code review 7005055: runtime: faster mcentral alloc.

runtime: faster mcentral alloc.

Reduce individual object handling by anticipating how much of them are servable.

Not a chunked transfer cache, but decent enough to make sure the bottleneck is not here.

Mac OSX, median of 10 runs:

benchmark                 old ns/op    new ns/op    delta
BenchmarkBinaryTree17    5358937333   4892813012   -8.70%
BenchmarkFannkuch11      3257752475   3315436116   +1.77%
BenchmarkGobDecode         23277349     23001114   -1.19%
BenchmarkGobEncode         14367327     14262925   -0.73%
BenchmarkGzip             441045541    440451719   -0.13%
BenchmarkGunzip           139117663    139622494   +0.36%
BenchmarkJSONEncode        45715854     45687802   -0.06%
BenchmarkJSONDecode       103949570    106530032   +2.48%
BenchmarkMandelbrot200      4542462      4548290   +0.13%
BenchmarkParse              7790558      7557540   -2.99%
BenchmarkRevcomp          831436684    832510381   +0.13%
BenchmarkTemplate         133789824    133007337   -0.58%

benchmark                  old MB/s     new MB/s  speedup
BenchmarkGobDecode            32.82        33.33    1.02x
BenchmarkGobEncode            53.42        53.86    1.01x
BenchmarkGzip                 43.70        44.01    1.01x
BenchmarkGunzip              139.09       139.14    1.00x
BenchmarkJSONEncode           42.69        42.56    1.00x
BenchmarkJSONDecode           18.78        17.91    0.95x
BenchmarkParse                 7.37         7.67    1.04x
BenchmarkRevcomp             306.83       305.70    1.00x
BenchmarkTemplate             14.57        14.56    1.00x

[Sebastien Paolacci, 2012-12-23 21:44:14]

Hello golang-dev@googlegroups.com,

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

[rsc, 2013-01-18 21:36:40]

LGTM

This is not quite the same as what we already had, in that it won't grab from
multiple spans during a single call. But it's probably good enough.

[rsc, 2013-01-18 21:39:25]

*** Submitted as https://code.google.com/p/go/source/detail?r=85eaf35d3760 ***

runtime: faster mcentral alloc.

Reduce individual object handling by anticipating how much of them are servable.

Not a chunked transfer cache, but decent enough to make sure the bottleneck is
not here.

Mac OSX, median of 10 runs:

benchmark                 old ns/op    new ns/op    delta
BenchmarkBinaryTree17    5358937333   4892813012   -8.70%
BenchmarkFannkuch11      3257752475   3315436116   +1.77%
BenchmarkGobDecode         23277349     23001114   -1.19%
BenchmarkGobEncode         14367327     14262925   -0.73%
BenchmarkGzip             441045541    440451719   -0.13%
BenchmarkGunzip           139117663    139622494   +0.36%
BenchmarkJSONEncode        45715854     45687802   -0.06%
BenchmarkJSONDecode       103949570    106530032   +2.48%
BenchmarkMandelbrot200      4542462      4548290   +0.13%
BenchmarkParse              7790558      7557540   -2.99%
BenchmarkRevcomp          831436684    832510381   +0.13%
BenchmarkTemplate         133789824    133007337   -0.58%

benchmark                  old MB/s     new MB/s  speedup
BenchmarkGobDecode            32.82        33.33    1.02x
BenchmarkGobEncode            53.42        53.86    1.01x
BenchmarkGzip                 43.70        44.01    1.01x
BenchmarkGunzip              139.09       139.14    1.00x
BenchmarkJSONEncode           42.69        42.56    1.00x
BenchmarkJSONDecode           18.78        17.91    0.95x
BenchmarkParse                 7.37         7.67    1.04x
BenchmarkRevcomp             306.83       305.70    1.00x
BenchmarkTemplate             14.57        14.56    1.00x

R=rsc, dvyukov
CC=golang-dev
https://codereview.appspot.com/7005055

Committer: Russ Cox <rsc@golang.org>

[dvyukov, 2013-01-19 07:12:24]

LGTM