At the end of `git checkout <pathspec>`, we get a message informing how
many entries were updated in the working tree. However, this number can
be inaccurate for two reasons:
1) Delayed entries currently get counted twice.
2) Failed entries are included in the count.
The first problem happens because the counter is first incremented
before inserting the entry in the delayed checkout queue, and once again
when finish_delayed_checkout() calls checkout_entry(). And the second
happens because the counter is incremented too early in
checkout_entry(), before the entry was in fact checked out. Fix that by
moving the count increment further down in the call stack and removing
the duplicate increment on delayed entries. Note that we have to keep
a per-entry reference for the counter (both on parallel checkout and
delayed checkout) because not all entries are always accumulated at the
same counter. See checkout_worktree(), at builtin/checkout.c for an
example.
Signed-off-by: Matheus Tavares <matheus.bernardino@usp.br>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Make parallel checkout configurable by introducing two new settings:
checkout.workers and checkout.thresholdForParallelism. The first defines
the number of workers (where one means sequential checkout), and the
second defines the minimum number of entries to attempt parallel
checkout.
To decide the default value for checkout.workers, the parallel version
was benchmarked during three operations in the linux repo, with cold
cache: cloning v5.8, checking out v5.8 from v2.6.15 (checkout I) and
checking out v5.8 from v5.7 (checkout II). The four tables below show
the mean run times and standard deviations for 5 runs in: a local file
system on SSD, a local file system on HDD, a Linux NFS server, and
Amazon EFS (all on Linux). Each parallel checkout test was executed with
the number of workers that brings the best overall results in that
environment.
Local SSD:
Sequential 10 workers Speedup
Clone 8.805 s ± 0.043 s 3.564 s ± 0.041 s 2.47 ± 0.03
Checkout I 9.678 s ± 0.057 s 4.486 s ± 0.050 s 2.16 ± 0.03
Checkout II 5.034 s ± 0.072 s 3.021 s ± 0.038 s 1.67 ± 0.03
Local HDD:
Sequential 10 workers Speedup
Clone 32.288 s ± 0.580 s 30.724 s ± 0.522 s 1.05 ± 0.03
Checkout I 54.172 s ± 7.119 s 54.429 s ± 6.738 s 1.00 ± 0.18
Checkout II 40.465 s ± 2.402 s 38.682 s ± 1.365 s 1.05 ± 0.07
Linux NFS server (v4.1, on EBS, single availability zone):
Sequential 32 workers Speedup
Clone 240.368 s ± 6.347 s 57.349 s ± 0.870 s 4.19 ± 0.13
Checkout I 242.862 s ± 2.215 s 58.700 s ± 0.904 s 4.14 ± 0.07
Checkout II 65.751 s ± 1.577 s 23.820 s ± 0.407 s 2.76 ± 0.08
EFS (v4.1, replicated over multiple availability zones):
Sequential 32 workers Speedup
Clone 922.321 s ± 2.274 s 210.453 s ± 3.412 s 4.38 ± 0.07
Checkout I 1011.300 s ± 7.346 s 297.828 s ± 0.964 s 3.40 ± 0.03
Checkout II 294.104 s ± 1.836 s 126.017 s ± 1.190 s 2.33 ± 0.03
The above benchmarks show that parallel checkout is most effective on
repositories located on an SSD or over a distributed file system. For
local file systems on spinning disks, and/or older machines, the
parallelism does not always bring a good performance. For this reason,
the default value for checkout.workers is one, a.k.a. sequential
checkout.
To decide the default value for checkout.thresholdForParallelism,
another benchmark was executed in the "Local SSD" setup, where parallel
checkout showed to be beneficial. This time, we compared the runtime of
a `git checkout -f`, with and without parallelism, after randomly
removing an increasing number of files from the Linux working tree. The
"sequential fallback" column below corresponds to the executions where
checkout.workers was 10 but checkout.thresholdForParallelism was equal
to the number of to-be-updated files plus one (so that we end up writing
sequentially). Each test case was sampled 15 times, and each sample had
a randomly different set of files removed. Here are the results:
sequential fallback 10 workers speedup
10 files 772.3 ms ± 12.6 ms 769.0 ms ± 13.6 ms 1.00 ± 0.02
20 files 780.5 ms ± 15.8 ms 775.2 ms ± 9.2 ms 1.01 ± 0.02
50 files 806.2 ms ± 13.8 ms 767.4 ms ± 8.5 ms 1.05 ± 0.02
100 files 833.7 ms ± 21.4 ms 750.5 ms ± 16.8 ms 1.11 ± 0.04
200 files 897.6 ms ± 30.9 ms 730.5 ms ± 14.7 ms 1.23 ± 0.05
500 files 1035.4 ms ± 48.0 ms 677.1 ms ± 22.3 ms 1.53 ± 0.09
1000 files 1244.6 ms ± 35.6 ms 654.0 ms ± 38.3 ms 1.90 ± 0.12
2000 files 1488.8 ms ± 53.4 ms 658.8 ms ± 23.8 ms 2.26 ± 0.12
From the above numbers, 100 files seems to be a reasonable default value
for the threshold setting.
Note: Up to 1000 files, we observe a drop in the execution time of the
parallel code with an increase in the number of files. This is a rather
odd behavior, but it was observed in multiple repetitions. Above 1000
files, the execution time increases according to the number of files, as
one would expect.
About the test environments: Local SSD tests were executed on an
i7-7700HQ (4 cores with hyper-threading) running Manjaro Linux. Local
HDD tests were executed on an Intel(R) Xeon(R) E3-1230 (also 4 cores
with hyper-threading), HDD Seagate Barracuda 7200.14 SATA 3.1, running
Debian. NFS and EFS tests were executed on an Amazon EC2 c5n.xlarge
instance, with 4 vCPUs. The Linux NFS server was running on a m6g.large
instance with 2 vCPUSs and a 1 TB EBS GP2 volume. Before each timing,
the linux repository was removed (or checked out back to its previous
state), and `sync && sysctl vm.drop_caches=3` was executed.
Co-authored-by: Jeff Hostetler <jeffhost@microsoft.com>
Signed-off-by: Matheus Tavares <matheus.bernardino@usp.br>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Use multiple worker processes to distribute the queued entries and call
write_pc_item() in parallel for them. The items are distributed
uniformly in contiguous chunks. This minimizes the chances of two
workers writing to the same directory simultaneously, which could affect
performance due to lock contention in the kernel. Work stealing (or any
other format of re-distribution) is not implemented yet.
The protocol between the main process and the workers is quite simple.
They exchange binary messages packed in pkt-line format, and use
PKT-FLUSH to mark the end of input (from both sides). The main process
starts the communication by sending N pkt-lines, each corresponding to
an item that needs to be written. These packets contain all the
necessary information to load, smudge, and write the blob associated
with each item. Then it waits for the worker to send back N pkt-lines
containing the results for each item. The resulting packet must contain:
the identification number of the item that it refers to, the status of
the operation, and the lstat() data gathered after writing the file (iff
the operation was successful).
For now, checkout always uses a hardcoded value of 2 workers, only to
demonstrate that the parallel checkout framework correctly divides and
writes the queued entries. The next patch will add user configurations
and define a more reasonable default, based on tests with the said
settings.
Co-authored-by: Nguyễn Thái Ngọc Duy <pclouds@gmail.com>
Co-authored-by: Jeff Hostetler <jeffhost@microsoft.com>
Signed-off-by: Matheus Tavares <matheus.bernardino@usp.br>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
This new interface allows us to enqueue some of the entries being
checked out to later uncompress them, apply in-process filters, and
write out the files in parallel. For now, the parallel checkout
machinery is enabled by default and there is no user configuration, but
run_parallel_checkout() just writes the queued entries in sequence
(without spawning additional workers). The next patch will actually
implement the parallelism and, later, we will make it configurable.
Note that, to avoid potential data races, not all entries are eligible
for parallel checkout. Also, paths that collide on disk (e.g.
case-sensitive paths in case-insensitive file systems), are detected by
the parallel checkout code and skipped, so that they can be safely
sequentially handled later. The collision detection works like the
following:
- If the collision was at basename (e.g. 'a/b' and 'a/B'), the framework
detects it by looking for EEXIST and EISDIR errors after an
open(O_CREAT | O_EXCL) failure.
- If the collision was at dirname (e.g. 'a/b' and 'A'), it is detected
at the has_dirs_only_path() check, which is done for the leading path
of each item in the parallel checkout queue.
Both verifications rely on the fact that, before enqueueing an entry for
parallel checkout, checkout_entry() makes sure that there is no file at
the entry's path and that its leading components are all real
directories. So, any later change in these conditions indicates that
there was a collision (either between two parallel-eligible entries or
between an eligible and an ineligible one).
After all parallel-eligible entries have been processed, the collided
(and thus, skipped) entries are sequentially fed to checkout_entry()
again. This is similar to the way the current code deals with
collisions, overwriting the previously checked out entries with the
subsequent ones. The only difference is that, since we no longer create
the files in the same order that they appear on index, we are not able
to determine which of the colliding entries will survive on disk (for
the classic code, it is always the last entry).
Co-authored-by: Nguyễn Thái Ngọc Duy <pclouds@gmail.com>
Co-authored-by: Jeff Hostetler <jeffhost@microsoft.com>
Signed-off-by: Matheus Tavares <matheus.bernardino@usp.br>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
