code review 9790043: runtime: inline MCache_Alloc() into mallocgc()

runtime: inline MCache_Alloc() into mallocgc()
benchmark                    old ns/op    new ns/op    delta
BenchmarkMalloc8                    68           62   -8.63%
BenchmarkMalloc16                   75           69   -7.94%
BenchmarkMallocTypeInfo8           102           98   -3.73%
BenchmarkMallocTypeInfo16          108          103   -4.63%

[dvyukov, 2013-05-26 16:33:11]

Hello golang-dev@googlegroups.com,

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

[dave_cheney.net, 2013-05-26 23:57:33]

On 2013/05/26 16:33:11, dvyukov wrote:
> Hello mailto:golang-dev@googlegroups.com,
> 
> I'd like you to review this change to
> https://dvyukov%2540google.com%40code.google.com/p/go/

This looks like a good change, do you have an profile data to show how much time
mallocgc() was waiting on MCache_Alloc() ?

[dvyukov, 2013-05-27 04:22:17]

On Mon, May 27, 2013 at 3:57 AM, <dave@cheney.net> wrote:

> On 2013/05/26 16:33:11, dvyukov wrote:
>
>> Hello mailto:golang-dev@**googlegroups.com <golang-dev@googlegroups.com>,
>>
>
>  I'd like you to review this change to
>> https://dvyukov%2540google.com**%40code.google.com/p/go/
>>
>
> This looks like a good change, do you have an profile data to show how
> much time mallocgc() was waiting on MCache_Alloc() ?
>


What do you mean "time mallocgc() was waiting on MCache_Alloc()"?
We save roughly that on the function call:
benchmark                    old ns/op    new ns/op    delta
BenchmarkMalloc8                    68           62   -8.63%
BenchmarkMalloc16                   75           69   -7.94%
BenchmarkMallocTypeInfo8           102           98   -3.73%
BenchmarkMallocTypeInfo16          108          103   -4.63%

[dave_cheney.net, 2013-05-27 04:30:50]

I'm just wondering if you have a figure for how large the overhead of
the function call was in relation to the time spent in MCache_Alloc.
Looking at the benchmarks you posted, [4,9) %

On Mon, May 27, 2013 at 2:22 PM, Dmitry Vyukov <dvyukov@google.com> wrote:
> On Mon, May 27, 2013 at 3:57 AM, <dave@cheney.net> wrote:
>>
>> On 2013/05/26 16:33:11, dvyukov wrote:
>>>
>>> Hello mailto:golang-dev@googlegroups.com,
>>
>>
>>> I'd like you to review this change to
>>> https://dvyukov%2540google.com%40code.google.com/p/go/
>>
>>
>> This looks like a good change, do you have an profile data to show how
>> much time mallocgc() was waiting on MCache_Alloc() ?
>
>
>
> What do you mean "time mallocgc() was waiting on MCache_Alloc()"?
> We save roughly that on the function call:
> benchmark                    old ns/op    new ns/op    delta
> BenchmarkMalloc8                    68           62   -8.63%
> BenchmarkMalloc16                   75           69   -7.94%
> BenchmarkMallocTypeInfo8           102           98   -3.73%
> BenchmarkMallocTypeInfo16          108          103   -4.63%
>
>

[dvyukov, 2013-05-27 06:34:52]

On Mon, May 27, 2013 at 8:30 AM, Dave Cheney <dave@cheney.net> wrote:

> I'm just wondering if you have a figure for how large the overhead of
> the function call was in relation to the time spent in MCache_Alloc.
> Looking at the benchmarks you posted, [4,9) %
>

I have not measured exactly. But function calls add significant overhead
with 6c. In this case it may be 50% or so.


On Mon, May 27, 2013 at 2:22 PM, Dmitry Vyukov <dvyukov@google.com> wrote:
> > On Mon, May 27, 2013 at 3:57 AM, <dave@cheney.net> wrote:
> >>
> >> On 2013/05/26 16:33:11, dvyukov wrote:
> >>>
> >>> Hello mailto:golang-dev@googlegroups.com,
> >>
> >>
> >>> I'd like you to review this change to
> >>> https://dvyukov%2540google.com%40code.google.com/p/go/
> >>
> >>
> >> This looks like a good change, do you have an profile data to show how
> >> much time mallocgc() was waiting on MCache_Alloc() ?
> >
> >
> >
> > What do you mean "time mallocgc() was waiting on MCache_Alloc()"?
> > We save roughly that on the function call:
> > benchmark                    old ns/op    new ns/op    delta
> > BenchmarkMalloc8                    68           62   -8.63%
> > BenchmarkMalloc16                   75           69   -7.94%
> > BenchmarkMallocTypeInfo8           102           98   -3.73%
> > BenchmarkMallocTypeInfo16          108          103   -4.63%
> >
> >
>

[dave_cheney.net, 2013-05-27 07:49:36]

Excellent. I think this change is good, but im not skilled enough to review it,
however I will try to generate some benchmarks on a variety of systems to
validate it. 

On 27/05/2013, at 16:34, Dmitry Vyukov <dvyukov@google.com> wrote:

> On Mon, May 27, 2013 at 8:30 AM, Dave Cheney <dave@cheney.net> wrote:
>> I'm just wondering if you have a figure for how large the overhead of
>> the function call was in relation to the time spent in MCache_Alloc.
>> Looking at the benchmarks you posted, [4,9) %
> 
> I have not measured exactly. But function calls add significant overhead with
6c. In this case it may be 50% or so.
> 
> 
>> On Mon, May 27, 2013 at 2:22 PM, Dmitry Vyukov <dvyukov@google.com> wrote:
>> > On Mon, May 27, 2013 at 3:57 AM, <dave@cheney.net> wrote:
>> >>
>> >> On 2013/05/26 16:33:11, dvyukov wrote:
>> >>>
>> >>> Hello mailto:golang-dev@googlegroups.com,
>> >>
>> >>
>> >>> I'd like you to review this change to
>> >>> https://dvyukov%2540google.com%40code.google.com/p/go/
>> >>
>> >>
>> >> This looks like a good change, do you have an profile data to show how
>> >> much time mallocgc() was waiting on MCache_Alloc() ?
>> >
>> >
>> >
>> > What do you mean "time mallocgc() was waiting on MCache_Alloc()"?
>> > We save roughly that on the function call:
>> > benchmark                    old ns/op    new ns/op    delta
>> > BenchmarkMalloc8                    68           62   -8.63%
>> > BenchmarkMalloc16                   75           69   -7.94%
>> > BenchmarkMallocTypeInfo8           102           98   -3.73%
>> > BenchmarkMallocTypeInfo16          108          103   -4.63%
>> >
>> >
> 

[dvyukov, 2013-05-27 08:07:35]

This is only a part of series of malloc-related changes.



On Mon, May 27, 2013 at 11:44 AM, Dave Cheney <dave@cheney.net> wrote:

> Excellent. I think this change is good, but im not skilled enough to
> review it, however I will try to generate some benchmarks on a variety of
> systems to validate it.
>
> On 27/05/2013, at 16:34, Dmitry Vyukov <dvyukov@google.com> wrote:
>
> On Mon, May 27, 2013 at 8:30 AM, Dave Cheney <dave@cheney.net> wrote:
>
>> I'm just wondering if you have a figure for how large the overhead of
>> the function call was in relation to the time spent in MCache_Alloc.
>> Looking at the benchmarks you posted, [4,9) %
>>
>
> I have not measured exactly. But function calls add significant overhead
> with 6c. In this case it may be 50% or so.
>
>
> On Mon, May 27, 2013 at 2:22 PM, Dmitry Vyukov <dvyukov@google.com> wrote:
>> > On Mon, May 27, 2013 at 3:57 AM, <dave@cheney.net> wrote:
>> >>
>> >> On 2013/05/26 16:33:11, dvyukov wrote:
>> >>>
>> >>> Hello mailto:golang-dev@googlegroups.com,
>> >>
>> >>
>> >>> I'd like you to review this change to
>> >>> https://dvyukov%2540google.com%40code.google.com/p/go/
>> >>
>> >>
>> >> This looks like a good change, do you have an profile data to show how
>> >> much time mallocgc() was waiting on MCache_Alloc() ?
>> >
>> >
>> >
>> > What do you mean "time mallocgc() was waiting on MCache_Alloc()"?
>> > We save roughly that on the function call:
>> > benchmark                    old ns/op    new ns/op    delta
>> > BenchmarkMalloc8                    68           62   -8.63%
>> > BenchmarkMalloc16                   75           69   -7.94%
>> > BenchmarkMallocTypeInfo8           102           98   -3.73%
>> > BenchmarkMallocTypeInfo16          108          103   -4.63%
>> >
>> >
>>
>
>

[dave_cheney.net, 2013-05-27 11:30:30]

linux/arm pandaboard

benchmark                    old ns/op    new ns/op    delta
BenchmarkMalloc8                   485          462   -4.74%
BenchmarkMalloc16                  548          526   -4.01%
BenchmarkMallocTypeInfo8           659          642   -2.58%
BenchmarkMallocTypeInfo16          714          690   -3.36%

nothing interesting in the go1 benchmarks

alarm(~/go/test/bench/go1) % ~/go/misc/benchcmp {old,new}.txt
benchmark                         old ns/op    new ns/op    delta
BenchmarkBinaryTree17           52538428682  50789882119   -3.33%
BenchmarkFannkuch11             38953379120  38700417265   -0.65%
BenchmarkFmtFprintfEmpty                963         1002   +4.05%
BenchmarkFmtFprintfString              2944         2896   -1.63%
BenchmarkFmtFprintfInt                 2304         2308   +0.17%
BenchmarkFmtFprintfIntInt              3539         3524   -0.42%
BenchmarkFmtFprintfPrefixedInt         3806         3772   -0.89%
BenchmarkFmtFprintfFloat               8308         8102   -2.48%
BenchmarkFmtManyArgs                  14363        14108   -1.78%
BenchmarkGobDecode                139490656    135971956   -2.52%
BenchmarkGobEncode                 69002525     69725796   +1.05%
BenchmarkGzip                    6480846723   6443431928   -0.58%
BenchmarkGunzip                   856426485    849669849   -0.79%
BenchmarkHTTPClientServer            494461       489987   -0.90%
BenchmarkJSONEncode               690281602    678056181   -1.77%
BenchmarkJSONDecode              1587260111   1545816969   -2.61%
BenchmarkMandelbrot200             57216559     57190924   -0.04%
BenchmarkGoParse                   56833256     56244263   -1.04%
BenchmarkRegexpMatchEasy0_32           1265         1263   -0.16%
BenchmarkRegexpMatchEasy0_1K           5113         5103   -0.20%
BenchmarkRegexpMatchEasy1_32           1331         1323   -0.60%
BenchmarkRegexpMatchEasy1_1K          12895        12869   -0.20%
BenchmarkRegexpMatchMedium_32          2386         2393   +0.29%
BenchmarkRegexpMatchMedium_1K        802739       800649   -0.26%
BenchmarkRegexpMatchHard_32           42871        42378   -1.15%
BenchmarkRegexpMatchHard_1K         1320015      1302758   -1.31%
BenchmarkRevcomp                  125495203    125284630   -0.17%
BenchmarkTemplate                1478128539   1467233693   -0.74%
BenchmarkTimeParse                    12890        13032   +1.10%
BenchmarkTimeFormat                   13037        12213   -6.32%

benchmark                          old MB/s     new MB/s  speedup
BenchmarkGobDecode                     5.50         5.64    1.03x
BenchmarkGobEncode                    11.12        11.01    0.99x
BenchmarkGzip                          2.99         3.01    1.01x
BenchmarkGunzip                       22.66        22.84    1.01x
BenchmarkJSONEncode                    2.81         2.86    1.02x
BenchmarkJSONDecode                    1.22         1.26    1.03x
BenchmarkGoParse                       1.02         1.03    1.01x
BenchmarkRegexpMatchEasy0_32          25.28        25.32    1.00x
BenchmarkRegexpMatchEasy0_1K         200.27       200.66    1.00x
BenchmarkRegexpMatchEasy1_32          24.04        24.18    1.01x
BenchmarkRegexpMatchEasy1_1K          79.41        79.57    1.00x
BenchmarkRegexpMatchMedium_32          0.42         0.42    1.00x
BenchmarkRegexpMatchMedium_1K          1.28         1.28    1.00x
BenchmarkRegexpMatchHard_32            0.75         0.76    1.01x
BenchmarkRegexpMatchHard_1K            0.78         0.79    1.01x
BenchmarkRevcomp                      20.25        20.29    1.00x
BenchmarkTemplate                      1.31         1.32    1.01x

[khr, 2013-05-28 06:28:54]

LGTM

[dvyukov, 2013-05-28 07:06:05]

*** Submitted as https://code.google.com/p/go/source/detail?r=598c883c08b8 ***

runtime: inline MCache_Alloc() into mallocgc()
benchmark                    old ns/op    new ns/op    delta
BenchmarkMalloc8                    68           62   -8.63%
BenchmarkMalloc16                   75           69   -7.94%
BenchmarkMallocTypeInfo8           102           98   -3.73%
BenchmarkMallocTypeInfo16          108          103   -4.63%

R=golang-dev, dave, khr
CC=golang-dev
https://codereview.appspot.com/9790043