Re: [PATCH] git gc: Speed it up by 18% via faster hash comparisons

From: Ingo Molnar <mingo@elte.hu>
Date: Apr 28, 2011, 10:36 UTC
Message-ID: <20110428103650.GA18530@elte.hu>
In-Reply-To: <20110428063625.GB952@elte.hu>

* Ingo Molnar <mingo@elte.hu> wrote:

Show 13 quoted lines

> Having said that, it would be nice if someone could test these two patches on a 
> modern AMD box, using the perf stat from here:
> 
>   http://people.redhat.com/mingo/tip.git/README
> 
>   cd tools/perf/
>   make -j install
> 
> and do something like this to test git gc's performance:
> 
>   $ perf stat --sync --repeat 10 ./git gc
> 
> ... to see whether these speedups are generic, or somehow Intel CPU specific.

To follow up on this angle, i've now tested both 'git gc' performance improvement patches on an Opteron box:

 #
 # Before:
 #

 Performance counter stats for './git gc' (10 runs):

       6400.903957 task-clock               #    0.977 CPUs utilized            ( +-  0.09% )
             1,537 context-switches         #    0.000 M/sec                    ( +-  3.12% )
               118 CPU-migrations           #    0.000 M/sec                    ( +-  7.08% )
            40,879 page-faults              #    0.006 M/sec                    ( +-  0.03% )
    14,768,724,649 cycles                   #    2.307 GHz                      ( +-  0.09% )
     9,185,135,676 instructions             #    0.62  insns per cycle          ( +-  0.08% )
     1,883,860,739 branches                 #  294.312 M/sec                    ( +-  0.16% )
        99,800,253 branch-misses            #    5.30% of all branches          ( +-  0.34% )

        6.548348691  seconds time elapsed  ( +-  1.01% )

 #
 # After:
 #

 Performance counter stats for './git gc' (10 runs):

       6078.191492 task-clock               #    0.979 CPUs utilized            ( +-  0.09% )
             1,511 context-switches         #    0.000 M/sec                    ( +-  2.85% )
               104 CPU-migrations           #    0.000 M/sec                    ( +-  6.29% )
            41,410 page-faults              #    0.007 M/sec                    ( +-  0.04% )
    14,024,602,415 cycles                   #    2.307 GHz                      ( +-  0.09% )
     <not counted> stalled-cycles          
    11,364,757,635 instructions             #    0.81  insns per cycle          ( +-  0.06% )
     2,612,173,056 branches                 #  429.762 M/sec                    ( +-  0.09% )
       109,257,789 branch-misses            #    4.18% of all branches          ( +-  0.74% )

        6.209749362  seconds time elapsed  ( +-  0.51% )

So they bring an about +5.3% improvement.

Thanks,

	Ingo