On 01/15/2015 04:00 PM, Dave Hansen wrote: > I/O devices are only getting faster. In fact, they're getting closer > and closer to memory in latency and bandwidth. But the VM is still > designed to do very orderly and costly procedures to reclaim memory, and > the existing algorithms don't parallelize particularly well. They hit > contention on mmap_sem or the lru locks well before all of the CPU > horsepower that we have can be brought to bear on reclaim. > > Once the latency to bring pages in and out of storage becomes low > enough, reclaiming the _right_ pages becomes much less important than > doing something useful with the CPU horsepower that we have. > > We need to talk about ways to do reclaim with lower CPU overhead and to > parallelize more effectively. > > There has been some research in this area by some folks at Intel and we > could quickly summarize what has been learned so far to help kick off a > discussion. I was actually planning to bring that up. Trinity can cause enough stress to a system that the hang watchdog triggers (with a 10 minute timeout!) inside reclaim code. Thanks, Sasha -- 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>