Re: Wow... (<-- blown away at performance)

From: Alexander Best <arundel_at_freebsd.org>
Date: Wed, 23 Feb 2011 13:45:19 +0000
On Wed Feb 23 11, John Baldwin wrote:
> On Tuesday, February 22, 2011 4:50:36 pm Alexander Best wrote:
> > On Tue Feb 22 11, Brandon Gooch wrote:
> > > On Tue, Feb 22, 2011 at 2:31 PM, Alexander Best <arundel_at_freebsd.org> 
> wrote:
> > > > On Tue Feb 22 11, Garrett Cooper wrote:
> > > >> On Tue, Feb 22, 2011 at 2:10 AM, Eir Nym <eirnym_at_gmail.com> wrote:
> > > >> > On 22 February 2011 11:15, Garrett Cooper <yanegomi_at_gmail.com> wrote:
> > > >> >>    I don't know what to say, but r218938 screams with flash videos
> > > >> >> (native Linux speed). Not sure if it's the new binutils or if it's 
> the
> > > >> >> new linuxulator patches, but I can run multiple instances of youtube
> > > >> >> in parallel (5 total with other miscellaneous flash animation) 
> without
> > > >> >> it totally lagging out Firefox/X11, and it appears to close the
> > > >> >> instances of firefox properly now. Hopefully this version fares 
> better
> > > >> >> than r218113 did (I think I hit a kernel bug after 2 weeks uptime,
> > > >> >> where my system just hardlocked for no apparent reason).
> > > >> >>    Anyhow, hope others have similar results.
> > > >> >> Cheers!
> > > >> >> -Garrett
> > > >> >>
> > > >> >> $ uname -a
> > > >> >> FreeBSD bayonetta.local 9.0-CURRENT FreeBSD 9.0-CURRENT #0 r218938M:
> > > >> >> Mon Feb 21 23:10:51 PST 2011
> > > >> >> gcooper_at_bayonetta.local:/usr/obj/usr/src/sys/BAYONETTA  amd64
> > > >> >
> > > >> > Which FlashPlayer version do you test? Adobe has made significant
> > > >> > performance changes in 10.2 (from 10.1). You can search for 
> StageVideo
> > > >> > performance to learn more about. Youtube already use them since 10.2
> > > >> > beta
> > > >>
> > > >>     linux-f10-flashplugin-10.1r102.65 . The performance increases are
> > > >> claimed to be "up to 85%" on the Stage Video site, but I'm seeing a
> > > >> more than 200% increase (now it actually scales between multiple
> > > >> instances, instead of croaks with one instance, tiling up and down the
> > > >> screen when moving the window slider for instance or switching tabs).
> > > >> Besides, it seems like it needs external support from the video
> > > >> driver, and I'm not sure that that bridge exists in the linuxulator.
> > > >>     Also, it looks like npviewer.bin still hangs to resources on until
> > > >> Firefox closes (or I kill it :)..), so something still needs to be
> > > >> resolved there, but that isn't a regression (it's acted that way for
> > > >> ages), and shouldn't be too hard to do.
> > > >
> > > > i think the problem with npviewer.bin having to be killed by hand is a 
> futex
> > > > issue in combination with a high number of threads. this is the output 
> of a
> > > > test from darren hart's futex test suite under freebsd 9.0:
> > > >
> > > > futex_wait: Measure FUTEX_WAIT operations per second
> > > >        Arguments: iterations=100000000 threads=1
> > > > Result: 18622 Kiter/s
> > > > futex_wait: Measure FUTEX_WAIT operations per second
> > > >        Arguments: iterations=100000000 threads=2
> > > > Result: 15469 Kiter/s
> > > > futex_wait: Measure FUTEX_WAIT operations per second
> > > >        Arguments: iterations=100000000 threads=3
> > > > Result: 12713 Kiter/s
> > > > futex_wait: Measure FUTEX_WAIT operations per second
> > > >        Arguments: iterations=100000000 threads=4
> > > > Result: 12247 Kiter/s
> > > > futex_wait: Measure FUTEX_WAIT operations per second
> > > >        Arguments: iterations=100000000 threads=5
> > > > Result: 11814 Kiter/s
> > > > futex_wait: Measure FUTEX_WAIT operations per second
> > > >        Arguments: iterations=100000000 threads=6
> > > > Result: 13468 Kiter/s
> > > > futex_wait: Measure FUTEX_WAIT operations per second
> > > >        Arguments: iterations=100000000 threads=8
> > > > Result: 12061 Kiter/s
> > > > futex_wait: Measure FUTEX_WAIT operations per second
> > > >        Arguments: iterations=100000000 threads=10
> > > > Result: 12854 Kiter/s
> > > > futex_wait: Measure FUTEX_WAIT operations per second
> > > >        Arguments: iterations=100000000 threads=12
> > > > Result: 12999 Kiter/s
> > > > futex_wait: Measure FUTEX_WAIT operations per second
> > > >        Arguments: iterations=100000000 threads=16
> > > > Result: 12402 Kiter/s
> > > > futex_wait: Measure FUTEX_WAIT operations per second
> > > >        Arguments: iterations=100000000 threads=24
> > > > Result: 9815 Kiter/s
> > > > futex_wait: Measure FUTEX_WAIT operations per second
> > > >        Arguments: iterations=100000000 threads=32
> > > > Result: 8518 Kiter/s
> > > > futex_wait: Measure FUTEX_WAIT operations per second
> > > >        Arguments: iterations=100000000 threads=64
> > > >        ERROR: Resource temporarily unavailable: pthread_create
> > > > Result: ERROR
> > > > futex_wait: Measure FUTEX_WAIT operations per second
> > > >        Arguments: iterations=100000000 threads=128
> > > >        ERROR: Resource temporarily unavailable: pthread_create
> > > > Result: ERROR
> > > > futex_wait: Measure FUTEX_WAIT operations per second
> > > >        Arguments: iterations=100000000 threads=256
> > > >        ERROR: Resource temporarily unavailable: pthread_create
> > > > Result: ERROR
> > > > futex_wait: Measure FUTEX_WAIT operations per second
> > > >        Arguments: iterations=100000000 threads=512
> > > >        ERROR: Resource temporarily unavailable: pthread_create
> > > > Result: ERROR
> > > > futex_wait: Measure FUTEX_WAIT operations per second
> > > >        Arguments: iterations=100000000 threads=1024
> > > >        ERROR: Resource temporarily unavailable: pthread_create
> > > > Result: ERROR
> > > >
> > > > cheers.
> > > > alex
> > > 
> > > Have you found any method or workaround to mitigate this issue?
> > 
> > no. i've increased the kern.threads.max_threads_per_proc value, but that had
> > no effect. so it seems to be a bug in the linuxulator's futex 
> implementation.
> 
> You can try http://www.freebsd.org/~jhb/patches/futex_share.patch
> 
> Not sure if it will solve your issue, but it might help if multiple
> processes (Linux processes, not just multiple FreeBSD processes
> masquerading as threads in a single Linux process) are involved.

no unfortunately. i don't think the failures with threads 64-1024 are futex
related, since linux_mmap2() returns -ENOMEN and thus seems responsible.

however even with your patch i still see lots of linux_sys_futex() calls
failing with -EAGAIN in the range of 2-32 threads. it seems nanosleep is
failing. either because some linux timespecs aren't converted to their *BSD
equivalents properly or another reason could be that CLOCK_MONOTONIC and
CLOCK_REALTIME aren't yet honoured in linux_futex.c, but simply discarded.

the actual place in linux_futex.c where EAGAIN gets handed over to the
linuxulator is here:

static int
futex_sleep(struct futex *f, struct waiting_proc *wp, int timeout)
{
        int error;
 
        FUTEX_ASSERT_LOCKED(f);
        LINUX_CTR4(sys_futex, "futex_sleep enter uaddr %p wp %p timo %d ref %d",
            f->f_uaddr, wp, timeout, f->f_refcount);
        error = sx_sleep(wp, &f->f_lck, PCATCH, "futex", timeout);
        // HERE EAGAIN GETS RETURNED

cheers.
alex

> 
> -- 
> John Baldwin

-- 
a13x
Received on Wed Feb 23 2011 - 12:45:19 UTC

This archive was generated by hypermail 2.4.0 : Wed May 19 2021 - 11:40:11 UTC