1.8 vs 1.9 performance

[ilya....@intel.com]

Hi everyone!

FYI I've run all benchmarks in src to check for performance vs 1.8.

Results are attached. Overall it looks great, but there are some significant ( > 10%) degradations.

I'm sure that at least in some cases lower performance is caused by changes in functionality, fixed bugs and general nosie, however some cases are vaild

At random I've checked encoding/base32/EncodeString  and it looks like real  degradation so I've created https://github.com/golang/go/issues/21262 and https://go-review.googlesource.com/c/52510

But I don't have time to analyze everything else right now, so I'm posting results here. 

[Alberto Donizetti]

That's nice, thanks for doing this.

I'll definitely try to investigate some of the worst ones for go1.10

A.

[Tylor Arndt]

I have not attempted to reproduce yet, but if a couple of these are the functions I think they are, these are pretty alarming!

time.Now() is something I know my applications have the potential to call very heavily, from setting net.Conn deadlines to calculating evetive dates from TTLs....
/localdisk/itocar/golang/src/time/Now-6                                                                 36.4ns ± 0%     72.8ns ± 1%   +99.93%  (p=0.000 n=8+8)
/localdisk/itocar/golang/src/time/NowUnixNano-6                                                         37.5ns ± 0%     74.9ns ± 1%   +99.73%  (p=0.000 n=8+8)

This could be a pretty dramatic increase in garbage generation for anything that is a public facing HTTPS server right?
/localdisk/itocar/golang/src/net/http/ClientServerParallelTLS4-6                                        4.97kB ± 0%    24.61kB ± 3%  +395.20%  (p=0.000 n=8+7)
/localdisk/itocar/golang/src/net/http/ClientServerParallelTLS4-6                                          64.0 ± 0%      258.3 ± 3%  +303.57%  (p=0.000 n=8+7)

Before my team adopts 1.9 we will have to run these tests on our prod HW and bench our apps, but at a glance the items above are a bit concerning.

[Brad Fitzpatrick]

On Fri, Aug 4, 2017 at 7:33 AM, Tylor Arndt <voidl...@gmail.com> wrote:

I have not attempted to reproduce yet, but if a couple of these are the functions I think they are, these are pretty alarming!

time.Now() is something I know my applications have the potential to call very heavily, from setting net.Conn deadlines to calculating evetive dates from TTLs....
/localdisk/itocar/golang/src/time/Now-6                                                                 36.4ns ± 0%     72.8ns ± 1%   +99.93%  (p=0.000 n=8+8)
/localdisk/itocar/golang/src/time/NowUnixNano-6                                                         37.5ns ± 0%     74.9ns ± 1%   +99.73%  (p=0.000 n=8+8)

That's expected from https://tip.golang.org/doc/go1.9#monotonic-time

I doubt you'll notice in real programs.

 

This could be a pretty dramatic increase in garbage generation for anything that is a public facing HTTPS server right?
/localdisk/itocar/golang/src/net/http/ClientServerParallelTLS4-6                                        4.97kB ± 0%    24.61kB ± 3%  +395.20%  (p=0.000 n=8+7)
/localdisk/itocar/golang/src/net/http/ClientServerParallelTLS4-6                                          64.0 ± 0%      258.3 ± 3%  +303.57%  (p=0.000 n=8+7)

Before my team adopts 1.9 we will have to run these tests on our prod HW and bench our apps, but at a glance the items above are a bit concerning.

Let us know what you find.

Feel free to file a bug too.

[ilya....@intel.com]

Josh Bleecher Snyder mentioned that benchmark parallelization in 1.9 may have hidden performance degradations, so I've rerun everything with -cpu=1

Results are attached.

[Philip Hofer]

What happened here?

/localdisk/itocar/golang/src/net/http/ClientServerParallelTLS64                                        126µs ± 2%     1800µs ±16%  +1333.24%  (p=0.000 n=8+8)

/localdisk/itocar/golang/src/net/http/ClientServerParallelTLS64                                       5.37kB ± 0%    60.32kB ±13%  +1022.65%  (p=0.000 n=7+8)

[Alberto Donizetti]

Looks like that benchmark's code was changed in CL 37771
(https://go-review.googlesource.com/c/37771) so now it's measuring
something different.