unshare operations.

--uprobe N
start N workers that trace the entry to libc function getpid() using the Linux uprobe kernel
tracing mechanism. This requires CAP_SYS_ADMIN capabilities and a modern Linux uprobe capable
kernel.

--uprobe-ops N
stop uprobe tracing after N trace events of the function that is being traced.

-u N, --urandom N
start N workers reading /dev/urandom (Linux only). This will load the kernel random number source.

--urandom-ops N
stop urandom stress workers after N urandom bogo read operations (Linux only).

--userfaultfd N
start N workers that generate write page faults on a small anonymously mapped memory region and
handle these faults using the user space fault handling via the userfaultfd mechanism. This will
generate a large quantity of major page faults and also context switches during the handling of
the page faults. (Linux only).

--userfaultfd-ops N
stop userfaultfd stress workers after N page faults.

--userfaultfd-bytes N
mmap N bytes per userfaultfd worker to page fault on, the default is 16MB. One can specify the
size as % of total available memory or in units of Bytes, KBytes, MBytes and GBytes using the
suffix b, k, m or g.

--usersyscall N
start N workers that exercise the Linux prctl userspace system call mechanism. A userspace system
call is handled by a SIGSYS signal handler and exercised with the system call disabled (ENOSYS)
and enabled (via SIGSYS) using prctl PR_SET_SYSCALL_USER_DISPATCH.

--usersyscall-ops N
stop after N successful userspace syscalls via a SIGSYS signal handler.

--utime N
start N workers updating file timestamps. This is mainly CPU bound when the default is used as the
system flushes metadata changes only periodically.

--utime-ops N
stop utime stress workers after N utime bogo operations.

--utime-fsync
force metadata changes on each file timestamp update to be flushed to disk. This forces the test
to become I/O bound and will result in many dirty metadata writes.

--vdso N
start N workers that repeatedly call each of the system call functions in the vDSO (virtual
dynamic shared object). The vDSO is a shared library that the kernel maps into the address space
of all user-space applications to allow fast access to kernel data to some system calls without
the need of performing an expensive system call.

--vdso-ops N
stop after N vDSO functions calls.

--vdso-func F
Instead of calling all the vDSO functions, just call the vDSO function F. The functions depend on
the kernel being used, but are typically clock_gettime, getcpu, gettimeofday and time.

--vecmath N
start N workers that perform various unsigned integer math operations on various 128 bit vectors.
A mix of vector math operations are performed on the following vectors: 16 × 8 bits, 8 × 16 bits,
4 × 32 bits, 2 × 64 bits. The metrics produced by this mix depend on the processor architecture
and the vector math optimisations produced by the compiler.

--vecmath-ops N
stop after N bogo vector integer math operations.

--vecwide N
start N workers that perform various 8 bit math operations on vectors of 4, 8, 16, 32, 64, 128,
256, 512, 1024 and 2048 bytes. With the -v option the relative compute performance vs the expected
compute performance based on total run time is shown for the first vecwide worker. The vecwide
stressor exercises various processor vector instruction mixes and how well the compiler can map
the vector operations to the target instruction set.

--vecwide-ops N
stop after N bogo vector operations (2048 iterations of a mix of vector instruction operations).

--verity N
start N workers that exercise read-only file based authenticy protection using the verity ioctls
FS_IOC_ENABLE_VERITY and FS_IOC_MEASURE_VERITY. This requires file systems with verity support
(currently ext4 and f2fs on Linux) with the verity feature enabled. The test attempts to creates a
small file with multiple small extents and enables verity on the file and verifies it. It also
checks to see if the file has verity enabled with the FS_VERITY_FL bit set on the file flags.

--verity-ops N
stop the verity workers after N file create, enable verity, check verity and unlink cycles.

--vfork N
start N workers continually vforking children that immediately exit.

--vfork-ops N
stop vfork stress workers after N bogo operations.

--vfork-max P
create P processes and then wait for them to exit per iteration. The default is just 1; higher
values will create many temporary zombie processes that are waiting to be reaped. One can
potentially fill up the process table using high values for --vfork-max and --vfork.

--vfork-vm
enable detrimental performance virtual memory advice using madvise on all pages of the vforked
process. Where possible this will try to set every page in the new process with using madvise
MADV_MERGEABLE, MADV_WILLNEED, MADV_HUGEPAGE and MADV_RANDOM flags. Linux only.

--vforkmany N
start N workers that spawn off a chain of vfork children until the process table fills up and/or
vfork fails. vfork can rapidly create child processes and the parent process has to wait until
the child dies, so this stressor rapidly fills up the process table.

--vforkmany-ops N
stop vforkmany stressors after N vforks have been made.

--vforkmany-vm
enable detrimental performance virtual memory advice using madvise on all pages of the vforked
process. Where possible this will try to set every page in the new process with using madvise
MADV_MERGEABLE, MADV_WILLNEED, MADV_HUGEPAGE and MADV_RANDOM flags. Linux only.

-m N, --vm N
start N workers continuously calling mmap(2)/munmap(2) and writing to the allocated memory. Note
that this can cause systems to trip the kernel OOM killer on Linux systems if not enough physical
memory and swap is not available.

--vm-bytes N
mmap N bytes per vm worker, the default is 256MB. One can specify the size as % of total available
memory or in units of Bytes, KBytes, MBytes and GBytes using the suffix b, k, m or g.

--vm-ops N
stop vm workers after N bogo operations.

--vm-hang N
sleep N seconds before unmapping memory, the default is zero seconds. Specifying 0 will do an
infinite wait.

--vm-keep
do not continually unmap and map memory, just keep on re-writing to it.

--vm-locked
Lock the pages of the mapped region into memory using mmap MAP_LOCKED (since Linux 2.5.37). This
is similar to locking memory as described in mlock(2).

--vm-madvise advice
Specify the madvise 'advice' option used on the memory mapped regions used in the vm stressor.
Non-linux systems will only have the 'normal' madvise advice, linux systems support 'dontneed',
'hugepage', 'mergeable' , 'nohugepage', 'normal', 'random', 'sequential', 'unmergeable' and
'willneed' advice. If this option is not used then the default is to pick random madvise advice
for each mmap call. See madvise(2) for more details.

--vm-method m
specify a vm stress method. By default, all the stress methods are exercised sequentially, however
one can specify just one method to be used if required. Each of the vm workers have 3 phases:

1. Initialised. The anonymously memory mapped region is set to a known pattern.

2. Exercised. Memory is modified in a known predictable way. Some vm workers alter memory
sequentially, some use small or large strides to step along memory.

3. Checked. The modified memory is checked to see if it matches the expected result.

The vm methods containing 'prime' in their name have a stride of the largest prime less than 2^64,
allowing to them to thoroughly step through memory and touch all locations just once while also
doing without touching memory cells next to each other. This strategy exercises the cache and page
non-locality.

Since the memory being exercised is virtually mapped then there is no guarantee of touching page
addresses in any particular physical order. These workers should not be used to test that all the
system's memory is working correctly either, use tools such as memtest86 instead.

The vm stress methods are intended to exercise memory in ways to possibly find memory issues and
to try to force thermal errors.

Available vm stress methods are described as follows:

Method Description
all iterate over all the vm stress methods as listed below.
cache-lines work through memory in 64 byte cache sized steps writing a single
byte per cache line. Once the write is complete, the memory is read
to verify the values are written correctly.
cache-stripe work through memory in 64 byte cache sized chunks, writing in
ascending address order on even offsets and descending address
order on odd offsets.
flip sequentially work through memory 8 times, each time just one bit in
memory flipped (inverted). This will effectively invert each byte
in 8 passes.
galpat-0 galloping pattern zeros. This sets all bits to 0 and flips just 1
in 4096 bits to 1. It then checks to see if the 1s are pulled down
to 0 by their neighbours or of the neighbours have been pulled up
to 1.
galpat-1 galloping pattern ones. This sets all bits to 1 and flips just 1 in
4096 bits to 0. It then checks to see if the 0s are pulled up to 1
by their neighbours or of the neighbours have been pulled down to
0.
gray fill the memory with sequential gray codes (these only change 1 bit
at a time between adjacent bytes) and then check if they are set
correctly.
grayflip fill memory with adjacent bytes of gray code and inverted gray code
pairs to change as many bits at a time between adjacent bytes and
check if these are set correctly.
incdec work sequentially through memory twice, the first pass increments
each byte by a specific value and the second pass decrements each
byte back to the original start value. The increment/decrement
value changes on each invocation of the stressor.
inc-nybble initialise memory to a set value (that changes on each invocation
of the stressor) and then sequentially work through each byte
incrementing the bottom 4 bits by 1 and the top 4 bits by 15.
rand-set sequentially work through memory in 64 bit chunks setting bytes in
the chunk to the same 8 bit random value. The random value changes
on each chunk. Check that the values have not changed.
rand-sum sequentially set all memory to random values and then summate the
number of bits that have changed from the original set values.
read64 sequentially read memory using 32 x 64 bit reads per bogo loop.
Each loop equates to one bogo operation. This exercises raw memory
reads.
ror fill memory with a random pattern and then sequentially rotate 64
bits of memory right by one bit, then check the final
load/rotate/stored values.
swap fill memory in 64 byte chunks with random patterns. Then swap each
64 chunk with a randomly chosen chunk. Finally, reverse the swap to
put the chunks back to their original place and check if the data
is correct. This exercises adjacent and random memory load/stores.
move-inv sequentially fill memory 64 bits of memory at a time with random
values, and then check if the memory is set correctly. Next,
sequentially invert each 64 bit pattern and again check if the
memory is set as expected.
modulo-x fill memory over 23 iterations. Each iteration starts one byte
further along from the start of the memory and steps along in 23
byte strides. In each stride, the first byte is set to a random
pattern and all other bytes are set to the inverse. Then it checks
see if the first byte contains the expected random pattern. This
exercises cache store/reads as well as seeing if neighbouring cells
influence each other.
mscan fill each bit in each byte with 1s then check these are set, fill
each bit in each byte with 0s and check these are clear.
prime-0 iterate 8 times by stepping through memory in very large prime
strides clearing just on bit at a time in every byte. Then check to
see if all bits are set to zero.
prime-1 iterate 8 times by stepping through memory in very large prime
strides setting just on bit at a time in every byte. Then check to
see if all bits are set to one.
prime-gray-0 first step through memory in very large prime strides clearing just
on bit (based on a gray code) in every byte. Next, repeat this but
clear the other 7 bits. Then check to see if all bits are set to
zero.
prime-gray-1 first step through memory in very large prime strides setting just
on bit (based on a gray code) in every byte. Next, repeat this but
set the other 7 bits. Then check to see if all bits are set to one.
rowhammer try to force memory corruption using the rowhammer memory stressor.
This fetches two 32 bit integers from memory and forces a cache
flush on the two addresses multiple times. This has been known to
force bit flipping on some hardware, especially with lower
frequency memory refresh cycles.
walk-0d for each byte in memory, walk through each data line setting them
to low (and the others are set high) and check that the written
value is as expected. This checks if any data lines are stuck.
walk-1d for each byte in memory, walk through each data line setting them
to high (and the others are set low) and check that the written
value is as expected. This checks if any data lines are stuck.
walk-0a in the given memory mapping, work through a range of specially
chosen addresses working through address lines to see if any
address lines are stuck low. This works best with physical memory
addressing, however, exercising these virtual addresses has some
value too.

walk-1a in the given memory mapping, work through a range of specially
chosen addresses working through address lines to see if any
address lines are stuck high. This works best with physical memory
addressing, however, exercising these virtual addresses has some
value too.
write64 sequentially write to memory using 32 x 64 bit writes per bogo
loop. Each loop equates to one bogo operation. This exercises raw
memory writes. Note that memory writes are not checked at the end
of each test iteration.
write64nt sequentially write to memory using 32 x 64 bit non-temporal writes
per bogo loop. Each loop equates to one bogo operation. This
exercises cacheless raw memory writes and is only available on x86
sse2 capable systems built with gcc and clang compilers. Note that
memory writes are not checked at the end of each test iteration.
write1024v sequentially write to memory using 1 x 1024 bit vector write per
bogo loop (only available if the compiler supports vector types).
Each loop equates to one bogo operation. This exercises raw memory
writes. Note that memory writes are not checked at the end of each
test iteration.
zero-one set all memory bits to zero and then check if any bits are not
zero. Next, set all the memory bits to one and check if any bits
are not one.

--vm-populate
populate (prefault) page tables for the memory mappings; this can stress swapping. Only available
on systems that support MAP_POPULATE (since Linux 2.5.46).

--vm-addr N
start N workers that exercise virtual memory addressing using various methods to walk through a
memory mapped address range. This will exercise mapped private addresses from 8MB to 64MB per
worker and try to generate cache and TLB inefficient addressing patterns. Each method will set the
memory to a random pattern in a write phase and then sanity check this in a read phase.

--vm-addr-ops N
stop N workers after N bogo addressing passes.

--vm-addr-method M
specify a vm address stress method. By default, all the stress methods are exercised sequentially,
however one can specify just one method to be used if required.

Available vm address stress methods are described as follows:

Method Description
all iterate over all the vm stress methods as listed below.
pwr2 work through memory addresses in steps of powers of two.
pwr2inv like pwr2, but with the all relevant address bits inverted.
gray work through memory with gray coded addresses so that each change
of address just changes 1 bit compared to the previous address.
grayinv like gray, but with the all relevant address bits inverted, hence
all bits change apart from 1 in the address range.
rev work through the address range with the bits in the address range
reversed.
revinv like rev, but with all the relevant address bits inverted.
inc work through the address range forwards sequentially, byte by byte.
incinv like inc, but with all the relevant address bits inverted.
dec work through the address range backwards sequentially, byte by
byte.
decinv like dec, but with all the relevant address bits inverted.

--vm-rw N
start N workers that transfer memory to/from a parent/child using process_vm_writev(2) and
process_vm_readv(2). This is feature is only supported on Linux. Memory transfers are only
verified if the --verify option is enabled.

--vm-rw-ops N
stop vm-rw workers after N memory read/writes.

--vm-rw-bytes N
mmap N bytes per vm-rw worker, the default is 16MB. One can specify the size as % of total
available memory or in units of Bytes, KBytes, MBytes and GBytes using the suffix b, k, m or g.

--vm-segv N
start N workers that create a child process that unmaps its address space causing a SIGSEGV on
return from the unmap.

--vm-segv-ops N
stop after N bogo vm-segv SIGSEGV faults.

--vm-splice N
move data from memory to /dev/null through a pipe without any copying between kernel address space
and user address space using vmsplice(2) and splice(2). This is only available for Linux.

--vm-splice-ops N
stop after N bogo vm-splice operations.

--vm-splice-bytes N
transfer N bytes per vmsplice call, the default is 64K. One can specify the size as % of total
available memory or in units of Bytes, KBytes, MBytes and GBytes using the suffix b, k, m or g.

--wait N
start N workers that spawn off two children; one spins in a pause(2) loop, the other continually
stops and continues the first. The controlling process waits on the first child to be resumed by
the delivery of SIGCONT using waitpid(2) and waitid(2).

--wait-ops N
stop after N bogo wait operations.

--watchdog N
start N workers that exercising the /dev/watchdog watchdog interface by opening it, perform
various watchdog specific ioctl(2) commands on the device and close it. Before closing the
special watchdog magic close message is written to the device to try and force it to never trip a
watchdog reboot after the stressor has been run. Note that this stressor needs to be run as root
with the --pathological option and is only available on Linux.

--watchdog-ops N
stop after N bogo operations on the watchdog device.

--wcs N
start N workers that exercise various libc wide character string functions on random strings.

--wcs-method wcsfunc
select a specific libc wide character string function to stress. Available string functions to
stress are: all, wcscasecmp, wcscat, wcschr, wcscoll, wcscmp, wcscpy, wcslen, wcsncasecmp,
wcsncat, wcsncmp, wcsrchr and wcsxfrm. The 'all' method is the default and will exercise all the
string methods.

--wcs-ops N
stop after N bogo wide character string operations.

--x86syscall N
start N workers that repeatedly exercise the x86-64 syscall instruction to call the getcpu(2),
gettimeofday(2) and time(2) system using the Linux vsyscall handler. Only for Linux.

--x86syscall-ops N
stop after N x86syscall system calls.

--x86syscall-func F
Instead of exercising the 3 syscall system calls, just call the syscall function F. The function F
must be one of getcpu, gettimeofday and time.

--xattr N
start N workers that create, update and delete batches of extended attributes on a file.

--xattr-ops N
stop after N bogo extended attribute operations.

-y N, --yield N
start N workers that call sched_yield(2). This stressor ensures that at least 2 child processes
per CPU exercise shield_yield(2) no matter how many workers are specified, thus always ensuring
rapid context switching.

--yield-ops N
stop yield stress workers after N sched_yield(2) bogo operations.

--zero N
start N workers reading /dev/zero.

--zero-ops N
stop zero stress workers after N /dev/zero bogo read operations.

--zlib N
start N workers compressing and decompressing random data using zlib. Each worker has two
processes, one that compresses random data and pipes it to another process that decompresses the
data. This stressor exercises CPU, cache and memory.

--zlib-ops N
stop after N bogo compression operations, each bogo compression operation is a compression of 64K
of random data at the highest compression level.

--zlib-level L
specify the compression level (0..9), where 0 = no compression, 1 = fastest compression and 9 =
best compression.

--zlib-method method
specify the type of random data to send to the zlib library. By default, the data stream is
created from a random selection of the different data generation processes. However one can
specify just one method to be used if required. Available zlib data generation methods are
described as follows:

Method Description
00ff randomly distributed 0x00 and 0xFF values.
ascii01 randomly distributed ASCII 0 and 1 characters.
asciidigits randomly distributed ASCII digits in the range of 0 and 9.
bcd packed binary coded decimals, 0..99 packed into 2 4-bit nybbles.
binary 32 bit random numbers.
brown 8 bit brown noise (Brownian motion/Random Walk noise).
double double precision floating point numbers from sin(θ).
fixed data stream is repeated 0x04030201.
gray 16 bit gray codes generated from an incrementing counter.
latin Random latin sentences from a sample of Lorem Ipsum text.
logmap Values generated from a logistical map of the equation Χn+1 = r ×
Χn × (1 - Χn) where r > ≈ 3.56994567 to produce chaotic data. The
values are scaled by a large arbitrary value and the lower 8 bits
of this value are compressed.
lfsr32 Values generated from a 32 bit Galois linear feedback shift
register using the polynomial x↑32 + x↑31 + x↑29 + x + 1. This
generates a ring of 2↑32 - 1 unique values (all 32 bit values
except for 0).
lrand48 Uniformly distributed pseudo-random 32 bit values generated from
lrand48(3).
morse Morse code generated from random latin sentences from a sample of
Lorem Ipsum text.
nybble randomly distributed bytes in the range of 0x00 to 0x0f.
objcode object code selected from a random start point in the stress-ng
text segment.
parity 7 bit binary data with 1 parity bit.

pink pink noise in the range 0..255 generated using the Gardner method
with the McCartney selection tree optimization. Pink noise is where
the power spectral density is inversely proportional to the
frequency of the signal and hence is slightly compressible.
random segments of the data stream are created by randomly calling the
different data generation methods.
rarely1 data that has a single 1 in every 32 bits, randomly located.
rarely0 data that has a single 0 in every 32 bits, randomly located.
text random ASCII text.
utf8 random 8 bit data encoded to UTF-8.
zero all zeros, compresses very easily.

--zlib-window-bits W
specify the window bits used to specify the history buffer size. The value is specified as the
base two logarithm of the buffer size (e.g. value 9 is 2^9 = 512 bytes). Default is 15.

Values:
-8-(-15): raw deflate format
8-15: zlib format
24-31: gzip format
40-47: inflate auto format detection using zlib deflate format

--zlib-mem-level L specify the reserved compression state memory for zlib. Default is 8.

Values:
1 = minimum memory usage
9 = maximum memory usage

--zlib-strategy S
specifies the strategy to use when deflating data. This is used to tune the compression algorithm.
Default is 0.

Values:
0: used for normal data (Z_DEFAULT_STRATEGY)
1: for data generated by a filter or predictor (Z_FILTERED)
2: forces huffman encoding (Z_HUFFMAN_ONLY)
3: Limit match distances to one run-length-encoding (Z_RLE)
4: prevents dynamic huffman codes (Z_FIXED)

--zlib-stream-bytes S
specify the amount of bytes to deflate until deflate should finish the block and return with
Z_STREAM_END. One can specify the size in units of Bytes, KBytes, MBytes and GBytes using the
suffix b, k, m or g. Default is 0 which creates and endless stream until stressor ends.

Values:
0: creates an endless deflate stream until stressor stops
n: creates an stream of n bytes over and over again.
Each block will be closed with Z_STREAM_END.

--zombie N
start N workers that create zombie processes. This will rapidly try to create a default of 8192
child processes that immediately die and wait in a zombie state until they are reaped. Once the
maximum number of processes is reached (or fork fails because one has reached the maximum allowed
number of children) the oldest child is reaped and a new process is then created in a first-in
first-out manner, and then repeated.

--zombie-ops N
stop zombie stress workers after N bogo zombie operations.

--zombie-max N
try to create as many as N zombie processes. This may not be reached if the system limit is less
than N.

EXAMPLES

       stress-ng --vm 8 --vm-bytes 80% -t 1h

              run 8 virtual memory stressors that combined use 80% of the available memory for 1 hour. Thus each
              stressor uses 10% of the available memory.

       stress-ng --cpu 4 --io 2 --vm 1 --vm-bytes 1G --timeout 60s

              runs for 60 seconds with 4 cpu stressors, 2 io stressors and 1 vm stressor using  1GB  of  virtual
              memory.

       stress-ng --iomix 2 --iomix-bytes 10% -t 10m

              runs  2  instances  of  the  mixed I/O stressors using a total of 10% of the available file system
              space for 10 minutes. Each stressor will use 5% of the available file system space.

       stress-ng --cyclic 1 --cyclic-dist 2500 --cyclic-method clock_ns --cyclic-prio 100  --cyclic-sleep  10000
       --hdd 0 -t 1m

              measures real time scheduling latencies created by the hdd stressor. This uses the high resolution
              nanosecond clock to measure latencies during sleeps of 10,000 nanoseconds. At the end of 1  minute
              of  stressing,  the latency distribution with 2500 ns intervals will be displayed. NOTE: this must
              be run with the CAP_SYS_NICE capability to  enable  the  real  time  scheduling  to  get  accurate
              measurements.

       stress-ng --cpu 8 --cpu-ops 800000

              runs 8 cpu stressors and stops after 800000 bogo operations.

       stress-ng --sequential 2 --timeout 2m --metrics

              run  2 simultaneous instances of all the stressors sequentially one by one, each for 2 minutes and
              summarise with performance metrics at the end.

       stress-ng --cpu 4 --cpu-method fft --cpu-ops 10000 --metrics-brief

              run 4 FFT cpu stressors, stop after 10000 bogo operations and produce a summary just for  the  FFT
              results.

       stress-ng --cpu -1 --cpu-method all -t 1h --cpu-load 90

              run  cpu  stressors on all online CPUs working through all the available CPU stressors for 1 hour,
              loading the CPUs at 90% load capacity.

       stress-ng --cpu 0 --cpu-method all -t 20m

              run cpu stressors on all configured CPUs working through all the available CPU  stressors  for  20
              minutes

       stress-ng --all 4 --timeout 5m

              run 4 instances of all the stressors for 5 minutes.

       stress-ng --random 64

              run 64 stressors that are randomly chosen from all the available stressors.

       stress-ng --cpu 64 --cpu-method all --verify -t 10m --metrics-brief

              run  64  instances of all the different cpu stressors and verify that the computations are correct
              for 10 minutes with a bogo operations summary at the end.

       stress-ng --sequential -1 -t 10m

              run all the stressors one by one for 10 minutes, with the number of  instances  of  each  stressor
              matching the number of online CPUs.

       stress-ng --sequential 8 --class io -t 5m --times

              run  all  the  stressors  in  the io class one by one for 5 minutes each, with 8 instances of each
              stressor running concurrently and show overall time utilisation statistics at the end of the run.

       stress-ng --all -1 --maximize --aggressive

              run all the stressors (1 instance of each per online CPU) simultaneously,  maximize  the  settings
              (memory sizes, file allocations, etc.) and select the most demanding/aggressive options.

       stress-ng --random 32 -x numa,hdd,key

              run 32 randomly selected stressors and exclude the numa, hdd and key stressors

       stress-ng --sequential 4 --class vm --exclude bigheap,brk,stack

              run  4  instances  of  the VM stressors one after each other, excluding the bigheap, brk and stack
              stressors

       stress-ng --taskset 0,2-3 --cpu 3

              run 3 instances of the CPU stressor and pin them to CPUs 0, 2 and 3.

EXIT STATUS

         Status     Description
           0        Success.
           1        Error; incorrect user options or a fatal resource issue in the  stress-ng  stressor  harness
                    (for example, out of memory).
           2        One or more stressors failed.
           3        One  or more stressors failed to initialise because of lack of resources, for example ENOMEM
                    (no memory), ENOSPC (no space on file system) or a missing or unimplemented system call.
           4        One or more stressors were not implemented on a specific architecture or operating system.
           5        A stressor has been killed by an unexpected signal.
           6        A stressor exited by exit(2) which  was  not  expected  and  timing  metrics  could  not  be
                    gathered.
           7        The bogo ops metrics maybe untrustworthy. This is most likely to occur when a stress test is
                    terminated during the update of a bogo-ops counter such as when it has been  OOM  killed.  A
                    less likely reason is that the counter ready indicator has been corrupted.

BUGS

       File bug reports at:
         https://github.com/ColinIanKing/stress-ng/issues

AUTHOR

       stress-ng  was  written  by Colin Ian King <colin.i.king@gmail.com> and is a clean room re-implementation
       and extension of the original stress tool by Amos Waterland. Thanks also  for  contributions  from  Abdul
       Haleem,  Aboorva  Devarajan,  Adrian  Ratiu,  André  Wild, Alexander Kanavin, Baruch Siach, Carlos Santos
       Christian Ehrhardt, Chunyu Hu, Danilo Krummrich, David Turner, Dominik B Czarnota, Fabien Malfoy, Fabrice
       Fontaine, Helmut Grohne, James Hunt, James Wang, Jianshen Liu, Jim Rowan, John Kacur, Joseph DeVincentis,
       Jules Maselbas, Khalid Elmously, Khem Raj, Luca Pizzamiglio, Luis Henriques, Manoj Iyer, Matthew Tippett,
       Mauricio Faria de Oliveira, Maxime Chevallier, Piyush Goyal, Ralf Ramsauer, Rob Colclaser, Thadeu Lima de
       Souza Cascardo, Thia Wyrod, Tim Gardner, Tim Orling, Tommi Rantala, Witold Baryluk, Zhiyi Sun and others.

NOTES

       Sending a SIGALRM, SIGINT or SIGHUP to stress-ng causes it to terminate all the  stressor  processes  and
       ensures temporary files and shared memory segments are removed cleanly.

       Sending a SIGUSR2 to stress-ng will dump out the current load average and memory statistics.

       Note  that  the  stress-ng cpu, io, vm and hdd tests are different implementations of the original stress
       tests and hence may produce different stress characteristics.  stress-ng does not support any GPU  stress
       tests.

       The bogo operations metrics may change with each release  because of bug fixes to the code, new features,
       compiler optimisations or changes in system call performance.

COPYRIGHT

       Copyright © 2013-2021 Canonical Ltd, Copyright © 2021 Colin Ian King.
       This is free software; see the source for copying  conditions.   There  is  NO  warranty;  not  even  for
       MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.

                                                  6 March 2022                                      STRESS-NG(1)