patch, 3 partsindex-pack threading defaults

I decided to revisit index-pack's hard-coded default of 3 threads. And it turns out that higher numbers perform much better. :)

That value was determined experimentally in 2012. I'm not sure of the exact reason it's different now (modern processors are better at parallelism, or modern git is better at parallelism, or the original experiment was just a fluke). But regardless, I can get on the order of a two-to-one speedup by bumping the thread count. See the final patch for timings and more specific discussion.

The first two patches are just perf-test improvements (actually, the slowness to value ratio of p5302 from the first patch is why I was looking at this at all).

  [1/3]: p5302: disable thread-count parameter tests by default
  [2/3]: p5302: count up to online-cpus for thread tests
  [3/3]: index-pack: adjust default threading cap

 builtin/index-pack.c       | 19 ++++++++++++---
 t/perf/README              |  9 ++++++++
 t/perf/p5302-pack-index.sh | 47 +++++++++++++++++++-------------------
 t/perf/perf-lib.sh         |  2 ++
 4 files changed, 51 insertions(+), 26 deletions(-)

-Peff

The primary function of the perf suite is to detect regressions (or improvements) between versions of Git. The only numbers we show a direct comparison for are timings between the same test run on two different versions.

However, it can sometimes be used to collect other information. For instance, p5302 runs the same index-pack operation with different thread counts. The output doesn't directly compare these, but anybody interested in working on index-pack can manually compare the results.

For a normal regression run of the full perf-suite, though, this incurs a significant cost to generate numbers nobody will actually look at; about 25% of the total time of the test suite is spent in p5302. And the low-thread-count runs are the most expensive part of it, since they're (unsurprisingly) not using as many threads.

Let's skip these tests by default, but make it possible for people working on index-pack to still run them by setting an environment variable. Rather than make this specific to p5302, let's introduce a generic mechanism. This makes it possible to run the full suite with every possible test if somebody really wants to burn some CPU.

Signed-off-by: Jeff King <peff@peff.net>
---
 t/perf/README              |  9 +++++++++
 t/perf/p5302-pack-index.sh | 10 +++++-----
 t/perf/perf-lib.sh         |  2 ++
 3 files changed, 16 insertions(+), 5 deletions(-)

When PERF_EXTRA is enabled, p5302 checks the performance of index-pack with various numbers of threads. This can be useful for deciding what the default should be (which is currently capped at 3 threads based on the results of this script).

However, we only go up to 8 threads, and modern machines may have more. Let's get the number of CPUs from test-tool, and test various numbers of threads between one and that maximum.

Note that the current tests aren't all identical, as we have to set GIT_FORCE_THREADS for the --threads=1 test (which measures the overhead of starting a single worker thread versus the "0" case of using the main thread). To keep the loop simple, we'll keep the "0" case out of it, and set GIT_FORCE_THREADS=1 for all of the other cases (it's a noop for all but the "1" case, since numbers higher than 1 would always need threads).

Note also that we could skip running "test-tool" if PERF_EXTRA isn't set. However, there's some small value in knowing the number of threads, so that we can mark each test as skipped in the output.

Signed-off-by: Jeff King <peff@peff.net>
---
 t/perf/p5302-pack-index.sh | 47 +++++++++++++++++++-------------------
 1 file changed, 24 insertions(+), 23 deletions(-)

On Fri, Aug 21, 2020 at 01:54:51PM -0400, Jeff King wrote:

I meant to generate this diff with --patience, which is much more readable:

Commit b8a2486f15 (index-pack: support multithreaded delta resolving, 2012-05-06) describes an experiment that shows that setting the number of threads for index-pack higher than 3 does not help.

I repeated that experiment using a more modern version of Git and a more modern CPU and got different results.

Here are timings for p5302 against linux.git run on my laptop, a Core i9-9880H with 8 cores plus hyperthreading (so online-cpus returns 16):

  5302.3: index-pack 0 threads                   256.28(253.41+2.79)
  5302.4: index-pack 1 threads                   257.03(254.03+2.91)
  5302.5: index-pack 2 threads                   149.39(268.34+3.06)
  5302.6: index-pack 4 threads                   94.96(294.10+3.23)
  5302.7: index-pack 8 threads                   68.12(339.26+3.89)
  5302.8: index-pack 16 threads                  70.90(655.03+7.21)
  5302.9: index-pack default number of threads   116.91(290.05+3.21)

You can see that wall-clock times continue to improve dramatically up to the number of cores, but bumping beyond that (into hyperthreading territory) does not help (and in fact hurts a little).

Here's the same experiment on a machine with dual Xeon 6230's, totaling 40 cores (80 with hyperthreading):

  5302.3: index-pack 0 threads                    310.04(302.73+6.90)
  5302.4: index-pack 1 threads                    310.55(302.68+7.40)
  5302.5: index-pack 2 threads                    178.17(304.89+8.20)
  5302.6: index-pack 5 threads                    99.53(315.54+9.56)
  5302.7: index-pack 10 threads                   72.80(327.37+12.79)
  5302.8: index-pack 20 threads                   60.68(357.74+21.66)
  5302.9: index-pack 40 threads                   58.07(454.44+67.96)
  5302.10: index-pack 80 threads                  59.81(720.45+334.52)
  5302.11: index-pack default number of threads   134.18(309.32+7.98)

The results are similar; things stop improving at 40 threads. Curiously, going from 20 to 40 really doesn't help much, either (and increases CPU time considerably). So that may represent an actual barrier to parallelism, where we lose out due to context-switching and loss of cache locality, but don't reap the wall-clock benefits due to contention of our coarse-grained locks.

So what's a good default value? It's clear that the current cap of 3 is too low; our default values are 42% and 57% slower than the best times on each machine. The results on the 40-core machine imply that 20 threads is an actual barrier regardless of the number of cores, so we'll take that as a maximum. We get the best results on these machines at half of the online-cpus value. That's presumably a result of the hyperthreading. That's common on multi-core Intel processors, but not necessarily elsewhere. But if we take it as an assumption, we can perform optimally on hyperthreaded machines and still do much better than the status quo on other machines, as long as we never half below the current value of 3.

So that's what this patch does.

Signed-off-by: Jeff King <peff@peff.net>
---
Somebody with an AMD machine with a bunch of cores may want to revisit
this. I don't think it makes anything worse, but they may want to not do
the halving. I don't know if there's an easy way for us to determine the
difference via online_cpus(), though.

 builtin/index-pack.c | 19 ++++++++++++++++---
 1 file changed, 16 insertions(+), 3 deletions(-)

On Fri, Aug 21, 2020 at 1:58 PM Jeff King <peff@peff.net> wrote:

I'm not familiar with the index-pack machinery, so this response may be silly, but the first question which came to my mind was whether or not SSD vs. spinning-platter disk impacts these results, and which of the two you were using for the tests (which I don't think was mentioned in any of the commit messages). So, basically, I'm wondering about the implication of this change for those of us still stuck with old spinning-platter disks.

On Fri, Aug 21, 2020 at 02:08:55PM -0400, Eric Sunshine wrote:

They were both SSD machines, but it wouldn't matter for these tests because they easily fit the whole pack into memory anyway.

But in the general case, I don't think disk performance would be relevant. Delta resolution is very CPU-bound, because it's de-compressing data and then computing its SHA-1. So linux.git, for instance, is looking at ~1.3GB on disk that expands to 87.5GB of bytes to run through SHA-1.

And it would be pretty unlikely to hit the disk anyway, as the thing we primarily index is incoming packs which we've literally just written. So I'd expect them to be in cache.

Of course, if you can get different numbers from p5302, I'd be curious to hear them. :)

A more plausible downside might be that memory usage would increase as we operate on multiple deltas at once. But pack-objects is already much more hungry here, as it runs online_cpus() delta-compression threads simultaneously, each of which may have up to window_size entries in memory at once.

-Peff

On 2020-08-21 at 17:58:00, Jeff King wrote:

I was going to ask if we could make the halving conditional on x86_64, but it turns out POWER and UltraSPARC also have SMT, so that doesn't make sense. I expect that most users who have more than 6 "cores" are going to be on one of those systems or possibly ARM64, and since the performance penalty of using half as many cores isn't that significant, I'm not sure it's worth worrying about further. This will be an improvement regardless.

Which is just a long way of saying, this patch seems fine to me.

On Sat, Aug 22, 2020 at 01:16:07AM +0000, brian m. carlson wrote:

OK, good. :) I agree there may be room for more improvement on those systems. But lacking access to any, my goal was to make things better on systems I _could_ test on, and not make things worse on other systems. So I'd be very happy if people on other platforms (especially non-intel ones) wanted to run:

  cd t/perf
  GIT_PERF_EXTRA=1 \
  GIT_PERF_LARGE_REPO=/path/to/clone/of/linux.git \
  ./p5302-pack-index.sh

and report the results.

I do have a slightly-old AMD machine with 4 cores (an A8-7600). Here's what it says:

  5302.3: index-pack 0 threads                   447.67(436.62+6.57)
  5302.4: index-pack 1 threads                   450.80(441.26+7.20)
  5302.5: index-pack 2 threads                   265.62(459.56+7.30)
  5302.6: index-pack 4 threads                   177.06(477.56+8.22)
  5302.7: index-pack default number of threads   202.60(473.15+7.61)

So it does get better with 4 threads (but we continue to cap it at 3). I wonder whether we should just do:

On Mon, Aug 24, 2020 at 01:37:35PM -0400, Jeff King wrote:

I tested on my 9.5 year old 4-core iMac with 2.5 GHz Intel Core i5, spinning platter hard disk, and 12 GB memory (recently upgraded from 4 GB). I used git.git rather than linux.git because I didn't want to wait several days for the results (plus this is my primary machine which I'm actively using). Results were similar to what you saw:

  5302.3: index-pack 0 threads                   14.85(14.10+0.66)
  5302.4: index-pack 1 threads                   14.65(13.83+0.63)
  5302.5: index-pack 2 threads                   9.11(14.40+0.80) 
  5302.6: index-pack 4 threads                   6.89(17.03+1.32) 
  5302.7: index-pack default number of threads   7.72(16.15+1.15) 

Thanks for providing exact instructions for running the test; I had never done it before, thus didn't know how.

On Fri, Aug 21, 2020 at 01:51:53PM -0400, Jeff King wrote:

After writing a response elsewhere in the thread, it occurred to me that a good candidate for explaining this may be that our modern sha1dc implementation is way slower than what we were using in 2012 (which would have been either block-sha1, or the even-faster openssl implementation). And since a good chunk of index-pack's time is going to computing sha1 hashes on the resulting objects, that means that since 2012, we're spending relatively more time in the hash computation (which parallelizes per-object) and less time in the other parts that happen under a lock.

-Peff

Jeff King <peff@peff.net> writes:

Believable conjecture that is. You could benchmark again with block-sha1 on today's hardware, but because the performance profile with sha1dc is what matters in the real world anyway...

Thanks.

On Fri, Aug 21, 2020 at 11:59:58AM -0700, Junio C Hamano wrote:

Yeah, I agree on the "real world" part, but I'm the curious sort, so here are numbers compiled against openssl (which is generally even faster than block-sha1, and would thus emphasize the results of our hypothesis):

  5302.3: index-pack 0 threads                   108.78(106.39+2.31)
  5302.4: index-pack 1 threads                   110.65(108.08+2.49)
  5302.5: index-pack 2 threads                   67.57(110.83+2.75) 
  5302.6: index-pack 4 threads                   48.18(123.82+3.02) 
  5302.7: index-pack 8 threads                   39.07(153.45+4.13) 
  5302.8: index-pack 16 threads                  38.38(265.78+7.71) 
  5302.9: index-pack default number of threads   54.64(117.35+2.73)

So it's actually pretty similar. Things continue getting faster as we go past 3 threads. Though our total improvement is less (29% better with 8 threads compared to 3, versus 42% better when using sha1dc). So I think it's _part_ of the reason, but not all of it.

-Peff