runtime: scheduler fix made Windows slower

[rsc]

See brainman's comments on https://golang.org/cl/10743044/.

[peterGo]

Comment 1:

Not true for amd64.
runtime: fix CPU underutilization. Revision: 417ffa35197c.
https://code.google.com/p/go/source/detail?r=417ffa35197c
Intel i5-2390T
$ go version
go version devel +08722921ea8c Thu Jul 11 15:55:08 2013 -0400 linux/amd64
$ go test runtime -short -timeout=240s -cpu=1,2,4
ok      runtime 16.754s
$ go version
go version devel +417ffa35197c Thu Jul 11 15:57:36 2013 -0400 linux/amd64
$ go test runtime -short -timeout=240s -cpu=1,2,4
ok      runtime 16.811s
C:\>go version
go version devel +08722921ea8c Thu Jul 11 15:55:08 2013 -0400 windows/amd64
C:\>go test runtime -short -timeout=240s -cpu=1,2,4
ok      runtime 21.528s
C:\>go version
go version devel +417ffa35197c Thu Jul 11 15:57:36 2013 -0400 windows/amd64
C:\>go test runtime -short -timeout=240s -cpu=1,2,4
ok      runtime 21.593s

[peterGo]

Comment 2:

Not confirmed for windows/386.
Intel Pentium 4 2.80 GHz
C:\>go version
go version devel +08722921ea8c Thu Jul 11 15:55:08 2013 -0400 windows/386
C:\>go test runtime -short -timeout=240s -cpu=1,2,4
ok      runtime 80.125s
C:\>go version
go version devel +417ffa35197c Thu Jul 11 15:57:36 2013 -0400 windows/386
C:\>go test runtime -short -timeout=240s -cpu=1,2,4
ok      runtime 81.688s

[dvyukov]

Comment 3:

Alex, can you run the runtime tests for GOMAXPROCS=1/2/4 separately to determine how
they are affected (you have 2 cores, right?). And then please look at what exactly tests
become slower.
Runtime contains either synthetic tests or exotic corner cases. With high probability it
does not affect real programs.
However, we can't get too far performance-wise without a performance bot... I will send
a separate email to golang-dev.

[alexbrainman]

Comment 4:

I run tests (see attached aaa.bat) as per your request against
changeset:   17239:08722921ea8c
user:        ChaiShushan <chaishushan@gmail.com>
date:        Thu Jul 11 15:55:08 2013 -0400
summary:     net/rpc: fix a test bug
and
changeset:   17240:417ffa35197c
user:        Dmitriy Vyukov <dvyukov@google.com>
date:        Thu Jul 11 15:57:36 2013 -0400
summary:     runtime: fix CPU underutilization
3 times each.
The output of these is in old1, old2 and old3 (for 17239:08722921ea8c) and new1, new2
and new3 (for 17240:417ffa35197c).
The difference between the two seems to be in new TestGoroutineParallelism. It takes
about 2.5s to run it here, and it runs 3 times because of "-cpu=1,2,3". So that pretty
much accounts for the increased overall test time. All other tests differ very little.
I don't think there is anything else to do here. I'm closing the issue.
Alex

Status changed to Invalid.

Attachments:

1. aaa.bat (456 bytes)
2. new1 (15197 bytes)
3. new2 (15197 bytes)
4. new3 (15197 bytes)
5. old1 (14940 bytes)
6. old2 (14940 bytes)
7. old3 (14940 bytes)