Using --prepare increases benchmark results

[hnorkowski]

When using hyperfine I noticed that it always reported worse performance when using the --prepare option to execute a command in between runs even if the command should not have any effect on the system or the benchmark command. I tested --prepare with commands like sleep 1 and curl <some address>. As benchmark command I used echo test. This is a very short command and hyperfines warns about it and suggests using --shell=none but even with that it happens. Here is a recording.

To make sure its not only with very short commands I also tested it with a slower command: Reading 3000000 bytes from /dev/urandom but it showed the same effect but the impact was smaller, probably because its a constant slowdown or it grows much slower than the overall execution time. Recording.

Is this expected behavior for some reason? Or is it a bug?

Also another question: Why is using --shell=none so much slower in the echo test example? Is it because of stdin and stdout connections?

---

Some system information:

❯ hyperfine --version
hyperfine 1.18.0

Arch Linux with 6.6.9-arch1-1 (btw)
CPU: Ryzen 5900x
Desktop: i3wm
Executed in: Alacritty -> Fish Shell -> Zellij

[sharkdp]

This is an interesting observation that I haven't made before. I can reproduce this.

Let's first rule out the shell as a possible source for this, which you also did by using --shell=none. Note that this will have an effect on which command we run, because echo and sleep are both shell builtins. So when running hyperfine "echo test" you are benchmarking the shell builtin. And when running hyperfine -N "echo test", you are benchmarking /usr/bin/echo, which is a proper executable. The former is going to be much faster, since we don't have to spawn a new process (note that hyperfine subtracts the time to spawn the shell). We could also do hyperfine "/usr/bin/echo test" which should be roughly equivalent to hyperfine -N "echo test".

So we are going to compare launching /usr/bin/echo test with and without an intermediate --prepare command. We can clearly see a difference, even more pronounced as in your results:

▶ hyperfine -N --time-unit microsecond --warmup 1000 --runs 2000 "echo test"
Benchmark 1: echo test
  Time (mean ± σ):     578.5 µs ±  49.1 µs    [User: 461.3 µs, System: 66.9 µs]
  Range (min … max):   491.5 µs … 953.0 µs    2000 runs

▶ hyperfine -N --time-unit microsecond --warmup 1000 --runs 2000 "echo test" --prepare "sleep 0.001"
Benchmark 1: echo test
  Time (mean ± σ):     2949.0 µs ± 400.6 µs    [User: 1097.9 µs, System: 1516.1 µs]
  Range (min … max):   590.0 µs … 3670.7 µs    2000 runs

Interestingly, if I use echo foo as the --prepare command, I do not see this effect:

▶ hyperfine -N --time-unit microsecond --warmup 1000 --runs 2000 "echo test" --prepare "echo foo" 
Benchmark 1: echo test
  Time (mean ± σ):     585.6 µs ±  54.9 µs    [User: 472.3 µs, System: 63.1 µs]
  Range (min … max):   495.8 µs … 924.7 µs    2000 runs

This seems to be a real effect which is not related to hyperfine.

Another thing I did was to do the run with --prepare "sleep 0.001" and record 10k runs, export the results as JSON, and plot a histogram:

We can see that the main peak is around 3 ms, but there is a clear second peak at the low (~500µs) runtime that we saw in the benchmark without --prepare. So even when running the prepare command, we sometimes see the fast runtime.

I don't know the real reason for this, but here are two wild guesses:

1. When using sleep as a --prepare command, we essentially tell the OS that it can use the current CPU core for scheduling another process. On the other hand, when running echo test in fast succession, there might be some OS-internal mechanism that pins hyperfine/echo on a single core.
2. There might be some (OS internal) caching effects that make running the same process over and over again fast.

Some evidence for hypothesis 1 comes from the following experiment, where I pin hyperfine on a certain core.. which speeds up the benchmarkee:

▶ taskset -c 3 hyperfine -N --time-unit microsecond --warmup 1000 --runs 2000 "echo test" --prepare "sleep 0.001"

Benchmark 1: echo test
  Time (mean ± σ):     738.5 µs ± 332.4 µs    [User: 555.6 µs, System: 85.6 µs]
  Range (min … max):   510.2 µs … 3666.1 µs    2000 runs

[hnorkowski]

I did not know that shells use builtin commands instead of the actual binary but when I think of if it is kinda obvious that they would do it for performance reasons. Is there a more convenient way to check if a built-in is used other than checking if a new process is created?

Hypothesis 1:

If that is the case a spinlock should not show the same behavior because it never signals the OS that it is idling.

To test this I created a simple loop with the asm macro to prevent compiler optimizations:

use std::arch::asm;

const COUNT: usize = 100_000_000;

fn main() {
    unsafe {
        asm! {
            "mov rax, {0:r}",
            "2:",
            "dec rax",
            "jnz 2b",
            in(reg) COUNT
        };
    }
}

Benchmarking this with hyperfine reported around 23ms runtime on my system and htop shows 100% usage of the cpu core.

Using this spinlock the effect is still present in the echo benchmark even with pinning:

❯ taskset -c 3 hyperfine -N '/usr/bin/echo test' --warmup 100 --runs 1000
Benchmark 1: /usr/bin/echo test
  Time (mean ± σ):     386.2 µs ±  57.4 µs    [User: 297.7 µs, System: 27.6 µs]
  Range (min … max):   352.7 µs … 931.4 µs    1000 runs

  Warning: Statistical outliers were detected. Consider re-running this benchmark on a quiet system without any interferences from other programs. It might help to use the '--warmup' or '--prepare' options.

❯ taskset -c 3 hyperfine -N '/usr/bin/echo test' --warmup 100 --runs 1000 --prepare target/release/spinlock
Benchmark 1: /usr/bin/echo test
  Time (mean ± σ):     413.2 µs ±  32.8 µs    [User: 318.0 µs, System: 33.8 µs]
  Range (min … max):   358.7 µs … 672.3 µs    1000 runs

  Warning: Statistical outliers were detected. Consider re-running this benchmark on a quiet system without any interferences from other programs.

However in the head random benchmark it is not present anymore:

❯ taskset -c 3 hyperfine -N '/usr/bin/head -c 3000000 /dev/urandom' --warmup 100 --runs 1000
Benchmark 1: /usr/bin/head -c 3000000 /dev/urandom
Time (mean ± σ):       6.4 ms ±   0.3 ms    [User: 0.2 ms, System: 6.2 ms]
Range (min … max):     6.2 ms …   7.8 ms    1000 runs

❯ taskset -c 3 hyperfine -N '/usr/bin/head -c 3000000 /dev/urandom' --warmup 100 --runs 1000 --prepare target/release/spinlock
Benchmark 1: /usr/bin/head -c 3000000 /dev/urandom
Time (mean ± σ):       6.4 ms ±   0.2 ms    [User: 0.2 ms, System: 6.1 ms]
Range (min … max):     6.2 ms …   7.8 ms    1000 runs

Afaik there should be no caching possible with /dev/urandom as it should always produce new random ouput? Maybe it can cache opening the "file" or something. So that this effect is still present in the echo benchmark could be related to caching effects or your second hypothesis.

Therefore I think the first hypothesis can be correct.

Hypothesis 2:

We can use the same command for --prepare and benchmarking. This way there will be always the same process executed and caching should not be affected:

❯ taskset -c 3 hyperfine -N 'echo test' --warmup 1000 --runs 10000
Benchmark 1: echo test
  Time (mean ± σ):     589.2 µs ±  87.3 µs    [User: 447.1 µs, System: 76.8 µs]
  Range (min … max):   519.1 µs … 1510.2 µs    10000 runs

  Warning: Statistical outliers were detected. Consider re-running this benchmark on a quiet system without any interferences from other programs. It might help to use the '--warmup' or '--prepare' options.

❯ taskset -c 3 hyperfine -N 'echo test' --warmup 1000 --runs 10000 --prepare 'echo test'
Benchmark 1: echo test
  Time (mean ± σ):     582.3 µs ±  87.6 µs    [User: 444.7 µs, System: 73.2 µs]
  Range (min … max):   517.9 µs … 1604.4 µs    10000 runs

  Warning: Statistical outliers were detected. Consider re-running this benchmark on a quiet system without any interferences from other programs.

And now it actually looks to be equal. So maybe both of your hypotheses are correct.

Confirming that it is not tool specific:

It is probably a good idea to try the same benchmarking with a different tool and see if it behaves the same. There we would need to find a tool that offers these capabilities. I can't think of one off the top of my head.

Adding a warning / notice about this behavior

I think it would be a good idea to document this behavior somewhere to make people aware of it even tho it might not be specific to this tool but rather to the system architecture.

Possible solutions

Maybe it is possible to gain exclusive access to a single core? Not sure if linux supports this. This would prevent the scheduler from running other processes on the core while the benchmarked process is waiting for I/O or otherwise idling.

[flying-sheep]

Maybe it is possible to gain exclusive access to a single core? Not sure if linux supports this.

Of course it does! https://access.redhat.com/solutions/480473

Using the kernel boot parameter isolcpus:

isolcpus=<cpu number>,...,<cpu number>
or
isolcpus=<cpu number>-<cpu number>
or
isolcpus=<cpu number>,...,<cpu number>-<cpu number>