Re: [Re: paging behavior in Linux]

[Neelam Saboo]

Another observation. I have two independent programs. One program incurring
page faults and another program just doing some work.
When work program run undependently it takes ~19 seconds of CPU time, but when
it is run along with page faulting program on the same machine, it takes ~32
seconds of CPU time. Doesnt this indicate that page faults in a program slows
down all the program on the machine and not only threads in the same process
?

neelam

Andrew Morton <[e-mail]> wrote:
> David Mansfield wrote:
> > 
> > Manfred Spraul wrote:
> > >
> > > >
> > > > When I run my program on a readhat linux machine, I dont get results
as
> > > > expected, work thread seems to be stuck when prefetch thread is
waiting on
> > > > a page fault
> > > >
> > > That's a known problem:
> > >
> > > The paging io for a process is controlled with a per-process semaphore.
> > > The semaphore is held while waiting for the actual io. Thus the paging
> > > in multi threaded applications is single threaded.
> > > Probably your prefetch thread is waiting for disk io, and the worker
> > > thread causes a minor pagefault --> worker thread sleeps until the disk
> > > io is completed.
> > 
> > This behavior is actually pretty annoying.  There can be cases where a
> > process wakes up from a page fault, does some work, goes back to sleep
> > on a page fault, thereby keeping it's mmap_sem locked at all times (i.e.
> > vmstat, top, ps unusable) on a UP system.  I posted this complaint a
> > while ago, it was discussed by Linus and Andrew Morton about how it also
> > boiled down to semaphore wakeup unfairness (and bugs?).  The current
> > semaphore was determined to be too ugly to even look at.  So it was
> > dropped.
> > 
> > Is there any way that the mmap_sem could be dropped during the blocking
> > on I/O, and reclaimed after the handle_mm_fault?  Probably not, or it'd
> > be done.
> > 
> > It can be a real DOS though, a 'well-written' clobbering program can
> > make ps/vmstat useless.  (it's actually /proc/pid/stat that's the
> > killer, IIRC).
> 
> Did the `goto inside' trick in the semaphore code actually
> fix this unfairness issue?
> 
> -

____________________________________________________________________
Get free email and a permanent address at http://www.netaddress.com/?N=1