Parallelize precompiles generation II

[petvana]

This follows up #47983 and introduce asynchronous read for repl precompiles. In contrast to IanButterworth@5ee671a?w=1 from @IanButterworth , it uses readavailable and BufferStream to handle end of lines correctly. It saves only a few seconds, but improves the fancy printing for impatient developers.

[IanButterworth]

It saves only a few seconds

For me the gains are bigger

before #47983

Generating REPL precompile statements... 40/40
Executing precompile statements... 1756/1760
Precompilation complete. Summary:
Generation ──  77.218106 seconds 79.4016%
Execution ───  20.031916 seconds 20.5984%
Total ───────  97.250022 seconds

#47983

Collecting and executing precompile statements
└ Collect (Basic: ✓ 730, REPL 40/40: ✓ 1659) => Execute ✓ 1751
Precompilation complete. Summary:
Total ───────  80.336407 seconds

This PR

Collecting and executing precompile statements
└ Collect (Basic: ✓ 730, REPL 40/40: ✓ 1657) => Execute ✓ 1750
Precompilation complete. Summary:
Total ───────  64.902187 seconds

[petvana]

For me the gains are bigger

Half a minute, as advertised in the original PR 😉

[IanButterworth]

Seems good to me.

Does the switch to serial mode for debugging with stdout rather than devnull still work?

[petvana]

Does the switch to serial mode for debugging with stdout rather than devnull still work?

Not sure, what you mean by stdout rather than devnull, but the switch still works and those 3 steps are done in serial order.

[IanButterworth]

I just grouped all the debug options together at the top of the file to make it clearer. Seems to work for me, if you disable pretty printing, disable parallel, and enable repl input viewing.

edit: Had to move them to inside the module after stdout had been reinitialized

[IanButterworth]

Seems good to me, but might be good to have a @vtjnash review too

[IanButterworth]

Out of curiosity I reordered repl_script to have Profile then Pkg then the ones in this file and saw 5 seconds faster.

Not sure if it's worth doing as it breaks the ordering (could be a followup PR), but at least it makes sense that doing the larger workloads first helps parallelization

Collecting and executing precompile statements
└ Collect (Basic: ✓ 730, REPL 40/40: ✓ 1661) => Execute ✓ 1753
Precompilation complete. Summary:
Total ───────  60.710451 seconds

[giordano]

what you mean by stdout rather than devnull

julia/contrib/generate_precompile.jl

Line 289 in cc5bc87

debug_output = devnull # or stdout

[petvana]

Out of curiosity I reordered repl_script to have Profile then Pkg then the ones in this file and saw 5 seconds faster.

I'm kind of surprised, it saved you so much time. I believe, we can reorder repl script as we want because it will remain deterministic.

[IanButterworth]

ok! Pushed

[IanButterworth]

Windows failures are unrelated. @vtjnash does this look ok to you?

[IanButterworth]

I'll merge as this appears to work as designed. If @vtjnash has any issue with the buffering design that can be addressed in a followup

[maleadt]

EDIT: I can't reproduce this, so let's hope it was a fluke 🙂

---

I just had a build hang (wasn't using the changes from this PR yet):

Collecting and executing precompile statements
└ Collect (Basic: ✓ 730, REPL 39/40: ◒ ) => Execute ◒ 494^C^C^C

As there was no CPU usage, but the main process was still happily spinning, I suspect this may be related to these parallel precompilation changes. The process also wasn't interruptible, i.e., CTRL-C is ignored (as you can see in the output). Are either of these issues known?

[StefanKarpinski]

We really need to get better at making ctrl-C work. I know that @vtjnash will tell me that it can't be made to work completely reliably, which, sucks, sure, but it can be made to work much better than it does and we should do that.

[IanButterworth]

@maleadt if you do hit it again, can you SIGINFO/SIGUSR1 profile it to see what it's doing. That's probably more reliable than getting a trace from an interrupt

[petvana]

I just had a build hang (wasn't using the changes from this PR yet):

I've noticed a hang when building without stdlibs, i.e., stdlibs = [] in sysimg.jl. It is reproduceble, but happened even before #47983. Thankfully, this can be fixed easily by disabling the whole REPL collection as have_repl = false.

[vtjnash]

LGTM. Copying streams (as done here) is one of the best actual use cases for readavailable