stress-ng puts load on the CPU, memory and disk for a set time — for experiments with load, OOM and disk I/O.
stress-ng --cpu 2 --timeout 60s
2 processes load the CPU
stress-ng --cpu 1 --cpu-load 30 -t 30s
one process at 30 %
stress-ng --vm 1 --vm-bytes 128M -t 30s
take 128 MiB of memory
stress-ng --hdd 1 --hdd-bytes 64M -t 30s --temp-path /var/tmp
stress-ng --io 1 -t 30s
sync() in a loop
stress-ng --cpu 1 -t 10s --metrics-brief
with a summary table
nice -n 19 stress-ng --cpu 1 -t 60s &
background, yielding
sudo systemd-run --unit=lab-vm -p MemoryMax=96M \
stress-ng --vm 1 --vm-bytes 160M -t 20s
in its own cgroup
stress-ng starts worker processes called stressors. Each loads its own
part of the system: --cpu computes, --vm allocates and walks through
memory, --hdd writes files, --io calls sync(). The number after a key
is how many workers to start. The training machine has version 0.17.
--timeout (-t) limits the run time. Without it stress-ng runs for a
day:
In ps and top the main process is stress-ng, the workers are
stress-ng-cpu, stress-ng-vm and so on. The main one only hands out
work: CPU and memory are used by the workers.
--metrics-brief added a line for the cpu stressor:
| bogo ops | real time | usr time | sys time | bogo ops/s (real) |
|---|---|---|---|---|
| 87462 | 70.00 | 68.68 | 0.02 | 1249.39 |
Bogo ops are the stressor's nominal operations: compare them only between
runs of the same stressor on the same machine. Two processes ran for 70
seconds on one CPU and got 68.68 seconds of usr together — half each.
| Code | Meaning |
|---|---|
| 0 | success |
| 1 | an error in the options or the temporary directory |
| 2 | a stressor failed |
| 3 | a stressor lacked resources — memory, space |
| 4 | a stressor is not supported on this system |
| 5 | a stressor was killed by an unexpected signal |
| Key | What it does | Where it shows |
|---|---|---|
--cpu N |
N processes compute | us in top, load average |
--cpu-load P |
each --cpu loads the CPU by P % |
the process's %CPU |
--vm N --vm-bytes S |
N processes take S of memory each and walk through it | free, RSS |
--hdd N --hdd-bytes S |
N processes write and delete files of S | wa, iostat |
--io N |
N processes call sync() |
bo in vmstat |
--cpu 0 starts one process per CPU. --vm-bytes takes a percentage of
available memory too: --vm-bytes 25%. --vm-keep holds the same memory
without freeing it between passes.
stress-ng gives its workers oom_score_adj 1000: when memory runs out,
the oom-killer picks them first. The main process restarts a killed
worker until the time is up. The --oomable key turns restarts off: the
stressor ends early, and stress-ng prints vm: WARNING: finished
prematurely after just 0.56 secs.
The main keys are in the table below. The full list of stressors is
stress-ng --stressors, the keys are in stress-ng --help.
| Option | What it does | Example |
|---|---|---|
-c --cpu N |
Start N processes that compute and load the CPU; 0 means one per CPU. | stress-ng --cpu 2 --timeout 60s |
-m --vm N |
Start N processes that allocate memory and walk through it. | stress-ng --vm 1 --vm-bytes 128M --timeout 30s |
--vm-bytes SIZE |
How much memory each --vm process takes (256M by default); a percentage of available memory such as 25% works too. | stress-ng --vm 1 --vm-bytes 25% --timeout 30s |
-t --timeout T |
Stop after T (seconds or 30s, 2m); without it stress-ng runs for a day. | stress-ng --cpu 1 --timeout 2m |
-l --cpu-load P |
Load the CPU by P percent per --cpu process instead of full load. | stress-ng --cpu 1 --cpu-load 30 --timeout 30s |
--vm-keep |
Keep the same memory the whole time instead of freeing and allocating it again. | stress-ng --vm 1 --vm-bytes 64M --vm-keep --timeout 30s |
-d --hdd N |
Start N processes that write and delete files in a temporary directory. | stress-ng --hdd 1 --hdd-bytes 64M --temp-path /var/tmp -t 30s |
--hdd-bytes SIZE |
How much data each --hdd process writes (1G by default). | stress-ng --hdd 1 --hdd-bytes 64M --temp-path /var/tmp -t 30s |
-i --io N |
Start N processes that call sync() in a loop, flushing the cache to disk. | stress-ng --io 1 --timeout 30s |
--metrics-brief |
Print a table of bogo ops and times per stressor at the end. | stress-ng --cpu 1 --timeout 10s --metrics-brief |
--temp-path DIR |
Directory for temporary files of --hdd and other stressors; the current one by default. | stress-ng --hdd 1 --temp-path /var/tmp --timeout 30s |
--ionice-class CLASS |
Disk priority class for worker processes — idle, besteffort, realtime. | stress-ng --hdd 1 --ionice-class idle --timeout 30s |
-q --quiet |
Do not print info messages. | stress-ng --cpu 1 --timeout 30s -q |
--oomable |
Do not restart a worker process killed by the OOM killer. | stress-ng --vm 1 --vm-bytes 160M --oomable --timeout 10s |
-v --verbose |
Print everything, including debug messages. | stress-ng --vm 1 --timeout 5s -v |
--aggressive |
Enable the heaviest stressor modes — do not run on a shared machine. dangerous | stress-ng --cpu 1 --aggressive --timeout 10s |
Nothing found. Try a shorter query.
With nice -n 19 the process takes all the free CPU (97.4 %) and yields it
to any process with the normal nice value. Memory in its own
cgroup leaves the rest of the machine alone: OOM happens inside the
lab-vm group (see oom-killer).
Three workers on one CPU push the load towards 3. The 1-minute number reaches about 1.9 in a minute and almost 3 in three (see load-average).
systemd stopped the whole service at the very first OOM (OOMPolicy=stop).
With -p OOMPolicy=continue the service lives until --timeout, and
stress-ng restarts the killed worker.
A shell job dies with the session. A systemd-run unit lives apart from
the terminal and shows in systemctl status lab-cpu. Ctrl+C stops
stress-ng with a summary: the main process collects the workers and
prints successful run completed.
--hdd writes files to the current directory, and the student cannot
write to /. Pass --temp-path /var/tmp or go to your home directory.
The full line is vm: gave up trying to mmap, no available memory, skipping
stressor. --vm-bytes is more than the machine can give (520M with 458 MiB
in total). The stressor is skipped, the exit code is 3. Make the size
smaller or give it as a percentage: --vm-bytes 50%.
--timeout is missing: by default a run lasts a day. Stop it with Ctrl+C,
kill %1 or sudo systemctl stop <unit> and start again with -t.
note: /proc/sys/kernel/sched_autogroup_enabled is 1… and for stable load
results, select a specific cpu stress method… are hints about the accuracy
of bogo ops. They do not change the load.
--vm without MemoryMax drives the machine into OOM, and neighbours may
suffer along with the stressor. --all, --aggressive and large N load the
machine so hard that everything slows down, your terminal included. On a
shared server always use --timeout, nice, ionice and a cgroup of its
own for memory.
vmstat, mpstat: what shows the load.free, available and swap.nice, renice, ionice.systemd-run.Found an inaccuracy? Report an error in the article