Re: thin packs ending up fat

[Jeff King <peff@peff.net>]

On Thu, Jan 12, 2012 at 05:15:23PM -0500, Jeff King wrote:

> It turns out that when packing a subset of a fully packed repo (as we do
> for a bundle or for a fetch), we tend not to make thin packs at all.
> The culprit is this logic in try_delta:
> 
>         /*
>          * We do not bother to try a delta that we discarded
>          * on an earlier try, but only when reusing delta data.
>          */
>         if (reuse_delta && trg_entry->in_pack &&
>             trg_entry->in_pack == src_entry->in_pack &&
>             trg_entry->in_pack_type != OBJ_REF_DELTA &&
>             trg_entry->in_pack_type != OBJ_OFS_DELTA)
>                 return 0;
> [...]
> Maybe it is enough to simply turn off this optimization if the potential
> delta source is not being included in the pack (i.e., we are using
> --thin and it is a boundary object). Because if both objects are being
> sent, we will just end up reusing the delta that goes in the reverse
> direction anyway.

Hmm. It turns out this is really easy, because we have already marked
such objects as preferred bases.

So with this patch:

diff --git a/builtin/pack-objects.c b/builtin/pack-objects.c
index 96c1680..d05e228 100644
--- a/builtin/pack-objects.c
+++ b/builtin/pack-objects.c
@@ -1439,6 +1439,7 @@ static int try_delta(struct unpacked *trg, struct unpacked *src,
 	 */
 	if (reuse_delta && trg_entry->in_pack &&
 	    trg_entry->in_pack == src_entry->in_pack &&
+	    !src_entry->preferred_base &&
 	    trg_entry->in_pack_type != OBJ_REF_DELTA &&
 	    trg_entry->in_pack_type != OBJ_OFS_DELTA)
 		return 0;

here are the numbers I get:

                  dataset
            | fetches | tags
---------------------------------
     before | 53358   | 2750977
size  after | 32398   | 2668479
     change |   -39%  |      -3%
---------------------------------
     before |  0.18   | 1.12
CPU   after |  0.18   | 1.15
     change |    +0%  |      +3%

So nearly all of the size benefit, but very little CPU change (even the
3% on the larger-pack case is close to the levels of run-to-run noise).
Obviously the size benefit in the larger-pack case isn't impressive, but
I think the "fetches" case is much more indicative of a real server
load.

-Peff