> Note that the lock_kernel() contention has been drastically reduced and
> we're now hitting semaphore contention.
>
> Running `dbench 32' on the quad Xeon, this patch took the context switch
> rate from 500/sec up to 125,000/sec.
note that there is _nothing_ wrong in doing 125,000 context switches per
sec, as long as performance increases over the lock_kernel() variant.
> I've asked Alex to put together a patch for spinlock-based locking in
> the block allocator (cut-n-paste from ext2).
sure, do this if it increases performance. But if it _decreases_
performance then it's plain pointless to do this just to avoid
context-switches. With the 2.4 scheduler i'd agree - avoid
context-switches like the plague. But context-switches are 100% localized
to the same CPU with the O(1) scheduler, they (should) cause (almost) no
scalability problem. The only thing this change will 'fix' is the
context-switch statistics.
plus someone might want to try some simple spin-sleep semaphore
implementation at this point. The context-switch takes roughly 2 usecs on
a typical x86 box, so i'd say spinning for 0.5 or 1.0 usecs could provide
some speedup.
Ingo
-
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/