Re: igc: Some benchmarks

[Stefan Kangas]

Helmut Eller <eller.helmut@gmail.com> writes:

> Probably because the script was buggy.  With the hopefully less buggy
> version of the same idea it looks like:
>
>       master   igc
> avg     28ms  16ms
> max    118ms  66ms
> min      0ms   0ms
> count     25    25

I note that both the maximum and average latency have been reduced by
approximately 45% with IGC.  Eli pointed out that the median is improved
by 37%, and I can only agree that this is indeed impressive.

These appear to be the first results we've seen that clearly demonstrate
IGC significantly improving latency in practice.  This will likely be a
key factor in deciding to proceed with merging the branch, though more
work obviously needs doing with stabilizing the branch too.

For context, one reference point is the stated goal of the Go runtime to
achieve GC pause times under 50ms.[1]  While 68ms for this particular
workload is good, we still lack detailed data on how much of this time
is spent solely on GC itself.  This result also comes from a specific
workload, and other workloads may present different outcomes.

Let's hope these results continue to hold up under closer scrutiny and
further benchmarking.  So far, things seem promising.

Thank you.

Footnotes:
[1]  https://www.youtube.com/watch?v=bMujSVMarqY