[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: Compiler memory consumption
From: |
Andy Wingo |
Subject: |
Re: Compiler memory consumption |
Date: |
Tue, 16 May 2017 22:14:09 +0200 |
User-agent: |
Gnus/5.13 (Gnus v5.13) Emacs/25.1 (gnu/linux) |
Hi!
Definitely sounds like we have an issue we should deal with.
On Tue 16 May 2017 18:19, address@hidden (Ludovic Courtès) writes:
> here is the gcprof output
I find gcprof useful when I want to improve runtime by removing
allocation. However I don't find it useful when dealing with memory use
issues; ymmv of course.
> while compiling gnu/packages/python.scm, which defines 841 package
> objects (structs) with 5 times more thunks of the form (lambda ()
> value):
You mention later:
> Also, for reference, loading python.go peaks at 315M RSS:
>
> $ \time ./pre-inst-env guile -c '(use-modules (gnu packages python))'
> 0.18user 0.02system 0:00.18elapsed 112%CPU (0avgtext+0avgdata
> 315648maxresident)k
> 0inputs+0outputs (0major+7784minor)pagefaults 0swaps
But this I don't understand. If I do a ./pre-inst-env guile and then
load (gnu packages python), I get a 20MB heap size, a 35MB total private
dirty memory and 52MB clean shared memory. (Measured using smaps via
https://wingolog.org/pub/mem_usage.py). This is at commit
60c9e80444421c412ae3d0e7b4b224ef0e32947f.
I just built the "time" package and I see similar numbers here. I can
only think that the "time" package's numbers are bogus.
> time(1) reports a maximum resident set size of 3.8G (though I see
> something around 900MiB in ‘top’.)
The only way Guile's memory usage can shrink in practice is if it
recurses a lot on the stack and then returns those pages to the OS. I
don't think libgc will return pages to the OS (though I could be
wrong). So that would be a possibility to look into, if time can be
trusted.
> When compiling python.scm #:to 'cps, we end up with 1G max RSS in 6s.
Measured with time? If this is the case it could be that python.scm is
just a lot of code. Any compiler would take a lot if the IR size is 1
GB.
> The only conclusion I can draw is that cps-to-bytecode compilation seems
> to be responsible for most of the memory consumption.
This is possible but I am not there yet. I don't see why compiling this
file to CPS should cause memory usage of 1GB. That is 9000 memory words
per textual line -- simply too much.
Many unknowns here!
Andy