Ben, I tested this (without the extra malloc), and this is a minor improvement on master, but a huge, in the order of ~40% performance increase on top of the NUMA patches I sent yesterday. Also, performance does not vary so much any more, and it is now actually beneficial to add more threads and netperfs. Now, with this and the NUMA level kernel stats, find_match_wc() is actually the top line on ‘perf’, as it should!
Based on testing only: Acked-by: Jarno Rajahalme <[email protected]> On Jan 8, 2014, at 12:31 PM, Ben Pfaff <[email protected]> wrote: > On Wed, Jan 08, 2014 at 11:07:38AM -0800, Alex Wang wrote: >> I suspect that this approach to a rwlock is faster when lots of threads >>> try to take the read lock at once, because it should avoid cross-core >>> contention on cache lines. >> >> >> >> For avoiding the contention, is that only because each thread will have its >> own lock >> or more precisely "fat_rwlock_perthread" which is aligned to cache line? > > Each thread having its own lock is the main benefit. Cache-line > alignment ensures that a lock's memory doesn't happen to overlap the > memory used for a different thread's lock (or other data used by > another thread). > >>> + base = xmalloc(sizeof *perthread + CACHE_LINE_SIZE - 1); >>> + perthread = (void *) ROUND_UP((uintptr_t) base, CACHE_LINE_SIZE); >>> + >>> >>> * + perthread = xmalloc(sizeof *perthread);* + perthread->base = >>> base; >>> + perthread->rwlock = rwlock; >>> + list_push_back(&rwlock->threads, &perthread->list_node); >>> + perthread->table = table; >>> + hmap_insert(&table->perthreads, &perthread->hmap_node, rwlock->id); >>> + ovs_mutex_init(&perthread->mutex); >>> + ovs_mutex_lock(&perthread->mutex); >>> + perthread->depth = 0; >>> >> >> Here, you want to align the perthread to cache line size, right? > > Yes. > >> What is the meaning of this second "perthread = xmalloc(sizeof >> *perthread);" ? > > That's a bug, thanks for reporting it. I've deleted that line now. > _______________________________________________ > dev mailing list > [email protected] > http://openvswitch.org/mailman/listinfo/dev
_______________________________________________ dev mailing list [email protected] http://openvswitch.org/mailman/listinfo/dev
