code review 88040051: bench: remove manual GC from garbage benchmark

bench: remove manual GC from garbage benchmark
Initially the garbage benchmark was flaky because
some runs trigger 3 GCs while others trigger 4 GCs.
This caused 33% flakes.
Then I've added the manual GC triggering that emulates
GOGC=100 but more stable. This fixed the flakiness.
Then we've changed runtime.GC so that it executes 2 GCs.
This increased GC time by 50%, and numbers between releases
become uncomparable.
I don't have a better idea than to remove the manual GC triggerring.

[dvyukov, 2014-04-17 14:30:15]

Hello adg@golang.org (cc: golang-codereviews@googlegroups.com, rsc@golang.org),

I'd like you to review this change to
https://code.google.com/p/go.benchmarks

[dvyukov, 2014-04-17 14:32:37]

Russ, do you have any ideas how to make stable garbage benchmark w/o calls to
runtime.GC?
Note that it runs with 1GB heap and GOMAXPROCS up to 16.

[rsc, 2014-04-17 14:58:39]

runtime.GC needs to do a full gc + sweep. The second GC is my clumsy way to
force the sweep. If you have a clean, simple way to force the sweep without
the second full GC, that's fine with me and would keep the timings
comparable.

Russ

[dvyukov, 2014-04-18 09:01:50]

On Thu, Apr 17, 2014 at 6:58 PM, Russ Cox <rsc@golang.org> wrote:

> runtime.GC needs to do a full gc + sweep. The second GC is my clumsy way
> to force the sweep. If you have a clean, simple way to force the sweep
> without the second full GC, that's fine with me and would keep the timings
> comparable.
>
>
The simple and clean way to force sweep is to start GC with a special flag
if called from runtime.GC. If the flag is set, GC will do eager sweep.

But what concerns me is that it will be a different mode than what happens
in real application, so we will measure and profile something potentially
uninteresting. The concurrent sweep is subtle enough performance-wise.

[adg, 2014-04-21 06:09:51]

LGTM

[dvyukov, 2014-04-21 14:23:55]

*** Submitted as
https://code.google.com/p/go/source/detail?r=a56e45af2aa7&repo=benchmarks ***

bench: remove manual GC from garbage benchmark
Initially the garbage benchmark was flaky because
some runs trigger 3 GCs while others trigger 4 GCs.
This caused 33% flakes.
Then I've added the manual GC triggering that emulates
GOGC=100 but more stable. This fixed the flakiness.
Then we've changed runtime.GC so that it executes 2 GCs.
This increased GC time by 50%, and numbers between releases
become uncomparable.
I don't have a better idea than to remove the manual GC triggerring.

LGTM=adg
R=adg
CC=golang-codereviews, rsc
https://codereview.appspot.com/88040051