On 11/15/20 2:43 PM, Matheus Tavares wrote:
Hi, everyone
I've been testing the parallel checkout code on some different machines,
to benchmark its performance against the sequential version. As
discussed in [1], the biggest speedups, on Linux, are usually seen on
SSDs and NFS volumes. (I haven't got the chance to benchmark on
Windows or OSX yet.)
Regarding NFS, I got some 2~3.4x speedups even when the NFS client and
server were both running on single-core machines. Here are some runtimes
for a linux-v5.8 clone (means of 5 cold-cache executions):
nfs 3.0 nfs 4.0 nfs 4.1
1: 183.708 s ± 3.290 s 205.851 s ± 0.844 s 217.317 s ± 3.047 s
2: 130.510 s ± 3.917 s 139.124 s ± 0.772 s 142.963 s ± 0.765 s
4: 89.611 s ± 1.032 s 102.701 s ± 1.665 s 94.728 s ± 1.014 s
8: 68.097 s ± 0.820 s 104.914 s ± 1.239 s 69.359 s ± 0.619 s
16: 63.999 s ± 0.820 s 104.808 s ± 2.279 s 64.843 s ± 0.587 s
32: 62.316 s ± 2.095 s 102.105 s ± 1.537 s 64.122 s ± 0.374 s
64: 63.699 s ± 0.841 s 103.103 s ± 1.319 s 63.532 s ± 0.734 s
The parallel version was also faster for some smaller checkouts. For
example, the following numbers come from a bat-v0.16.0 clone
(251 files, ~3MB):
nfs 3.0 nfs 4.0 nfs 4.1
1: 0.853 s ± 0.080 s 0.814 s ± 0.020 s 0.876 s ± 0.065 s
2: 0.671 s ± 0.020 s 0.702 s ± 0.030 s 0.705 s ± 0.030 s
4: 0.530 s ± 0.024 s 0.595 s ± 0.020 s 0.570 s ± 0.030 s
8: 0.470 s ± 0.033 s 0.609 s ± 0.025 s 0.510 s ± 0.031 s
16: 0.469 s ± 0.037 s 0.616 s ± 0.022 s 0.513 s ± 0.030 s
32: 0.487 s ± 0.030 s 0.639 s ± 0.018 s 0.527 s ± 0.028 s
64: 0.520 s ± 0.022 s 0.680 s ± 0.028 s 0.562 s ± 0.026 s
While looking at these numbers with Geert (in CC), he had the idea that
we could try to detect when the checkout path is within an NFS mount,
and auto-enable paralellism for this case. This way, users in NFS would
get the best performance by default. And it seems that using ~16 workers
would produce good results regardless of the NFS version that they might
be running.
The major downside is that detecting the file system type is quite
platform-dependent, so there is no simple and portable solution. (Also,
I'm not sure if the optimal number of workers would be the same on
different OSes). But we decided to give it a try, so this is a
rough prototype that would work for Linux:
https://github.com/matheustavares/git/commit/2e2c787e2a1742fed8c35dba185b7cd208603de9
Any thoughts on this idea? Or alternative suggestions?
Thanks,
Matheus
[1]: https://lore.kernel.org/git/815137685ac3e41444201316f537db9797dcacd2.1604521276.git.matheus.bernardino@xxxxxx/
I can't really speak to NFS performance, but I have to wonder if there's
not something else affecting the results -- 4 and/or 8 core results are
better than 16+ results in some columns. And we get diminishing returns
after ~16.
I'm wondering if during these test runs, you were IO vs CPU bound and if
VM was a problem. I'm wondering if setting thread affinity would help
here.
Jeff
