Messages in this thread | | | Date | Mon, 24 Mar 2003 08:02:19 +0100 (CET) | From | Ingo Molnar <> | Subject | Re: 2.5.65-mm4 |
| |
On Sun, 23 Mar 2003, Andrew Morton wrote:
> 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/
| |