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

From: Jeff Hostetler <git@jeffhostetler.com>
Date: Apr 7, 2017, 13:51 UTC
Message-ID: <fa94cfb3-ac5e-13f1-4c06-f0f3f1a2d1df@jeffhostetler.com>
In-Reply-To: <20170407051935.4usaxtmco7lqm4ew@sigill.intra.peff.net>

On 4/7/2017 1:19 AM, Jeff King wrote:

Show 38 quoted lines

> On Thu, Apr 06, 2017 at 03:48:07PM -0700, Stefan Beller wrote:
>
>> 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.
>
> I don't think this is using lookup_object() at all, though. It goes
> straight to fill_tree_descriptor(), which will read the object contents
> from disk. So our time is going to:
>
>   1. Finding the object in the odb (ideally a binary search in a single
>      pack index, but less optimal when there are many packs).
>
>   2. Reconstructing the object. This means zlib-inflating from disk, but
>      it may also mean delta reconstruction.
>
> I think there _are_ some caches at play here, though, when you look up
> the same tree back to back. The pack-mru means that we'll always look in
> the correct pack first. And in theory the delta base cache means that
> we'll already have the whole thing reconstructed in memory (though I
> have often been confused by exactly when we put items into that cache,
> so I might be wrong).

This change shaved 500K calls to fill_tree_descriptor() (on the Windows source tree on a "checkout -b" to the same commit). I was surprised it didn't give more of a speed up, so some of that caching may be in play here, but it's hard to tell.

Also, on my repo I have 100GB+ of packfiles, so that may be messing with things a bit.

Show 8 quoted lines

>
> So in theory, this is not all that different than the "just allocate and
> copy the bytes" optimization that's happening here (though I'm not
> surprised that doing it at a higher level can produce some speedup).
>
> I think the more interesting optimization is "just use the same buffer
> without bothering to copy", which is hard for the low-level code to do
> (since it doesn't know what lifetime the caller is expecting).

My first draft did just borrow the buffer right there in traverse_ and it seemed to work just fine. I was being cautious and copied it properly in the lower layer. Since the bufs are freed at the bottom it felt safe, but I didn't want to overlook anything. I'll switch it back.

Show 12 quoted lines

>
>> object.c has its own hash table, I presume for historical and
>> performance reasons, this would be split up to multiple hash
>> tables.
>
> So I don't think lookup_object() is really relevant here. But I'm also
> not sure that multiple hash tables would really buy us much. In theory
> hash tables are O(1), so multiple smaller tables doesn't help (and might
> hurt, since now we have four O(1) lookups to do). Of course that's a
> minor fiction because of collisions, but we keep the load factor on the
> object.c table pretty low (and that's why things like quadratic probing
> and cuckoo hashing never showed great improvement).

I do wonder now about the initial hash table size and any limits on it, but that is a question for another day.

Thanks Jeff