Re: [PATCH] lookup_object(): Speed up 'git gc' by 12%, by reducing hash chain length

From: Ingo Molnar <mingo@elte.hu>
Date: Apr 29, 2011, 07:38 UTC
Message-ID: <20110429073825.GA16941@elte.hu>
In-Reply-To: <20110429072604.GA16371@elte.hu>

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

Show 8 quoted lines

> Find below a debug patch i use to run with a configurable spread.
> 
> Note, i just ran the patch on a different system and there the effect was 
> much less pronounced. So i'd prefer independent confirmation as well that it 
> speeds up things for others as well.
> 
> I'll run more numbers - maybe we are just very sensitive to the exact layout 
> of the object hash and a 16x spread created a different, more optimal layout.

Here are those numbers:

 $ for ((size=2; size<24; size++)); do printf "%5d: " $size; perf stat -e instructions:u -e cycles:u -e task-clock --sync --repeat 10 ./git --object-hash-spread $size gc 2>&1 | grep cycles; done

    2:      9,362,801,669 cycles:u                 #    2.982 GHz                      ( +-  0.25% )
    3:      9,464,946,158 cycles:u                 #    2.993 GHz                      ( +-  1.17% )
    4:      9,382,214,358 cycles:u                 #    2.981 GHz                      ( +-  0.26% )
    5:      9,373,537,954 cycles:u                 #    2.986 GHz                      ( +-  0.24% )
    6:      9,492,635,404 cycles:u                 #    2.988 GHz                      ( +-  1.25% )
    7:      9,427,037,835 cycles:u                 #    2.982 GHz                      ( +-  0.19% )
    8:      9,311,764,604 cycles:u                 #    2.987 GHz                      ( +-  0.23% )
    9:      9,384,331,920 cycles:u                 #    2.985 GHz                      ( +-  0.27% )
   10:      9,388,460,044 cycles:u                 #    2.983 GHz                      ( +-  0.31% )
   11:      9,374,380,165 cycles:u                 #    2.984 GHz                      ( +-  0.25% )
   12:      9,417,466,827 cycles:u                 #    2.984 GHz                      ( +-  0.27% )
   13:      9,348,550,619 cycles:u                 #    2.982 GHz                      ( +-  0.12% )
   14:      9,369,435,508 cycles:u                 #    2.982 GHz                      ( +-  0.31% )
   15:      9,361,127,598 cycles:u                 #    2.983 GHz                      ( +-  0.27% )
   16:      9,402,077,866 cycles:u                 #    2.987 GHz                      ( +-  0.20% )
   17:      9,390,950,850 cycles:u                 #    2.985 GHz                      ( +-  0.27% )
   18:      9,355,126,542 cycles:u                 #    2.986 GHz                      ( +-  0.30% )
   19:      9,357,143,371 cycles:u                 #    2.974 GHz                      ( +-  0.33% )
   20:      9,372,977,607 cycles:u                 #    2.985 GHz                      ( +-  0.34% )
   21:      9,355,406,722 cycles:u                 #    2.985 GHz                      ( +-  0.45% )
   22:      9,342,730,882 cycles:u                 #    2.982 GHz                      ( +-  0.31% )
   23:      9,372,321,792 cycles:u                 #    2.982 GHz                      ( +-  0.28% )

They are utterly unconvincing - there seems to be no improvement, it's all within noise.

Thanks,

	Ingo