Try to integrate fork of Chili parallel runtime

[zetanumbers]

Chili is a rust implementation of a parallel runtime with the heartbeat scheduling.Due to my bit of experience of working on rustc parallel runtime, I've tried to do a quick integration of chili into rustc to check the performance. My hacky modifications to Chili consist of:

• Addition of TLV;
• Deadlock detection with mark_(un)blocked methods;
• Addition of another thread pool creation method scoped_with_config to mimic rayon's design;
  • As such spawning exact number of worker threads as specified, no longer considering parent thread to be a worker;
  • Add install method to run code on the worker threads.
• Global ThreadPool was replaced with thread local context respecting subroutine calls.

I've also removed parallel! macro as I couldn't figure out how to do it without causing ambiguity error:

error: local ambiguity when calling macro `parallel`: multiple parsing options: built-in NTs expr ('blocks') or expr ('endblock').

Original runtime repo: https://github.com/dragostis/chili

Modifications to Chili to accommodate rustc: dragostis/chili@main...zetanumbers:chili:rustc

Related zulip topic: #t-compiler/parallel-rustc > use heartbeat scheduling to improve parallel frontend

[rustbot]

r? @fee1-dead

rustbot has assigned @fee1-dead.
They will have a look at your PR within the next two weeks and either review your PR or reassign to another reviewer.

Use r? to explicitly pick a reviewer

[zetanumbers]

Currently this code fails from inside of Chili: dragostis/chili#29
Currently this code just hangs up

[bors]

☔ The latest upstream changes (presumably #139983) made this pull request unmergeable. Please resolve the merge conflicts.

[zetanumbers]

Prototype integration is ready for benchmarks

[Zoxc]

7 threads:

Benchmark	Before	After		Before	After		Before	After
	Time	Time	%	Physical Memory	Physical Memory	%	Committed Memory	Committed Memory	%
🟣 clap:check	0.4688s	0.6971s	💔 48.70%	200.59 MiB	200.55 MiB	-0.02%	278.42 MiB	276.75 MiB	-0.60%
🟣 hyper:check	0.1370s	0.1852s	💔 35.20%	127.18 MiB	126.30 MiB	-0.69%	200.79 MiB	199.72 MiB	-0.53%
🟣 regex:check	0.2954s	0.4575s	💔 54.88%	161.01 MiB	159.40 MiB	-1.00%	229.10 MiB	227.50 MiB	-0.70%
🟣 syn:check	0.5696s	0.7758s	💔 36.20%	194.07 MiB	192.92 MiB	-0.59%	264.84 MiB	263.13 MiB	-0.64%
Total	1.4707s	2.1155s	💔 43.84%	682.84 MiB	679.17 MiB	-0.54%	973.16 MiB	967.10 MiB	-0.62%
Summary	1.0000s	1.4375s	💔 43.75%	1 byte	0.99 bytes	-0.57%	1 byte	0.99 bytes	-0.62%

Benchmark	Before	After		Before	After		Before	After
	Time	Time	%	Physical Memory	Physical Memory	%	Committed Memory	Committed Memory	%
🟣 clap:check:unchanged	0.3761s	0.3591s	💚 -4.51%	133.04 MiB	132.72 MiB	-0.24%	234.70 MiB	234.40 MiB	-0.13%
🟣 hyper:check:unchanged	0.1509s	0.1418s	💚 -5.99%	85.57 MiB	85.37 MiB	-0.23%	191.57 MiB	191.32 MiB	-0.13%
🟣 regex:check:unchanged	0.2882s	0.2698s	💚 -6.39%	112.93 MiB	112.51 MiB	-0.37%	212.20 MiB	211.58 MiB	-0.29%
🟣 syn:check:unchanged	0.5831s	0.5616s	💚 -3.69%	157.73 MiB	157.05 MiB	-0.43%	260.27 MiB	259.25 MiB	-0.39%
Total	1.3982s	1.3323s	💚 -4.71%	489.27 MiB	487.65 MiB	-0.33%	898.74 MiB	896.54 MiB	-0.25%
Summary	1.0000s	0.9486s	💚 -5.14%	1 byte	1.00 bytes	-0.32%	1 byte	1.00 bytes	-0.24%

2 threads:

Benchmark	Before	After		Before	After		Before	After
	Time	Time	%	Physical Memory	Physical Memory	%	Committed Memory	Committed Memory	%
🟣 clap:check	0.8644s	1.0451s	💔 20.90%	193.95 MiB	194.38 MiB	0.22%	265.88 MiB	264.76 MiB	-0.42%
🟣 hyper:check	0.1822s	0.2146s	💔 17.73%	124.16 MiB	124.16 MiB	-0.00%	197.65 MiB	197.43 MiB	-0.11%
🟣 regex:check	0.5183s	0.6032s	💔 16.37%	155.93 MiB	155.86 MiB	-0.05%	223.65 MiB	223.58 MiB	-0.03%
🟣 syn:check	0.8861s	0.9679s	💔 9.23%	187.97 MiB	188.41 MiB	0.23%	258.04 MiB	258.30 MiB	0.10%
Total	2.4511s	2.8307s	💔 15.49%	662.02 MiB	662.81 MiB	0.12%	945.22 MiB	944.06 MiB	-0.12%
Summary	1.0000s	1.1606s	💔 16.06%	1 byte	1.00 bytes	0.10%	1 byte	1.00 bytes	-0.12%

Benchmark	Before	After		Before	After		Before	After
	Time	Time	%	Physical Memory	Physical Memory	%	Committed Memory	Committed Memory	%
🟣 clap:check:unchanged	0.3342s	0.3383s	💔 1.22%	132.49 MiB	132.61 MiB	0.09%	234.06 MiB	234.17 MiB	0.05%
🟣 hyper:check:unchanged	0.1370s	0.1392s	💔 1.58%	84.94 MiB	85.03 MiB	0.11%	190.79 MiB	190.84 MiB	0.03%
🟣 regex:check:unchanged	0.2531s	0.2542s	0.47%	112.21 MiB	111.99 MiB	-0.20%	211.32 MiB	211.05 MiB	-0.13%
🟣 syn:check:unchanged	0.5317s	0.5372s	💔 1.03%	156.78 MiB	156.85 MiB	0.04%	259.15 MiB	258.92 MiB	-0.09%
Total	1.2559s	1.2689s	💔 1.03%	486.42 MiB	486.48 MiB	0.01%	895.31 MiB	894.98 MiB	-0.04%
Summary	1.0000s	1.0108s	💔 1.08%	1 byte	1.00 bytes	0.01%	1 byte	1.00 bytes	-0.03%

[zetanumbers]

Hm, well I forgot to mention there's no jobserver integrated yet, so it's only relevant to benchmark a single rustc run, instead of a whole cargo check. Sorry for that. I would like to add this next.

Also there is a constant parameter on join_with_heartbeat_every we should keep in mind, which corresponds to how many join operations are actually parallel, unless worker's job queue was almost empty (len < 3) at that time. Simple join has it set to one out of 64 joins, which is bad for heavy code chunks.

And do not worry about doing benchmarks for me, unless you want to. I may not have time to collect benchmarks before day ends, so I just update the progress as it is. :)

And thank you for your measurements!

[Zoxc]

Hm, well I forgot to mention there's no jobserver integrated yet, so it's only relevant to benchmark a single rustc run, instead of a whole cargo check

The benchmarks are for a single rustc run. I'm using rcb (rcb bench --check --incr-none -n 10 --details none --rflag=a:-Zthreads=7 --threads=a master~~e master~chili~8).

The unchanged are interesting as the improvement doesn't seem to be reduced per-thread overhead, which would show up with 2 threads too.

[zetanumbers]

Either this implementation of a parallel runtime with the heartbeat scheduling or the heartbeat scheduling itself has introduced regressions after its integration into rustc. I see no reason trying to further salvage this idea.