Re: [PATCH v1] unpack-trees: avoid duplicate ODB lookups during checkout

[Stefan Beller <sbeller@google.com>]

On Thu, Apr 6, 2017 at 1:37 PM,  <git@jeffhostetler.com> wrote:
> From: Jeff Hostetler <jeffhost@microsoft.com>
>
> Teach traverse_trees_recursive() to not do redundant ODB
> lookups when both directories refer to the same OID.

And the reason for this is that omitting the second lookup
saves time, i.e. a lookup in the ODB of a sufficiently large
repo is slow.

My kneejerk line of thinking:
* yes, it sounds good to reduce the number of ODB accesses.
* But if we consider ODB lookups to be slow and we perform
  a structured access, how about a cache in front of the ODB?
* We already have that! (sort of..) 9a414486d9 (lookup_object:
  prioritize recently found objects, 2013-05-01)
* Instead of improving the caching, maybe change the
  size of the problem: We could keep the objects of different types
  in different hash-tables.

object.c has its own hash table, I presume for historical and
performance reasons, this would be split up to multiple hash
tables.

Additionally to "object *lookup_object(*sha1)", we'd have
a function

"object *lookup_object(*sha1, enum object_type hint)"
which looks into the correct the hash table.

If you were to call just  lookup_object with no hint, then you'd
look into all the different tables (I guess there is a preferrable
order in which to look, haven't thought about that).

>
> In operations such as read-tree, checkout, and merge when
> the differences between the commits are relatively small,
> there will likely be many directories that have the same
> SHA-1.  In these cases we can avoid hitting the ODB multiple
> times for the same SHA-1.

This would explain partially why there was such a good
performance boost in the referenced commit above as we
implicitly lookup the same object multiple times.

Peff is really into getting this part faster, c.f.
https://public-inbox.org/git/20160914235547.h3n2otje2hec6u7k@sigill.intra.peff.net/

> TODO This change is a first attempt to test that by comparing
> TODO the hashes of name[i] and name[i-i] and simply copying
> TODO the tree-descriptor data.  I was thinking of the n=2
> TODO case here.  We may want to extend this to the n=3 case.

>
> ================
> On the Windows repo (500K trees, 3.1M files, 450MB index),
> this reduced the overall time by 0.75 seconds when cycling
> between 2 commits with a single file difference.
>
> (avg) before: 22.699
> (avg) after:  21.955
> ===============

So it shaves off 4% of the time needed. it doesn't sound like a
break through, but I guess these small patches add up. :)

>         for (i = 0; i < n; i++, dirmask >>= 1) {
> -               const unsigned char *sha1 = NULL;
> -               if (dirmask & 1)
> -                       sha1 = names[i].oid->hash;
> -               buf[i] = fill_tree_descriptor(t+i, sha1);
> +               if (i > 0 && (dirmask & 1) && names[i].oid && names[i-1].oid &&
> +                       !hashcmp(names[i].oid->hash, names[i-1].oid->hash)) {

Why do we need to check for dirmask & 1 here?
This ought to be covered by the hashcmp already IIUC.
So maybe we can pull out the
                         if (dirmask & 1)
                                 sha1 = names[i].oid->hash;
out of the else when dropping that dirmask check?


Thanks,
Stefan