01:43 eternaleye joined 03:31 woosley joined 03:55 woosley joined 04:13 colomon joined
lue is almost surprised the macro-happy-ish source didn't factor out sigs beginning with (MVMThreadContext *tc, to SIG_TC( :) 05:46
07:11 FROGGS joined
FROGGS lue: jnthn is not fond of overmacroing it :o) 07:21
10:33 jnthn joined
FROGGS hmmm, nqp-m -e '1' takes 0m0.175s while nqp-p slurps in a 35kb text file within 0m0.112s 10:56
nqp-m needs 0m0.209s to slurp it
tadzik man the profilers! :P 11:03
FROGGS already doing :o) 11:07
but lunch first
&
12:14 tgt joined 12:34 woolfy left
FROGGS jnthn: here is the heap profile for `nqp-m -e 1` btw froggs.de/perl6/nqpm-e1.ps 12:45
though, might not that readable on screen for you :/ 12:46
timotimo i can't get that to render in okular nor evince 12:52
ah, converted it to a pdf, now it works
timotimo doesn't know what to do with that info 12:53
jnthn hm...can I have that PDF? :) 12:54
nwc10 can confirm that it renders OK here, but no good idea what to do with that info, even if I have time, which I don't
timotimo jnthn: you don't have ps2pdf installed? ;)
gimme a sec.
where do i upload this to :|
timotimo mails it 12:55
jnthn oh, nice! 12:58
FROGGS: When I see somehting like:
timotimo so, that was memory usage, not time usage, right? 12:59
jnthn deserialize_frames
0.1 (1.4%)
of 0.6 (5.4%)
timotimo yeah, heap profile
jnthn What are the first and second lines about?
incl/excl?
timotimo it seems to me one is the time the thing itself takes, the other the thing plus its children
jnthn But that doesn't make sense to me either...
timotimo and one compared to its parents
fwiw, 10 MB is pretty decent
jnthn Yeah. 13:00
That add_page mallocs most is not so surprising. 13:01
But we may want to tweak bin sizing or something. 13:02
13:04 FROGGS joined
FROGGS jnthn: there is also other stuff that I can generate, see code.google.com/p/gperftools/wiki/G...manceTools 13:05
timotimo ah, it's perf!
perf is so excellent :)
oh, wait, that's not the thing i was thinking of 13:06
jnthn FROGGS: Hm, what does the CPUPROFILE thing give as output? 13:07
FROGGS jnthn: I'll look into that after $meeting
13:44 tgt joined 14:12 jnap joined 15:02 tgt joined 15:15 FROGGS[mobile] joined 15:58 FROGGS joined 18:13 ssutch joined 19:20 lizmat joined 19:27 tgt joined
FROGGS jnthn: froggs.de/perl6/cpu_nqpm-e1.pdf 19:42
nwc10 that was a long meeting :-) 19:45
[yes, strictly, "after" doesn't mean "immediately after"] 19:46
FROGGS froggs.de/perl6/cpu_nqpm_slurp35kb.pdf 19:47
nwc10: I had to find out first why it didn't work :o) 19:48
jnthn FROGGS: Normally 'cus nobody listens, or there's not enough coffee... 19:49
FROGGS hmm? 19:50
jnthn Why meetings don't work :P
FROGGS hehe
well, yeah
jnthn OK, I can see things to do from this graph :) 19:51
FROGGS it was kinda ridiculous
that is a nice renderer btw
jnthn One outstanding task is to get the serialization blog dragged out of the string heap and into its own segment. 19:52
uh, blob.
heh, serialization blog. That's be fascinating. "I just wrote an int32 to disk!" "I just wrote an int16 to disk!" "I just wrote a string heap index!" 19:53
nwc10 instead couldn't you send it to Twitter, and see if it improves the signal to noise ratio? 19:54
and find out how many followers you pick up
19:59 lizmat joined
diakopter FROGGS: the numbers in that graph seem much too small to indicate anything 20:18
sample size, I mean 20:19
only 15 samples?
FROGGS well, I'd need a better test case then :o) 20:20
jnthn If the test case is "startup"... :) 20:21
The exact percentages may be a bit off (sampling profilers trade that in for low overhead), but I think the big picture is quite believable.
FROGGS gperf I used earlier lets you rerun the test and merge the results to get better percentages 20:23
jnthn Ah, that'd do nicely too :)
FROGGS btw, the first trap I fell into was when looking at the heapmap: why does add_page take so long? 20:28
I know, it was silly of me :o) 20:29
jnthn It's fast, just hungry. Like a cheetah that didn't eat in a while. 20:32
FROGGS diakopter: I was able to increase the sample size... 20:56
diakopter ok :)
url? 20:57
FROGGS froggs.de/perl6/cpu_nqpm-e1.pdf
same link as before, I hope it does not cache it
diakopter invalid pdf
(web server cached some partial write?) 20:58
FROGGS ohh, the upload got stuck
hmmm 21:00
diakopter: now?
nwc10 thought I had a few days ago - PyPy 2.2 recently released emphasising that it's just gained incremental GC: morepypy.blogspot.co.uk/2013/11/pyp...alism.html 21:10
diakopter yeah, but 39 is still way too small; I posted a link to the output of somehting that instrumented and logged millions of calls
nwc10 Incremetnal GC seemed to be a big new feature of Rubinius recently
jnthn diakopter: sampling != instrumental 21:11
nwc10 superficially this looks like "everyone needs incremental GC"
diakopter I know.
jnthn: my point was that it's possible to sampl all of them.
and that 39 is miniscule compared to what it could get
nwc10 but I think it actually means "you can be a sucessful project for at least 5 years before incremental GC becomes the most important thing to fix"
FROGGS diakopter: then I'd need to set the sample-per-seconds to a gazillion
jnthn diakopter: Yes, mine is that IME even a smallish number of samples (granted, normally 100s, not 10s) can be enough to give a good idea... :) 21:12
diakopter FROGGS: definitely... try it.. set it to 10,000
at least
or 1e6
FROGGS that was the command: LD_PRELOAD="/usr/lib/libprofiler.so" CPUPROFILE_FREQUENCY=1000000 CPUPROFILE=_cpuprofile nqp-m -e 1
I can add more zeros though :o)
diakopter but anyway the total startup time is quite good; why optimize that? 21:13
(it's like 8ms on windows msvc)
at least 50s goes to loading shared libraries
jnthn nwc10: Incremental is a bit of a trade-off too. In general, shorter pause times trade off against overall throughput.
FROGGS the startup time is at 0.2s while nqp-p can startup and slurp a 35kb test file within 0.125s 21:14
nwc10 jnthn: yes, I remember you saying that, and it seems intuitive
FROGGS diakopter: nqp-m's startup time?
diakopter yes?
nwc10 I guess what's been bugging them is that they have users who want to write games or stuff with some degree of desired responsiveness
but, anyway, at least 5 years until it matters :-) 21:15
jnthn As another data point, .Net ships with a client GC and a server GC
diakopter I vehemently disagree; there are tons of definitions of "successful project" and I dont think pypy meets very many of them
FROGGS I would say that C is successful 21:16
diakopter who wouldn't?
FROGGS :o)
diakopter if there were only 39 samples, and it was trying to make 1e6 of them per second, then it seems it only profiled for 39 microseconds 21:19
FROGGS which is weird that the samples only increase to 41 when I add several zeros 21:20
diakopter well, it's reached some limit
which isn't helpful for such short-lived programs 21:21
I already reported that 90% of the startup time (in code, not loading shared libraries) was spent in deserialization
(and nearly all of that in decoding and initialization strings) 21:22
*initializing
nwc10: from what the rubinius x guy said, pypy basically failed to garner hardly any users 21:23
nwc10 diakopter: oh interesting. Where did he say that? I'd not spotted that
implied - Rubinius has won itself a userbase? 21:24
diakopter he didn't seem to imply that 21:25
nwc10 oh, OK. 21:26
diakopter ergh
21:45 benabik joined 22:07 benabik joined 22:18 lizmat joined 22:41 BenGoldberg joined, woolfy joined

Powered by Cro and the Raku® Programming Language.

Please report any issues / comments / feature requests as an issue on App::Raku::Log.

Thank you!