threads / discuss / 31661

Re: Commit cache to speed up rev-list and merge

Subject: Re: Commit cache to speed up rev-list and merge

## tl;dr

10 messages between Sep 27, 2012 and Oct 1, 2012.

replies: 9people: 3as markdown or json

Shawn Pearce· Sep 27, 2012, 15:51 UTC · lore
On Thu, Sep 27, 2012 at 5:17 AM, Nguyen Thai Ngoc Duy <pclouds@gmail.com> wrote:
> I'd like to see some sort of extension mechanism like in
> $GIT_DIR/index, so that we don't have to increase pack index version
> often. What I have in mind is optional commit cache to speed up
> rev-list and merge, which could be stored in pack index too.
Can you share some of your ideas?

In Linus' Linux kernel tree there are currently about 323,178 commits. If we store just the pre-parsed commit time as an int32 field this is an additional 1.2 MiB of data in the pack-*.idx file, assuming we can use additional data like pack offset position to correlate commit to the parsed int. If we stored parent pointers in a similar way you probably need at least 3.6 MiB of additional disk space on the index. For example, use 12 bytes for each commit to store enough of the parsed commit time to sort commits, and up to 2 parent pointers per commit.... with a reserved magic value for octopus merges to mean the commit itself has to be parsed to get the graph structure correct.

Jeff King· Sep 27, 2012, 17:39 UTC · re: Shawn Pearce · lore
On Thu, Sep 27, 2012 at 08:51:51AM -0700, Shawn O. Pearce wrote:
Show 7 quoted lines
> On Thu, Sep 27, 2012 at 5:17 AM, Nguyen Thai Ngoc Duy <pclouds@gmail.com> wrote:
> > I'd like to see some sort of extension mechanism like in
> > $GIT_DIR/index, so that we don't have to increase pack index version
> > often. What I have in mind is optional commit cache to speed up
> > rev-list and merge, which could be stored in pack index too.
> 
> Can you share some of your ideas?
Some of it is here:
  http://article.gmane.org/gmane.comp.version-control.git/203308
-Peff
Shawn Pearce· Sep 27, 2012, 17:45 UTC · re: Jeff King · lore
On Thu, Sep 27, 2012 at 10:39 AM, Jeff King <peff@peff.net> wrote:
Show 12 quoted lines
> On Thu, Sep 27, 2012 at 08:51:51AM -0700, Shawn O. Pearce wrote:
>> On Thu, Sep 27, 2012 at 5:17 AM, Nguyen Thai Ngoc Duy <pclouds@gmail.com> wrote:
>> > I'd like to see some sort of extension mechanism like in
>> > $GIT_DIR/index, so that we don't have to increase pack index version
>> > often. What I have in mind is optional commit cache to speed up
>> > rev-list and merge, which could be stored in pack index too.
>>
>> Can you share some of your ideas?
>
> Some of it is here:
>
>   http://article.gmane.org/gmane.comp.version-control.git/203308
Quoting from that patch:
On  2012-08-12 Nguyen Thai Ngoc Duy <pclouds@gmail.com> wrote:
> Long term we might gain slight lookup speedup if we know object type
> as search region is made smaller. But for that to happen, we need to
> propagate object type hint down to find_pack_entry_one() and friends.
> Possible thing to do, I think.

I'm not sure reclustering the index by object type is going to make a worthwhile difference. Of 2.2m objects in the Linux tree, 320k are commits. The difference between doing the binary search through all objects vs. just commits is only 2 iterations more of binary search if we assume the per-type ranges have their own fan-out tables.

Show 5 quoted lines
> The main reason to group objects by type is to make it possible to
> create another sha1->something mapping for a particular object type,
> without wasting space for storing sha-1 keys again. For example, we
> can store commit caches, tree caches... at the end of the index as
> extensions.

Using ordinal position in the pack also works, and doesn't require clustering objects by type.

Jeff King· Sep 27, 2012, 18:32 UTC · re: Shawn Pearce · lore
On Thu, Sep 27, 2012 at 10:45:32AM -0700, Shawn O. Pearce wrote:
Show 11 quoted lines
> On  2012-08-12 Nguyen Thai Ngoc Duy <pclouds@gmail.com> wrote:
> > Long term we might gain slight lookup speedup if we know object type
> > as search region is made smaller. But for that to happen, we need to
> > propagate object type hint down to find_pack_entry_one() and friends.
> > Possible thing to do, I think.
> 
> I'm not sure reclustering the index by object type is going to make a
> worthwhile difference. Of 2.2m objects in the Linux tree, 320k are
> commits. The difference between doing the binary search through all
> objects vs. just commits is only 2 iterations more of binary search if
> we assume the per-type ranges have their own fan-out tables.

To me the big win would be implicit indexing for items that are present for every instance of a particular object type. So if we wanted to keep the timestamp for every commit, you could have a "pack-*.timestamps" that is literally just a packed list of uint32's, one per commit, where the position of a commit's timestamp in the list is the same as its position in the index of sha1s in the pack index.

That's simple to do if your index is just commits. But if it includes all objects, then your list is sparse. So either you waste space by making an empty slot for the non-commit objects, or you have an extra level of indirection mapping the commit into the packed list, which is going to double the storage in this case (though you could reuse that extra mapping for the parent, generation number, etc, so it at least gets amortized as you store more data). Or is there some clever solution I'm missing?

For your extension, I don't think it matters. You're sparse even in the commit-object space, so you have to store the mapping anyway. And your data is big enough that the overhead isn't too painful.

-Peff
Nguyen Thai Ngoc Duy· Sep 28, 2012, 01:43 UTC · re: Jeff King · lore
On Fri, Sep 28, 2012 at 12:39 AM, Jeff King <peff@peff.net> wrote:
Show 13 quoted lines
> On Thu, Sep 27, 2012 at 08:51:51AM -0700, Shawn O. Pearce wrote:
>
>> On Thu, Sep 27, 2012 at 5:17 AM, Nguyen Thai Ngoc Duy <pclouds@gmail.com> wrote:
>> > I'd like to see some sort of extension mechanism like in
>> > $GIT_DIR/index, so that we don't have to increase pack index version
>> > often. What I have in mind is optional commit cache to speed up
>> > rev-list and merge, which could be stored in pack index too.
>>
>> Can you share some of your ideas?
>
> Some of it is here:
>
>   http://article.gmane.org/gmane.comp.version-control.git/203308
And an experiment was done earlier
http://thread.gmane.org/gmane.comp.version-control.git/194306/focus=194596
-- 
Duy
Nguyen Thai Ngoc Duy· Sep 28, 2012, 02:14 UTC · re: Shawn Pearce · lore
On Thu, Sep 27, 2012 at 10:51 PM, Shawn Pearce <spearce@spearce.org> wrote:
Show 10 quoted lines
> In Linus' Linux kernel tree there are currently about 323,178 commits.
> If we store just the pre-parsed commit time as an int32 field this is
> an additional 1.2 MiB of data in the pack-*.idx file, assuming we can
> use additional data like pack offset position to correlate commit to
> the parsed int. If we stored parent pointers in a similar way you
> probably need at least 3.6 MiB of additional disk space on the index.
> For example, use 12 bytes for each commit to store enough of the
> parsed commit time to sort commits, and up to 2 parent pointers per
> commit.... with a reserved magic value for octopus merges to mean the
> commit itself has to be parsed to get the graph structure correct.

This is much better than my naive approach (storing sha-1 and timestamps). We could use less space by storing parent pointer of non-merge commits only. Merge commits linux-2.6 is 6% the number of commits. git.git has higher percentage, 21%. I bet many projects do not merge as much and the number of merge commits is less than 5%.

-- 
Duy
Shawn Pearce· Oct 1, 2012, 01:49 UTC · re: Nguyen Thai Ngoc Duy · lore
On Thu, Sep 27, 2012 at 7:14 PM, Nguyen Thai Ngoc Duy <pclouds@gmail.com> wrote:
Show 17 quoted lines
> On Thu, Sep 27, 2012 at 10:51 PM, Shawn Pearce <spearce@spearce.org> wrote:
>> In Linus' Linux kernel tree there are currently about 323,178 commits.
>> If we store just the pre-parsed commit time as an int32 field this is
>> an additional 1.2 MiB of data in the pack-*.idx file, assuming we can
>> use additional data like pack offset position to correlate commit to
>> the parsed int. If we stored parent pointers in a similar way you
>> probably need at least 3.6 MiB of additional disk space on the index.
>> For example, use 12 bytes for each commit to store enough of the
>> parsed commit time to sort commits, and up to 2 parent pointers per
>> commit.... with a reserved magic value for octopus merges to mean the
>> commit itself has to be parsed to get the graph structure correct.
>
> This is much better than my naive approach (storing sha-1 and
> timestamps). We could use less space by storing parent pointer of
> non-merge commits only. Merge commits linux-2.6 is 6% the number of
> commits. git.git has higher percentage, 21%. I bet many projects do
> not merge as much and the number of merge commits is less than 5%.

Some projects merge quite often. Android's frameworks/base repository has a very large number of merges. Out of 79905 commits reachable from the master branch, 65.3% are merges. So actually there are more merge commits in the Android history than there are code commits. A cache of only non-merges may be worthless on such a history.

Nguyen Thai Ngoc Duy· Oct 1, 2012, 02:05 UTC · re: Shawn Pearce · lore
On Mon, Oct 1, 2012 at 8:49 AM, Shawn Pearce <spearce@spearce.org> wrote:
Show 24 quoted lines
> On Thu, Sep 27, 2012 at 7:14 PM, Nguyen Thai Ngoc Duy <pclouds@gmail.com> wrote:
>> On Thu, Sep 27, 2012 at 10:51 PM, Shawn Pearce <spearce@spearce.org> wrote:
>>> In Linus' Linux kernel tree there are currently about 323,178 commits.
>>> If we store just the pre-parsed commit time as an int32 field this is
>>> an additional 1.2 MiB of data in the pack-*.idx file, assuming we can
>>> use additional data like pack offset position to correlate commit to
>>> the parsed int. If we stored parent pointers in a similar way you
>>> probably need at least 3.6 MiB of additional disk space on the index.
>>> For example, use 12 bytes for each commit to store enough of the
>>> parsed commit time to sort commits, and up to 2 parent pointers per
>>> commit.... with a reserved magic value for octopus merges to mean the
>>> commit itself has to be parsed to get the graph structure correct.
>>
>> This is much better than my naive approach (storing sha-1 and
>> timestamps). We could use less space by storing parent pointer of
>> non-merge commits only. Merge commits linux-2.6 is 6% the number of
>> commits. git.git has higher percentage, 21%. I bet many projects do
>> not merge as much and the number of merge commits is less than 5%.
>
> Some projects merge quite often. Android's frameworks/base repository
> has a very large number of merges. Out of 79905 commits reachable from
> the master branch, 65.3% are merges. So actually there are more merge
> commits in the Android history than there are code commits. A cache of
> only non-merges may be worthless on such a history.

The good thing about these cache is it's configurable. Merge-preferred projects can choose to cache the first two parents. Non-merge projects can choose to cache just the first parent. We don't need a fixed format for both.

-- 
Duy
Shawn Pearce· Oct 1, 2012, 02:27 UTC · re: Nguyen Thai Ngoc Duy · lore
On Sun, Sep 30, 2012 at 7:05 PM, Nguyen Thai Ngoc Duy <pclouds@gmail.com> wrote:
Show 30 quoted lines
> On Mon, Oct 1, 2012 at 8:49 AM, Shawn Pearce <spearce@spearce.org> wrote:
>> On Thu, Sep 27, 2012 at 7:14 PM, Nguyen Thai Ngoc Duy <pclouds@gmail.com> wrote:
>>> On Thu, Sep 27, 2012 at 10:51 PM, Shawn Pearce <spearce@spearce.org> wrote:
>>>> In Linus' Linux kernel tree there are currently about 323,178 commits.
>>>> If we store just the pre-parsed commit time as an int32 field this is
>>>> an additional 1.2 MiB of data in the pack-*.idx file, assuming we can
>>>> use additional data like pack offset position to correlate commit to
>>>> the parsed int. If we stored parent pointers in a similar way you
>>>> probably need at least 3.6 MiB of additional disk space on the index.
>>>> For example, use 12 bytes for each commit to store enough of the
>>>> parsed commit time to sort commits, and up to 2 parent pointers per
>>>> commit.... with a reserved magic value for octopus merges to mean the
>>>> commit itself has to be parsed to get the graph structure correct.
>>>
>>> This is much better than my naive approach (storing sha-1 and
>>> timestamps). We could use less space by storing parent pointer of
>>> non-merge commits only. Merge commits linux-2.6 is 6% the number of
>>> commits. git.git has higher percentage, 21%. I bet many projects do
>>> not merge as much and the number of merge commits is less than 5%.
>>
>> Some projects merge quite often. Android's frameworks/base repository
>> has a very large number of merges. Out of 79905 commits reachable from
>> the master branch, 65.3% are merges. So actually there are more merge
>> commits in the Android history than there are code commits. A cache of
>> only non-merges may be worthless on such a history.
>
> The good thing about these cache is it's configurable. Merge-preferred
> projects can choose to cache the first two parents. Non-merge projects
> can choose to cache just the first parent. We don't need a fixed
> format for both.

Git has enough magic switches. It doesn't need yet another magic switch that one group of users needs to set, and another can safely ignore because their project's usage just happens to align with Linus Torvald's current world view.

Nguyen Thai Ngoc Duy· Oct 1, 2012, 05:16 UTC · re: Shawn Pearce · lore
On Mon, Oct 1, 2012 at 9:27 AM, Shawn Pearce <spearce@spearce.org> wrote:
> Git has enough magic switches. It doesn't need yet another magic
> switch that one group of users needs to set, and another can safely
> ignore because their project's usage just happens to align with Linus
> Torvald's current world view.

I see it as tuning, not switching. It's like setting the number of commits where bitmaps are taken. We can see the commit cache as a table, where columns are commit properties. We have a column for date, one for the first parent. Users can choose to have a third column for the second parent, or another one for root tree sha-1. The implementation could be made generic to support caching any <n> sha1-columns.

-- 
Duy

← back to recent threads