• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1# Reducing Variance
2
3<a name="disabling-cpu-frequency-scaling" />
4
5## Disabling CPU Frequency Scaling
6
7If you see this error:
8
9```
10***WARNING*** CPU scaling is enabled, the benchmark real time measurements may be noisy and will incur extra overhead.
11```
12
13you might want to disable the CPU frequency scaling while running the
14benchmark, as well as consider other ways to stabilize the performance of
15your system while benchmarking.
16
17See [Reducing Variance](reducing_variance.md) for more information.
18
19Exactly how to do this depends on the Linux distribution,
20desktop environment, and installed programs.  Specific details are a moving
21target, so we will not attempt to exhaustively document them here.
22
23One simple option is to use the `cpupower` program to change the
24performance governor to "performance".  This tool is maintained along with
25the Linux kernel and provided by your distribution.
26
27It must be run as root, like this:
28
29```bash
30sudo cpupower frequency-set --governor performance
31```
32
33After this you can verify that all CPUs are using the performance governor
34by running this command:
35
36```bash
37cpupower frequency-info -o proc
38```
39
40The benchmarks you subsequently run will have less variance.
41
42<a name="reducing-variance" />
43
44## Reducing Variance in Benchmarks
45
46The Linux CPU frequency governor [discussed
47above](user_guide#disabling-cpu-frequency-scaling) is not the only source
48of noise in benchmarks.  Some, but not all, of the sources of variance
49include:
50
511. On multi-core machines not all CPUs/CPU cores/CPU threads run the same
52   speed, so running a benchmark one time and then again may give a
53   different result depending on which CPU it ran on.
542. CPU scaling features that run on the CPU, like Intel's Turbo Boost and
55   AMD Turbo Core and Precision Boost, can temporarily change the CPU
56   frequency even when the using the "performance" governor on Linux.
573. Context switching between CPUs, or scheduling competition on the CPU the
58   benchmark is running on.
594. Intel Hyperthreading or AMD SMT causing the same issue as above.
605. Cache effects caused by code running on other CPUs.
616. Non-uniform memory architectures (NUMA).
62
63These can cause variance in benchmarks results within a single run
64(`--benchmark_repetitions=N`) or across multiple runs of the benchmark
65program.
66
67Reducing sources of variance is OS and architecture dependent, which is one
68reason some companies maintain machines dedicated to performance testing.
69
70Some of the easier and and effective ways of reducing variance on a typical
71Linux workstation are:
72
731. Use the performance governor as [discussed
74above](user_guide#disabling-cpu-frequency-scaling).
751. Disable processor boosting by:
76   ```sh
77   echo 0 | sudo tee /sys/devices/system/cpu/cpufreq/boost
78   ```
79   See the Linux kernel's
80   [boost.txt](https://www.kernel.org/doc/Documentation/cpu-freq/boost.txt)
81   for more information.
822. Set the benchmark program's task affinity to a fixed cpu.  For example:
83   ```sh
84   taskset -c 0 ./mybenchmark
85   ```
863. Disabling Hyperthreading/SMT.  This can be done in the Bios or using the
87   `/sys` file system (see the LLVM project's [Benchmarking
88   tips](https://llvm.org/docs/Benchmarking.html)).
894. Close other programs that do non-trivial things based on timers, such as
90   your web browser, desktop environment, etc.
915. Reduce the working set of your benchmark to fit within the L1 cache, but
92   do be aware that this may lead you to optimize for an unrelistic
93   situation.
94
95Further resources on this topic:
96
971. The LLVM project's [Benchmarking
98   tips](https://llvm.org/docs/Benchmarking.html).
991. The Arch Wiki [Cpu frequency
100scaling](https://wiki.archlinux.org/title/CPU_frequency_scaling) page.
101