The fundamental data structure of the delta base cache is a
hash table mapping pairs of "(packfile, offset)" into
structs containing the actual object data. The hash table
implementation dates back to e5e0161 (Implement a simple
delta_base cache, 2007-03-17), and uses a fixed-size table.
The current size is a hard-coded 256 entries.
Because we need to be able to remove objects from the hash
table, entry lookup does not do any kind of probing to
handle collisions. Colliding items simply replace whatever
is in their slot. As a result, we have fewer usable slots
than even the 256 we allocate. At half full, each new item
has a 50% chance of displacing another one. Or another way
to think about it: every item has a 1/256 chance of being
ejected due to hash collision, without regard to our LRU
strategy.
So it would be interesting to see the effect of increasing
the cache size on the runtime for some common operations. As
with the previous patch, we'll measure "git log --raw" for
tree-only operations, and "git log -Sfoo --raw" for
operations that touch trees and blobs. All times are
wall-clock best-of-3, done against fully packed repos with
--depth=50, and the default core.deltaBaseCacheLimit of
96MB.
Here are timings for various values of MAX_DELTA_CACHE
against git.git (the asterisk marks the minimum time for
each operation):
MAX_DELTA_CACHE log-raw log-S
--------------- --------- ---------
256 0m02.227s 0m12.821s
512 0m02.143s 0m10.602s
1024 0m02.127s 0m08.642s
2048 0m02.148s 0m07.123s
4096 0m02.194s 0m06.448s*
8192 0m02.239s 0m06.504s
16384 0m02.144s* 0m06.502s
32768 0m02.202s 0m06.622s
65536 0m02.230s 0m06.677s
The log-raw case isn't changed much at all here (probably
because our trees just aren't that big in the first place,
or possibly because we have so _few_ trees in git.git that
the 256-entry cache is enough). But once we start putting
blobs in the cache, too, we see a big improvement (almost
50%). The curve levels off around 4096, which means that we
can hold about that many entries before hitting the 96MB
memory limit (or possibly that the workload is small enough
that there is simply no more work to be optimized out by
caching more).
(As a side note, I initially timed my existing git.git pack,
which was a base of --aggressive combined with some pulls on
top. So it had quite a few deeper delta chains. The
256-cache case was more like 15s, and it still dropped to
~6.5s in the same way).
Here are the timings for linux.git:
MAX_DELTA_CACHE log-raw log-S
--------------- --------- ---------
256 0m41.661s 5m12.410s
512 0m39.547s 5m07.920s
1024 0m37.054s 4m54.666s
2048 0m35.871s 4m41.194s*
4096 0m34.646s 4m51.648s
8192 0m33.881s 4m55.342s
16384 0m35.190s 5m00.122s
32768 0m35.060s 4m58.851s
65536 0m33.311s* 4m51.420s
As we grow we see a nice 20% speedup in the tree traversal,
and more modest 10% in the log-S. This is probably an
indication that we are bound less by the number of entries,
and more by the memory limit (more on that below). What is
interesting is that the numbers bounce around a bit;
increasing the number of entries isn't always a strict
improvement.
Partially this is due to noise in the measurement. But it
may also be an indication that our LRU ejection scheme is
not optimal. The smaller cache sizes introduce some
randomness into the ejection (due to collisions), which may
sometimes work in our favor (and sometimes not!).
So what is the optimal setting of MAX_DELTA_CACHE? The
"bouncing" in the linux.git log-S numbers notwithstanding,
it mostly seems like bigger is better. And even if we were
to try to find a "sweet spot", these are just two
repositories, that are not necessarily representative. The
shape of history, the size of trees and blobs, the memory
limit configuration, etc, all will affect the outcome.
Rather than trying to find the "right" number, another
strategy is to just switch to a hash table that can actually
store collisions: namely our hashmap.h implementation.
Here are numbers for that compared to the "best" we saw from
adjusting MAX_DELTA_CACHE:
| log-raw | log-S
| best hashmap | best hashmap
| --------- --------- | --------- ---------
git | 0m02.144s 0m02.144s | 0m06.448s 0m06.688s
linux | 0m33.311s 0m33.092s | 4m41.194s 4m57.172s
We can see the results are similar in most cases, which is
what we'd expect. We're not ejecting due to collisions at
all, so this is purely representing the LRU. So really, we'd
expect this to model most closely the larger values of the
static MAX_DELTA_CACHE limit. And that does seem to be
what's happening, including the "bounce" in the linux log-S
case.
So while the value for that case _isn't_ as good as the
optimal one measured above (which was 2048 entries), given
the bouncing I'm hesitant to suggest that 2048 is any kind
of optimum (not even for linux.git, let alone as a general
rule). The generic hashmap has the appeal that it drops the
number of tweakable numbers by one, which means we can focus
on tuning other elements, like the LRU strategy or the
core.deltaBaseCacheLimit setting.
And indeed, if we bump the cache limit to 1G (which is
probably silly for general use, but maybe something people
with big workstations would want to do), the linux.git log-S
time drops to 3m32s. That's something you really _can't_ do
easily with the static hash table, because the number of
entries needs to grow in proportion to the memory limit (so
2048 is almost certainly not going to be the right value
there).
This patch takes that direction, and drops the static hash
table entirely in favor of using the hashmap.h API.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Git - fast, scalable, distributed revision control system
Git is a fast, scalable, distributed revision control system with an
unusually rich command set that provides both high-level operations
and full access to internals.
Git is an Open Source project covered by the GNU General Public
License version 2 (some parts of it are under different licenses,
compatible with the GPLv2). It was originally written by Linus
Torvalds with help of a group of hackers around the net.
Please read the file INSTALL for installation instructions.
Many Git online resources are accessible from http://git-scm.com/
including full documentation and Git related tools.
See Documentation/gittutorial.txt to get started, then see
Documentation/giteveryday.txt for a useful minimum set of commands, and
Documentation/git-.txt for documentation of each command.
If git has been correctly installed, then the tutorial can also be
read with man gittutorial or git help tutorial, and the
documentation of each command with man git-<commandname> or git help <commandname>.
CVS users may also want to read Documentation/gitcvs-migration.txt
(man gitcvs-migration or git help cvs-migration if git is
installed).
The maintainer frequently sends the "What's cooking" reports that
list the current status of various development topics to the mailing
list. The discussion following them give a good reference for
project status, development direction and remaining tasks.
The name "git" was given by Linus Torvalds when he wrote the very
first version. He described the tool as "the stupid content tracker"
and the name as (depending on your mood):
random three-letter combination that is pronounceable, and not
actually used by any common UNIX command. The fact that it is a
mispronunciation of "get" may or may not be relevant.
stupid. contemptible and despicable. simple. Take your pick from the
dictionary of slang.
"global information tracker": you're in a good mood, and it actually
works for you. Angels sing, and a light suddenly fills the room.
"goddamn idiotic truckload of sh*t": when it breaks
8261e1f139