threads / discuss / 34381

expanding pack idx fanout table

Subject: expanding pack idx fanout table

## tl;dr

4 messages between Jul 8, 2013 and Jul 9, 2013.

replies: 3people: 3as markdown or json

Shawn Pearce· Jul 8, 2013, 15:54 UTC · lore

Has anyone studied the impact of converting the pack idx fanout table from 256 entries to 65536 entries?

Back of the envelope estimates for 3.1M objects in linux.git suggests a 2^16 fanout table would decrease the number of binary search iterations from ~14 to ~6. The increased table costs an extra 255 KiB of disk. On a 70M idx file this is noise.

I'm starting to wonder if increasing the fanout table once the object count is above a certain threshold is a reasonable optimization for larger repositories.

Junio C Hamano· Jul 8, 2013, 17:37 UTC · re: Shawn Pearce · lore

Re: expanding pack idx fanout table

Shawn Pearce <spearce@spearce.org> writes:

Show 11 quoted lines

> Has anyone studied the impact of converting the pack idx fanout table
> from 256 entries to 65536 entries?
>
> Back of the envelope estimates for 3.1M objects in linux.git suggests
> a 2^16 fanout table would decrease the number of binary search
> iterations from ~14 to ~6. The increased table costs an extra 255 KiB
> of disk. On a 70M idx file this is noise.
>
> I'm starting to wonder if increasing the fanout table once the object
> count is above a certain threshold is a reasonable optimization for
> larger repositories.

Yeah, and I do not think we have to be worried too much about backward compatibility for .idx files, as they are local and can be regenerated if an older version cannot read it.

I also wonder if we can generate a finer-grained fan-out table on the fly, perhaps lazily, without changing the on-disk format ;-)

Shawn Pearce· Jul 8, 2013, 17:58 UTC · re: Junio C Hamano · lore

Re: expanding pack idx fanout table

On Mon, Jul 8, 2013 at 10:37 AM, Junio C Hamano <gitster@pobox.com> wrote:

Show 20 quoted lines

> Shawn Pearce <spearce@spearce.org> writes:
>
>> Has anyone studied the impact of converting the pack idx fanout table
>> from 256 entries to 65536 entries?
>>
>> Back of the envelope estimates for 3.1M objects in linux.git suggests
>> a 2^16 fanout table would decrease the number of binary search
>> iterations from ~14 to ~6. The increased table costs an extra 255 KiB
>> of disk. On a 70M idx file this is noise.
>>
>> I'm starting to wonder if increasing the fanout table once the object
>> count is above a certain threshold is a reasonable optimization for
>> larger repositories.
>
> Yeah, and I do not think we have to be worried too much about
> backward compatibility for .idx files, as they are local and can be
> regenerated if an older version cannot read it.
>
> I also wonder if we can generate a finer-grained fan-out table on
> the fly, perhaps lazily, without changing the on-disk format ;-)

Heh. Yes. The reader at runtime could expand the 256 fanout table to a 65536 fanout table without changing the on disk format. Unfortunately doing the expansion will require O(N) time as the 2nd byte of each SHA-1 must be examined from every bucket. If you are going to perform O(N) lookups this expansion might save time as it lowers the "log N" bound of the O(N log N) algorithm for N lookups. It doesn't help enough when doing only N/20 lookups.

Jeff King· Jul 9, 2013, 04:19 UTC · re: Shawn Pearce · lore

Re: expanding pack idx fanout table

On Mon, Jul 08, 2013 at 08:54:24AM -0700, Shawn O. Pearce wrote:

Show 7 quoted lines

> Has anyone studied the impact of converting the pack idx fanout table
> from 256 entries to 65536 entries?
> 
> Back of the envelope estimates for 3.1M objects in linux.git suggests
> a 2^16 fanout table would decrease the number of binary search
> iterations from ~14 to ~6. The increased table costs an extra 255 KiB
> of disk. On a 70M idx file this is noise.

No, I hadn't considered it, but I think your analysis is correct that it would decrease the lookup cost. Having run "perf" on git a lot of times, I do see find_pack_entry_one turn up as a noticeable hot spot, but usually not more than a few percent.

Which kind of makes sense, because of the way we cache objects in memory. If you are doing something like `rev-list --objects`, you are going to do O(n) packfile lookups, but you end up doing _way_ more lookups in the object hash. Not just one per tree entry, but one per tree entry for each commit in which the entry was untouched.

I tried (maybe a year or so ago) to make the object hash faster using a similar fan-out trick, but I don't remember it having any real impact (because we keep our hash tables relatively collision free already). I think I also tried a full prefix-trie, with no luck.

Obviously the exact results would depend on your workload, but in general I'd expect the object hash to take the brunt of the load for repeated lookups, and for non-repeated lookups to be dominated by the time to actually access the object, as opposed to saving a few hashcmps. That may change as we start using the pack .idx for more clever things than simply accessing the objects, though (e.g., it might make a difference when coupled with my commit cache patches).

-Peff