lkml.org 
[lkml]   [2005]   [Mar]   [16]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
    /
    Date
    From
    SubjectRe: [patch 0/3] j_state_lock, j_list_lock, remove-bitlocks
    Ingo Molnar <mingo@elte.hu> wrote:
    >
    > > ooh, I'd rather not. I spent an intense three days removing all the
    > > sleeping locks from ext3 (and three months debugging the result).
    > > Ended up gaining 1000% on 16-way.
    > >
    > > Putting them back in will really hurt the SMP performance.
    >
    > ah. Yeah. Sniff.
    >
    > if we gain 1000% on a 16-way then there's something really wrong about
    > semaphores (or scheduling) though. A semaphore is almost a spinlock, in
    > the uncontended case - and even under contention we really (should) just
    > spend the cycles that we'd spend spinning. There will be some
    > intermediate contention level where semaphores hurt, but 1000% sounds
    > truly excessive.

    I forget how much of the 1000% came from that, but it was quite a lot.

    Removing the BKL was the first step. That took the context switch rate
    under high load from ~10,000/sec up to ~300,000/sec. Because the first
    thing a CPU hit on entry to the fs was then a semaphore. Performance rather
    took a dive.

    Of course the locks also became much finer-grained, so the contention
    opportunities lessened. But j_list_lock and j_state_lock have fs-wide
    scope, so I'd expect the context switch rate to go up quite a lot again.

    The hold times are short, and a context switch hurts rather ore than a quick
    spin.

    -
    To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
    the body of a message to majordomo@vger.kernel.org
    More majordomo info at http://vger.kernel.org/majordomo-info.html
    Please read the FAQ at http://www.tux.org/lkml/

    \
     
     \ /
      Last update: 2005-03-22 14:11    [W:5.050 / U:0.332 seconds]
    ©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site