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 negative or zero value will select the number of processors available as defined by sysconf(_SC_NPROCESSORS_CONF).

OPTIONS

General stress-ng control options:

--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

start N instances of each stressor. 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.

-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.

-n, --dry-run

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

-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.

-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.

--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.

--metrics-brief

enable metrics and only output metrics that are non-zero.

--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-advise

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

--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."

-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.

--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.

--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.

-t N, --timeout N

stop stress test after N 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.

--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 fuzzinesss 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.

-V, --version

show version.

-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:

--affinity N

start N workers that rapidly change CPU affinity (only on Linux). Rapidly switching CPU affinity can contribute to poor cache behaviour.

--affinity-ops N

stop affinity workers after N bogo affinity operations (only on Linux).

--affinity-rand

switch CPU affinity randomly rather than the default of sequentially.

--af-alg N

start N workers that exercise the AF_ALG socket domain by hashing and encrypting various sized random messages. This exercises the SHA1, SHA224, SHA256, SHA384, SHA512, MD4, MD5, RMD128, RMD160, RMD256, RMD320, WP256, WP384, WP512, TGR128, TGR160, TGR192 hashes and the cbc(aes), lrw(aes), ofb(aes), xts(twofish), xts(serpent), xts(cast6), xts(camellia), lrw(twofish), lrw(cast6), lrw(camellia), salsa20 skcipers.

--af-alg-ops N

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

--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.

--apparmor N

start N workers that exercise various parts of the AppArmor interface. Currently one needs root permission to run this particular test. This test is only available on Linux systems with AppArmor support.

--apparmor-ops

stop the AppArmor workers after N bogo 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.

--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.

--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-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-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-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 capabililties via calls to capget(2) (Linux only).

--cap-ops N

stop after N cap bogo operations.

--chdir N

start N workers that change directory between 8192 directories using chdir(2).

--chdir-ops N

stop after N chdir bogo operations.

--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.

--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) system call). 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 operarions. 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.

--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 N 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 copyied 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 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.

--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.

--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, 10), 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
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
correlate	perform a 16384 × 1024 correlation of random doubles
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.
djb2a	128 rounds of hash DJB2a (Dan Bernstein hash using the xor variant) on 128 to 1 bytes of random strings
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)
fibonacci	compute Fibonacci sequence of 0, 1, 1, 2, 5, 8...
fft	4096 sample Fast Fourier Transform
float	1000 iterations of a mix of floating point operations
fnv1a	128 rounds of hash FNV-1a (Fowler–Noll–Vo hash using the xor then multiply variant) on 128 to 1 bytes of random strings
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)
jenkin	Jenkin's integer hash on 128 rounds of 128..1 bytes of random data
jmp	Simple unoptimised compare >, <, == and jmp branching
ln2	compute ln(2) based on series:  1 - 1/2 + 1/3 - 1/4 + 1/5 - 1/6 ...
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, and the lookup table ways (2 variations).
phi	compute the Golden Ratio φ using series
pi	compute π using the Srinivasa Ramanujan fast convergence algorithm
pjw	128 rounds of hash pjw function on 128 to 1 bytes of random strings
prime	find all the primes in the range 1..1000000 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..12
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)
sdbm	128 rounds of hash sdbm (as used in the SDBM database and GNU awk) on 128 to 1 bytes of random strings
sieve	find the primes in the range 1..10000000 using the sieve of Eratosthenes
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.

--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.

--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.

-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.

--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.

--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.

--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.

--exec N

start N workers continually forking children that exec stress-ng and then exit almost immediately.

--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.

-F N, --fallocate N

start N workers continually fallocating (preallocating file space) and ftuncating (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 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.

--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-size N

specify the size of the fiemap'd file in bytes. One can specify the size 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.

--filename N

start N workers that exercise file creation using various length filenames containing a range of allower 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 specifed 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 the process table using high values for --fork-max and --fork.

--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.

--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.

--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 all the get*(2) system calls.

--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 only).

--getrandom-ops N

stop getrandom workers after N bogo get operations.

--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.

-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 --ggressive 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 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).

--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.

--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.

--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.

--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.

--kcmp N

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

--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. 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.

--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.

--lockbus N

start N workers that rapidly lock and increment 64 bytes of randomly chosen memory from a 16MB mmap'd region (Intel x86 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).

--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.

--madvice-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 mememory (OOM) killer.

--malloc-bytes N

maximum per allocation/reallocation size. Allocations are randomly selected from 1 to N bytes. One can specify the size 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 ramdom 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-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.

--matrix N

start N workers that perform various matrix operations on floating point values. 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
hadamard	Hadamard product of two N × N matrices
frobenius	Frobenius product of two N × N matrices
mean	arithmetic mean of two N × N matrices
mult	multiply an N × N matrix by a scalar
prod	product of two N × N matrices
sub	subtract one N × N matrix from another N × N matrix
trans	transpose 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.

--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.

--memfd N

start N workers that create 256 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 units of Bytes, KBytes, MBytes and GBytes using the suffix b, k, m

--memfd-ops N

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

--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 of the page mapped and also is resident in memory using mincore(2).

--mincore-ops N

stop after N mincore bogo operations. One mincore bogo op is equivalent to a 1000 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.

--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.

--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 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-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.

--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.

--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.

--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.

--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.

--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 in units of Bytes, KBytes, MBytes and GBytes using the

--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.

--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

--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 executate illegal 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.

--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.

-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.

-P N, --poll N

start N workers that perform zero timeout polling via the poll(2), select(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.

--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 65536 pseudoterminals and perform various pty ioctls upon the ptys before closing them.

--pty-ops N

stop pty workers after N pty bogo operations.

-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.

--quota-ops N

stop quota stress workers after N bogo quotactl operations.

--rdrand N

start N workers that read the Intel hardware random number generator (Intel Ivybridge processors upwards).

--rdrand-ops N

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

--readahead N

start N workers that randomly seeks and performs 512 byte read/write I/O operations on a file with readahead. The default file size is 1 GB. 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 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.

--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.

--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.

--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.

--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 stessor 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.

--seccomp-ops N

stop seccomp stress workers after N seccomp filter tests.

--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.

--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 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 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.

--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.

--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.

--signpending-ops N

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

--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.

--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.

--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.

-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-ops N

stop socket stress workers after N bogo operations.

--sock-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.

--sock-type [ stream | seqpacket ]

specify the socket type to use. The default type is stream. seqpacket currently only works for the unix socket domain.

--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.

--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.

--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 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 alloca(3).

--stack-full

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-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-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.

-s N, --switch N

start N workers that send messages via pipe to a child to force context switching.

--switch-ops N

stop context switching workers after N bogo operations.

--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 in units of Bytes, KBytes, MBytes and GBytes using the suffix b, k, m or g.

--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.

--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.

--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).

--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 upto 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.

--tsc N

start N workers that read the Time Stamp Counter (TSC) 256 times per loop iteration (bogo operation). Available only on Intel x86 platforms.

--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.

--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 ipv4 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.

-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 quanity 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 in units of Bytes, KBytes, MBytes and GBytes using the suffix b, k, m or g.

--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.

--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.

--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 the process table using high values for --vfork-max and --vfork.

-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 in units of Bytes, KBytes, MBytes and GBytes using the suffix b, k, m or g.

--vm-stride N

deprecated since version 0.03.02

--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-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.
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.
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.
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 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.
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-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 in units of Bytes, KBytes, MBytes and GBytes using the suffix b, k, m or g.

--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 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.

--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.

--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 exercice 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.

--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 --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 --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 0 --cpu-method all -t 1h

run cpu stressors on all online CPUs working through all the available CPU stressors for 1 hour.

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 0 -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 0 --maximize --aggressive

run all the stressors (1 instance of each per 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.

AUTHOR

stress-ng was written by Colin King <[email protected]> and is a clean room re-implementation and extension of the original stress tool by Amos Waterland <[email protected]>. Thanks also for contributions from Christian Ehrhardt, James Hunt, Jim Rowan, Tim Gardner and Luca Pizzamiglio.

NOTES

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 © 2013-2016 Canonical Ltd.
This is free software; see the source for copying conditions. There is NO warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.