code review 6441134: net/http: add parallel client/server benchmark

net/http: add parallel client/server benchmark
R=bradfitz@golang.org

[dvyukov, 2012-08-14 21:47:07]

Hello golang-dev@googlegroups.com,

I'd like you to review this change to
https://dvyukov%40google.com@code.google.com/p/go/

[dvyukov, 2012-08-14 21:56:38]

Is it a representative benchmark?

I see strange results on it:
BenchmarkClientServerParallel	   50000	    123094 ns/op
BenchmarkClientServerParallel-2	  100000	     81216 ns/op
BenchmarkClientServerParallel-4	  100000	     55722 ns/op
BenchmarkClientServerParallel-8	   10000	    302047 ns/op  <<-- here
BenchmarkClientServerParallel-16	   10000	    324270 ns/op  <<-- and here

However if I change
-P := runtime.GOMAXPROCS(-1) * 32
+P := 100
then it looks reasonable:
BenchmarkClientServerParallel	   50000	    136820 ns/op
BenchmarkClientServerParallel-2	  100000	     73536 ns/op
BenchmarkClientServerParallel-4	  100000	     43542 ns/op
BenchmarkClientServerParallel-8	  200000	     33484 ns/op
BenchmarkClientServerParallel-16	  100000	     37460 ns/op

I see absolutely the same thing on TCP net benchmarks. Something with high
number of goroutines...

[dave_cheney.net, 2012-08-16 11:23:48]

Looks reasonable to me. I think you should add a P=100 case as well if you think
it shows something interesting.

[dvyukov, 2012-08-17 17:30:25]

On 2012/08/16 11:23:48, dfc wrote:
> Looks reasonable to me. I think you should add a P=100 case as well if you
think
> it shows something interesting.

PTAL
I've added cases with GOMAXPROCS*4 and GOMAXPROCS*64. There is indeed something
interesting. I have no explanation for now. When supplied with high number of
descriptors epoll sometimes returns in 4mks and sometimes in 30000mks. Does
somebody see the same picture?

BenchmarkClientServerParallel4		  50000			175669 ns/op
BenchmarkClientServerParallel4-2	  100000	     82637 ns/op
BenchmarkClientServerParallel4-4	  100000	     52903 ns/op
BenchmarkClientServerParallel4-8	  100000	    294324 ns/op
BenchmarkClientServerParallel4-16	  100000	    454013 ns/op

BenchmarkClientServerParallel64	       10000	    153067 ns/op
BenchmarkClientServerParallel64-2	   20000	     87302 ns/op
BenchmarkClientServerParallel64-4	   10000	    302023 ns/op
BenchmarkClientServerParallel64-8	   10000	    325881 ns/op
BenchmarkClientServerParallel64-16	   10000	    904687 ns/op

[bradfitz, 2012-08-20 05:51:29]

LGTM

http://codereview.appspot.com/6441134/diff/9001/src/pkg/net/http/serve_test.go
File src/pkg/net/http/serve_test.go (right):

http://codereview.appspot.com/6441134/diff/9001/src/pkg/net/http/serve_test.g...
src/pkg/net/http/serve_test.go:1303: P := runtime.GOMAXPROCS(-1) * conc
capital P? how about just numProcs? it's not used often.

http://codereview.appspot.com/6441134/diff/9001/src/pkg/net/http/serve_test.g...
src/pkg/net/http/serve_test.go:1312: b.Logf("Get: %v", err)
Logf or Errorf? Seems like you want Errorf.

http://codereview.appspot.com/6441134/diff/9001/src/pkg/net/http/serve_test.g...
src/pkg/net/http/serve_test.go:1317: b.Logf("ReadAll: %v", err)
Errorf?

[gobot, 2012-08-20 06:16:50]

R=dsymonds (assigned by bradfitz)

[gobot, 2012-08-20 06:20:57]

R=bradfitz (assigned by bradfitz)

[dvyukov, 2012-08-20 09:13:19]

PTAL

http://codereview.appspot.com/6441134/diff/9001/src/pkg/net/http/serve_test.go
File src/pkg/net/http/serve_test.go (right):

http://codereview.appspot.com/6441134/diff/9001/src/pkg/net/http/serve_test.g...
src/pkg/net/http/serve_test.go:1303: P := runtime.GOMAXPROCS(-1) * conc
On 2012/08/20 05:51:29, bradfitz wrote:
> capital P? how about just numProcs? it's not used often.

Done.

http://codereview.appspot.com/6441134/diff/9001/src/pkg/net/http/serve_test.g...
src/pkg/net/http/serve_test.go:1312: b.Logf("Get: %v", err)
On 2012/08/20 05:51:29, bradfitz wrote:
> Logf or Errorf? Seems like you want Errorf.

I am not sure. First, Errorf() in a child goroutine does not abort benchmark.
Second, I am not sure that I want to abort it here, because occasionally I see
some transient errors on Linux/Mac. Sometimes I just can't get a clean run with
-cpu=1,2,4,8,12,16 -benchtime=10, it's better if you may decide that few errors
are OK for such benchmark.
Note that there is a panic if response body is incorrect.

[dvyukov, 2012-08-24 10:20:01]

*** Submitted as http://code.google.com/p/go/source/detail?r=219989b731b8 ***

net/http: add parallel client/server benchmark
R=bradfitz@golang.org

R=bradfitz
CC=bradfitz, dave, dsymonds, gobot, golang-dev
http://codereview.appspot.com/6441134