From: Jonathan Nieder <jrnieder@gmail.com>
Date: Tue, 23 Nov 2010 23:17:18 GMT
Subject: Re: [PATCH/RFC] fast-import: insert new object entries at start of hash bucket
Message-ID: <20101123231718.GA4317@burratino>
In-Reply-To: <AANLkTikqUjjjMRzWTcEOs+2PGu=-9VVbdn0YgpabFaDu@mail.gmail.com>

Shawn Pearce wrote:
> On Mon, Nov 22, 2010 at 11:53 PM, Jonathan Nieder <jrnieder@gmail.com> wrote:

>> Other aspects to investigate: choice of hash function;
>
> Why?  SHA-1 is pretty uniform in its distribution.

I got distracted for a moment by the atom table, but since that does
not have a big effect on performance it's probably not worth spending
time on.  Sorry about that; please ignore.

[...]
>                                                           The way I
> read this store_tree() code, every subdirectory is recursed into even
> if no modifications were made inside of that subdirectory during the
> current commit.

Doesn't the is_null_sha1 check avoid that?

To further explain the workload: svn-fe receives its blobs from svn
in the form of deltas.  So the conversation might go like this:

	S	commit refs/heads/master
	S	mark :10000
	S	committer felicity <felicity@local>
	S	data 74
	S	bug 3097: switch spamd from doing 'fork per message' to a 'prefork' model
	S	cat incubator/spamassassin/trunk/spamd/spamd.raw
	F	89d56462577b8b7b4f4115f2a47f0b3da22b791a blob 63633
	F	#!/usr/bin/perl -w -T
	...
	S	M 100644 inline incubator/spamassassin/trunk/spamd/spamd.raw
	S	data 62114
	...

Current svn-fe in vcs-svn-pu requests the preimage blobs using marks,
but the idea is the same.

If this proves a bottleneck I suppose we could cache the content of
frequently-requested old blobs and keep pointers to that in the
in-core tree.