stress-ng is a tool designed for imposing a high load on various computer subsystems. This tool is capable of loading and stressing multiple aspects of a computer, including the CPU, cache, disk, memory, socket, and pipe I/O, as well as scheduling. Developed as an enhanced version of Amos Waterland’s original stress tool, stress-ng introduces numerous additional features. These include the ability to define the number of bogo operations, gather execution metrics, verify stress on memory and compute operations, and implement a wider range of stress mechanisms.
The primary function of stress-ng is to conduct stress tests on a computer system through different methods. It targets both the physical subsystems of a computer and the interfaces of the operating system kernel. Furthermore, stress-ng contains an array of CPU-focused stress tests, which target various operations including floating-point calculations, integer processing, bit manipulation, and control flow.
Originally, the purpose of stress-ng was to heavily load a system to identify hardware issues like thermal overruns and operating system bugs that emerge under extreme conditions. Caution is advised when using stress-ng, as some tests may cause systems with poor design to overheat or experience significant system overload, potentially leading to difficult-to-stop conditions.
In addition to its primary stress-testing functions, stress-ng can be used to measure test throughput rates. This feature is helpful for observing performance variances across different hardware or operating system versions. However, stress-ng is not designed as a benchmarking tool and should not be used for precise benchmark testing.
Executing stress-ng with root privileges on Linux systems allows for the adjustment of out-of-memory settings, rendering stressors unkillable in low-memory scenarios. This feature should be used carefully. With the necessary privileges, stress-ng also permits the modification of ‘ionice’ classes and levels, which again requires cautious application.
Users can specify the number of processes for each stress test type. Setting this number to zero defaults to the number of processors available as defined by _sysconf( SC_NPROCESSORS_CONF ). If this cannot be determined, the number of online CPUs is used. A negative value defaults to the number of online CPUs.
stress-ng includes over 220 different stress tests (stressors), encompassing a broad spectrum of system components:
The tool offers a wide range of stress mechanisms, each with a detailed description available in the manual page. This document serves as a quick-start reference guide, outlining common use cases for stress-ng.
Exercise caution when using the stress-ng tool. Certain tests have the potential to affect thermal zone trip points in systems with suboptimal hardware design. This can lead to degraded system performance and excessive system thrashing, which may present challenges in halting. Additionally, some fault stressors may render the operating system inoperable, necessitating manual intervention. This is an important consideration for users who do not have direct physical access to the system.
| # | Debian | Version |
|---|---|---|
| 13 | 0.17.05 | |
| 12 | Bookworm | 0.15.06-2 |
| 11 | Bullseye | 0.12.06 |
| 10 | Buster | 0.09.50 |
Installation of the distribution package is straight forward.
aptitude install stress-ngAmazingly stress-ng has approximately over 926 command line options1. Consequently, this guide will not attempt to cover all of them. The manual page for stress-ng includes 18 usage examples. The usage examples provided here are configured for brief durations, allowing for quick testing. However, for substantive stress testing, longer run times are typically recommended.
# Run 1 CPU threads (stressor) for 1 minute
stress-ng --cpu 1 --timeout 1m
stress-ng: info: [106383] setting to a 60 second run per stressor
stress-ng: info: [106383] dispatching hogs: 1 cpu
stress-ng: info: [106383] successful run completed in 60.00s (1 min, 0.00 secs)In case you would like to run a special CPU test, you can query which tests are available.
stress-ng --cpu-method which
cpu-method must be one of: all ackermann apery bitops callfunc cdouble \
cfloat clongdouble collatz correlate crc16 decimal32 decimal64 \
decimal128 dither div8 div16 div32 div64 div128 double euler explog \
factorial fibonacci fft fletcher16 float float32 float64 float80 \
float128 floatconversion gamma gcd gray hamming hanoi hyperbolic idct \
int128 int64 int32 int16 int8 int128float int128double int128longdouble \
int128decimal32 int128decimal64 int128decimal128 int64float int64double \
int64longdouble int32float int32double int32longdouble intconversion \
ipv4checksum jmp lfs r32 ln2 logmap longdouble loop matrixprod nsqrt \
omega parity phi pi prime psi queens rand rand48 rgb sieve stats sqrt \
trig union zetaExecute two virtual memory stressors, adjusting the number to align with the core or thread count of your system. These stressors should collectively utilize 80% of the available memory for a duration of one minute. Accordingly, each stressor will consume approximately 40% of the total available memory.
stress-ng --vm 2 --vm-bytes 80% -t 1m
stress-ng: info: [119819] setting to a 60 second run per stressor
stress-ng: info: [119819] dispatching hogs: 2 vm
stress-ng: info: [119819] successful run completed in 60.12s (1 min, 0.12 secs)Execute two instances of mixed I/O stressors, allocating a total of 10% of the available file system space for a duration of 1 minute. Each stressor is configured to utilize 5% of the available file system space.
stress-ng --iomix 2 --iomix-bytes 10% -t 1m
stress-ng: info: [106940] setting to a 60 second run per stressor
stress-ng: info: [106940] dispatching hogs: 2 iomix
stress-ng: info: [106940] successful run completed in 60.01s (1 min, 0.01 secs)Measure real-time scheduling latencies induced by the HDD stressor using a high-resolution nanosecond clock. This measurement involves assessing latencies during sleep intervals of 10,000 nanoseconds. After a one-minute stress period, the tool displays the latency distribution in 2500 nanosecond intervals. It is important to note that this test requires the CAP_SYS_NICE capability for enabling real-time scheduling, which is essential for obtaining accurate measurements.
stress-ng --cyclic 1 --cyclic-dist 2500 --cyclic-method clock_ns \
--cyclic-prio 100 --cyclic-sleep 10000 --hdd 0 -t 1m
stress-ng: info: [107070] setting to a 60 second run per stressor
stress-ng: info: [107070] dispatching hogs: 1 cyclic, 4 hdd
stress-ng: info: [107071] cyclic: sched SCHED_DEADLINE: 10000 ns delay, \
10000 samples
stress-ng: info: [107071] cyclic: mean: 4312.11 ns, mode: 3641 ns
stress-ng: info: [107071] cyclic: min: 3564 ns, max: 24051 ns, std.dev. \
992.30
stress-ng: info: [107071] cyclic: latency percentiles:
stress-ng: info: [107071] cyclic: 25.00%: 3668 ns
stress-ng: info: [107071] cyclic: 50.00%: 3993 ns
stress-ng: info: [107071] cyclic: 75.00%: 4667 ns
stress-ng: info: [107071] cyclic: 90.00%: 5289 ns
stress-ng: info: [107071] cyclic: 95.40%: 6014 ns
stress-ng: info: [107071] cyclic: 99.00%: 8153 ns
stress-ng: info: [107071] cyclic: 99.50%: 9663 ns
stress-ng: info: [107071] cyclic: 99.90%: 12810 ns
stress-ng: info: [107071] cyclic: 99.99%: 24051 ns
stress-ng: info: [107071] cyclic: latency distribution (2500 ns intervals):
stress-ng: info: [107071] cyclic: (for the first 10 buckets of 10)
stress-ng: info: [107071] cyclic: latency (ns) frequency
stress-ng: info: [107071] cyclic: 0 0
stress-ng: info: [107071] cyclic: 2500 8476
stress-ng: info: [107071] cyclic: 5000 1389
stress-ng: info: [107071] cyclic: 7500 92
stress-ng: info: [107071] cyclic: 10000 31
stress-ng: info: [107071] cyclic: 12500 8
stress-ng: info: [107071] cyclic: 15000 2
stress-ng: info: [107071] cyclic: 17500 1
stress-ng: info: [107071] cyclic: 20000 0
stress-ng: info: [107071] cyclic: 22500 1
stress-ng: info: [107071] cyclic: Note: --cyclic-samples needed to be \
4374911 to capture all the data for this run
stress-ng: info: [107070] successful run completed in 60.75s (1 min, 0.75 secs)Executing timers at a high frequency can result in a substantial increase in interrupt load. Utilizing the --timer stressor with a suitably chosen timer frequency can induce a high number of interrupts per second. It is important to note that executing this operation requires root permissions.
stress-ng --timer 32 --timer-freq 1000000 -t 1m
stress-ng: info: [108527] setting to a 60 second run per stressor
stress-ng: info: [108527] dispatching hogs: 32 timer
stress-ng: info: [108527] successful run completed in 60.69s (1 min, 0.69 secs)In this example, stress-ng tests 32 instances at 1MHz.
The combination of tasks (stressors) is easily possible by adding more command line options. Example CPU + IO + Memory.
stress-ng --cpu 4 --io 2 --vm 1 --vm-bytes 1G --timeout 60s
stress-ng: info: [108632] setting to a 60 second run per stressor
stress-ng: info: [108632] dispatching hogs: 4 cpu, 2 io, 1 vm
stress-ng: info: [108637] io: this is a legacy I/O sync stressor, \
consider using iomix instead
stress-ng: info: [108632] successful run completed in 60.06s (1 min, 0.06 secs)
# Replacing io with iomix:
stress-ng --cpu 4 --iomix 2 --vm 1 --vm-bytes 1G --timeout 60s
stress-ng: info: [108673] setting to a 60 second run per stressor
stress-ng: info: [108673] dispatching hogs: 4 cpu, 2 iomix, 1 vm
stress-ng: info: [108673] successful run completed in 60.67s (1 min, 0.67 secs)For system benchmarking or evaluating systems produced by an Original Equipment Manufacturer (OEM), injecting faults can be a useful approach to observe system reactions.
Page Faults:
stress-ng enables the testing and analysis of page fault rates by generating major page faults in pages that are not currently loaded in memory. In newer kernel versions, the userfaultfd mechanism alerts fault-handling threads about page faults within the virtual memory layout of a process. Executing these operations requires root privileges on the system.
# For older kernel (output was run on newer kernel, stress-ng + PID
# removed)
# Linux 6.1.0-18-amd64 #1 SMP PREEMPT_DYNAMIC Debian 6.1.76-1 \
# (2024-02-01) x86_64 GNU/Linux
# 1. Simple Run without performance reporting (we can not see that it do
# not fit well)
stress-ng --fault 0 -t 1m
setting to a 60 second run per stressor
dispatching hogs: 4 fault
successful run completed in 60.00s (1 min, 0.00 secs)# 2. Run with performance reporting (we can see that it do not fit well: a
# lot of 0.000 /sec reporting
stress-ng --fault 0 --perf -t 1m
setting to a 60 second run per stressor
dispatching hogs: 4 fault
fault:
544,855,786,816 CPU Cycles 8.794 B/sec
371,921,049,104 Instructions 6.003 B/sec (0.683 instr. per cycle)
70,548,210,352 Branch Instructions 1.139 B/sec
3,463,128,704 Branch Misses 55.894 M/sec ( 4.909%)
21,793,985,384 Bus Cycles 0.352 B/sec
479,480,334,820 Total Cycles 7.739 B/sec
14,189,251,132 Cache References 0.229 B/sec
878,261,112 Cache Misses 14.175 M/sec ( 6.190%)
108,783,796,528 Cache L1D Read 1.756 B/sec
4,686,096,088 Cache L1D Read Miss 75.633 M/sec ( 4.308%)
70,957,241,756 Cache L1D Write 1.145 B/sec
4,238,030,032 Cache L1I Read Miss 68.401 M/sec
1,720,397,160 Cache LL Read 27.767 M/sec
151,743,796 Cache LL Read Miss 2.449 M/sec ( 8.820%)
415,481,916 Cache LL Write 6.706 M/sec
30,201,260 Cache LL Write Miss 0.487 M/sec ( 7.269%)
108,737,787,928 Cache DTLB Read 1.755 B/sec
59,822,784 Cache DTLB Read Miss 0.966 M/sec ( 0.055%)
70,922,004,348 Cache DTLB Write 1.145 B/sec
17,035,460 Cache DTLB Write Miss 0.275 M/sec ( 0.024%)
70,553,791,700 Cache BPU Read 1.139 B/sec
3,457,202,928 Cache BPU Read Miss 55.799 M/sec ( 4.900%)
150,661,108 Cache NODE Read 2.432 M/sec
0 Cache NODE Read Miss 0.000 /sec ( 0.000%)
30,356,440 Cache NODE Write 0.490 M/sec
0 Cache NODE Write Miss 0.000 /sec ( 0.000%)
219,716,211,724 CPU Clock 3.546 B/sec
219,750,070,380 Task Clock 3.547 B/sec
2,590,976 Page Faults Total 41.818 K/sec
1,554,604 Page Faults Minor 25.091 K/sec
1,036,372 Page Faults Major 16.727 K/sec
48,180 Context Switches 777.619 /sec
48,052 Cgroup Switches 775.553 /sec
284 CPU Migrations 4.584 /sec
0 Alignment Faults 0.000 /sec
0 Emulation Faults 0.000 /sec
2,590,972 Page Faults User 41.818 K/sec
4 Page Faults Kernel 0.065 /sec
6,737,088 System Call Enter 0.109 M/sec
6,737,084 System Call Exit 0.109 M/sec
2,082,852 Kmalloc 33.617 K/sec
2,776,412 Kfree 44.811 K/sec
17,315,696 Kmem Cache Alloc 0.279 M/sec
18,768,992 Kmem Cache Free 0.303 M/sec
1,640,580 MM Page Alloc 26.479 K/sec
1,568,584 MM Page Free 25.317 K/sec
7,254,672 MMAP lock start 0.117 M/sec
7,254,672 MMAP lock release 0.117 M/sec
670,928 RCU Utilization 10.829 K/sec
0 RCU Stall Warning 0.000 /sec
912 Sched Migrate Task 14.720 /sec
0 Sched Move NUMA 0.000 /sec
47,784 Sched Wakeup 771.228 /sec
0 Sched Proc Exec 0.000 /sec
0 Sched Proc Exit 0.000 /sec
0 Sched Proc Fork 0.000 /sec
8 Sched Proc Free 0.129 /sec
0 Sched Proc Hang 0.000 /sec
0 Sched Proc Wait 0.000 /sec
48,184 Sched Switch 777.684 /sec
0 New Task 0.000 /sec
0 Context User Exit 0.000 /sec
8 Signal Generate 0.129 /sec
4 Signal Deliver 0.065 /sec
736 IRQ Entry 11.879 /sec
736 IRQ Exit 11.879 /sec
288,772 Soft IRQ Entry 4.661 K/sec
288,772 Soft IRQ Exit 4.661 K/sec
48 NMI handler 0.775 /sec
1,424 Block BIO Complete 22.983 /sec
0 IO uring submit 0.000 /sec
0 IO uring complete 0.000 /sec
0 Writeback Dirty Inode 0.000 /sec
0 Migrate MM Pages 0.000 /sec
36 SKB Consume 0.581 /sec
0 SKB Kfree 0.000 /sec
26,652 Lock Contention Begin 430.160 /sec
26,652 Lock Contention End 430.160 /sec
0 IOMMU IO Page Fault 0.000 /sec
132 IOMMU Map 2.130 /sec
532 IOMMU Unmap 8.586 /sec
518,188 Filemap page-cache add 8.363 K/sec
1,036,376 Filemap page-cache del 16.727 K/sec
0 OOM Compact Retry 0.000 /sec
0 OOM Wake Reaper 0.000 /sec
0 OOM Score Adjust Update 0.000 /sec
0 Thermal Zone Trip 0.000 /sec
successful run completed in 61.96s (1 min, 1.96 secs)# For newer kernel
# (output was run on newer kernel, stress-ng + PID removed)
# Linux 6.1.0-18-amd64 #1 SMP PREEMPT_DYNAMIC Debian 6.1.76-1 \
# (2024-02-01) x86_64 GNU/Linux
# 1. Simple run
stress-ng --userfaultfd 0 -t 1m
setting to a 60 second run per stressor
dispatching hogs: 4 userfaultfd
successful run completed in 60.00s (1 min, 0.00 secs)# 2. Run with performance report
stress-ng --userfaultfd 0 --perf -t 1m
setting to a 60 second run per stressor
dispatching hogs: 4 userfaultfd
userfaultfd:
573,784,529,308 CPU Cycles 9.563 B/sec
367,401,565,752 Instructions 6.123 B/sec (0.640 instr. per cycle)
75,247,240,040 Branch Instructions 1.254 B/sec
597,506,500 Branch Misses 9.958 M/sec ( 0.794%)
22,975,717,548 Bus Cycles 0.383 B/sec
505,493,763,492 Total Cycles 8.425 B/sec
16,891,417,900 Cache References 0.282 B/sec
778,519,852 Cache Misses 12.975 M/sec ( 4.609%)
100,779,216,856 Cache L1D Read 1.680 B/sec
12,832,061,780 Cache L1D Read Miss 0.214 B/sec (12.733%)
76,371,736,040 Cache L1D Write 1.273 B/sec
5,657,176,416 Cache L1I Read Miss 94.282 M/sec
3,251,015,820 Cache LL Read 54.181 M/sec
46,779,708 Cache LL Read Miss 0.780 M/sec ( 1.439%)
1,282,146,056 Cache LL Write 21.368 M/sec
504,428,128 Cache LL Write Miss 8.407 M/sec (39.342%)
101,289,571,044 Cache DTLB Read 1.688 B/sec
57,563,032 Cache DTLB Read Miss 0.959 M/sec ( 0.057%)
76,447,085,328 Cache DTLB Write 1.274 B/sec
85,701,532 Cache DTLB Write Miss 1.428 M/sec ( 0.112%)
75,143,922,784 Cache BPU Read 1.252 B/sec
598,705,764 Cache BPU Read Miss 9.978 M/sec ( 0.797%)
49,948,888 Cache NODE Read 0.832 M/sec
0 Cache NODE Read Miss 0.000 /sec ( 0.000%)
493,044,204 Cache NODE Write 8.217 M/sec
0 Cache NODE Write Miss 0.000 /sec ( 0.000%)
235,288,674,364 CPU Clock 3.921 B/sec
235,309,819,424 Task Clock 3.922 B/sec
8,465,072 Page Faults Total 0.141 M/sec
224 Page Faults Minor 3.733 /sec
8,464,844 Page Faults Major 0.141 M/sec
16,939,336 Context Switches 0.282 M/sec
15,748 Cgroup Switches 262.455 /sec
1,204 CPU Migrations 20.066 /sec
0 Alignment Faults 0.000 /sec
0 Emulation Faults 0.000 /sec
8,465,052 Page Faults User 0.141 M/sec
20 Page Faults Kernel 0.333 /sec
33,959,376 System Call Enter 0.566 M/sec
33,959,376 System Call Exit 0.566 M/sec
66,364 Kmalloc 1.106 K/sec
99,728 Kfree 1.662 K/sec
134,364 Kmem Cache Alloc 2.239 K/sec
138,432 Kmem Cache Free 2.307 K/sec
16,940,184 MM Page Alloc 0.282 M/sec
16,929,232 MM Page Free 0.282 M/sec
25,395,468 MMAP lock start 0.423 M/sec
25,395,468 MMAP lock release 0.423 M/sec
34,085,752 RCU Utilization 0.568 M/sec
0 RCU Stall Warning 0.000 /sec
3,312 Sched Migrate Task 55.197 /sec
0 Sched Move NUMA 0.000 /sec
13,425,252 Sched Wakeup 0.224 M/sec
0 Sched Proc Exec 0.000 /sec
4 Sched Proc Exit 0.067 /sec
8 Sched Proc Fork 0.133 /sec
4 Sched Proc Free 0.067 /sec
0 Sched Proc Hang 0.000 /sec
8 Sched Proc Wait 0.133 /sec
16,939,336 Sched Switch 0.282 M/sec
8 New Task 0.133 /sec
0 Context User Exit 0.000 /sec
12 Signal Generate 0.200 /sec
8 Signal Deliver 0.133 /sec
404 IRQ Entry 6.733 /sec
404 IRQ Exit 6.733 /sec
61,384 Soft IRQ Entry 1.023 K/sec
61,384 Soft IRQ Exit 1.023 K/sec
40 NMI handler 0.667 /sec
20 Block BIO Complete 0.333 /sec
0 IO uring submit 0.000 /sec
0 IO uring complete 0.000 /sec
4 Writeback Dirty Inode 0.067 /sec
0 Migrate MM Pages 0.000 /sec
316 SKB Consume 5.266 /sec
4 SKB Kfree 0.067 /sec
783,916 Lock Contention Begin 13.065 K/sec
783,908 Lock Contention End 13.065 K/sec
0 IOMMU IO Page Fault 0.000 /sec
4 IOMMU Map 0.067 /sec
68 IOMMU Unmap 1.133 /sec
0 Filemap page-cache add 0.000 /sec
0 Filemap page-cache del 0.000 /sec
0 OOM Compact Retry 0.000 /sec
0 OOM Wake Reaper 0.000 /sec
4 OOM Score Adjust Update 0.067 /sec
0 Thermal Zone Trip 0.000 /sec
successful run completed in 60.00s (1 min, 0.00 secs)In the output below stress-ng: REALM: [PID] was removed.
--metrics:
In comparison to --metrics-brief this adds CPU used per instance (%) and RSS Max.
stress-ng --cpu 1 --timeout 1m --metrics
setting to a 60 second run per stressor
dispatching hogs: 1 cpu
stressor bogo ops real time usr time sys time bogo ops/s bogo ops/s \
CPU used per RSS Max
(secs) (secs) (secs) (real time) (usr+sys time) \
instance (%) (KB)
cpu 74433 60.00 60.00 0.00 1240.49 1240.59 \
99.99 5824
successful run completed in 60.00s (1 min, 0.00 secs)--metrics-brief:
Compared to --metrics the output fits better a 80 char terminal and it leaves out CPU used per instance (%) and RSS Max.
stress-ng --cpu 1 --timeout 1m --metrics-brief
setting to a 60 second run per stressor
dispatching hogs: 1 cpu
stressor bogo ops real time usr time sys time bogo ops/s bogo ops/s
(secs) (secs) (secs) (real time) (usr+sys time)
cpu 71860 60.00 60.00 0.00 1197.59 1197.68
successful run completed in 60.01s (1 min, 0.01 secs)--tz: (temperature)
stress-ng --cpu 0 --cpu-method all --verify -t 1m --tz
stress-ng: info: [107199] setting to a 60 second run per stressor
stress-ng: info: [107199] dispatching hogs: 4 cpu
stress-ng: info: [107199] cpu:
stress-ng: info: [107199] pch_wildcat_point 47.50 C (320.65 K)
stress-ng: info: [107199] x86_pkg_temp 64.50 C (337.65 K)
stress-ng: info: [107199] successful run completed in 60.01s (1 min, 0.01 secs)--perf:
Measure processor and system activity using perf events, a functionality specific to Linux. However, exercise caution as per the guidance in perf_event_open(2), which advises to “Always double-check your results! Various generalized events have had wrong values.” It’s important to note that as of Linux version 4.7, CAP_SYS_ADMIN capabilities are required for this option to function2.
--verify:
Most stressors have a verification mode --verify to sanity check test operations. This adds overhead to bogo-ops rate so don’t use it for benchmarking.
Test memory with different test patterns for 1 hour:
stress-ng --vm 1 --vm-bytes 2G --verify -v -t 1h
1 hour CPU computation soak test:
stress-ng --cpu 0 --verify -t 1h
--perf:
See previous section Reporting and Faults.
When executing stress-ng, it is important to check the exit status. The tool documents eight exit conditions, which are detailed in the manual page. Refer to the bottom of the manual page for comprehensive information on these conditions.
# With ERROR:
stress-ng --cpu 0 --timeout 60s.;echo "Exit code: $?"
Illegal time specifier .
Exit code: 1
# With SUCCESS:
stress-ng --cpu 0 --timeout 60s;echo "Exit code: $?"
stress-ng: info: [108944] setting to a 60 second run per stressor
stress-ng: info: [108944] dispatching hogs: 4 cpu
stress-ng: info: [108944] successful run completed in 60.01s (1 min, 0.01 secs)
Exit code: 0| Version | Date | Notes |
|---|---|---|
| 0.1.1 | 2024-03-16 | Fix example descriptions, formatting |
| 0.1.0 | 2024-03-16 | Initial release (to GitHub) |