2012/7/5 Eric Dumazet <eric.dumazet@xxxxxxxxx>: > On Thu, 2012-07-05 at 00:48 +0900, JoonSoo Kim wrote: >> 2012/7/5 Eric Dumazet <eric.dumazet@xxxxxxxxx>: >> > Its the slow path. I am not convinced its useful on real workloads (not >> > a benchmark) >> > >> > I mean, if a workload hits badly slow path, some more important work >> > should be done to avoid this at a higher level. >> > >> >> In hackbench test, fast path allocation is about to 93%. >> Is it insufficient? > > 7% is insufficient I am afraid. > > One prefetch() in the fast path serves 93% of the allocations, > so added icache pressure is ok. > > One prefetch() in slow path serves 7% of the allocations, do you see the > difference ? > > Adding a prefetch() is usually a win when a benchmark uses the path one > million times per second. > > But adding prefetches also increases kernel size and it hurts globally. > (Latency of the kernel depends on its size, when cpu caches are cold) > Okay. Thanks for comments which is very helpful to me. -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to majordomo@xxxxxxxxx. For more info on Linux MM, see: http://www.linux-mm.org/ . Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>