On Thu, Oct 14, 2010 at 11:50:28AM +0900, KOSAKI Motohiro wrote: > > On Wed, Oct 13, 2010 at 04:10:43PM +0900, KOSAKI Motohiro wrote: > > > When memory shortage, we are using drain_pages() for flushing per cpu > > > page cache. In this case, per cpu stat should be flushed too. because > > > now we are under memory shortage and we need to know exact free pages. > > > > > > Otherwise get_page_from_freelist() may fail even though pcp was flushed. > > > > > > > With my patch adjusting the threshold to a small value while kswapd is awake, > > it seems less necessary. > > I agree this. > > > It's also very hard to predict the performance of > > this. We are certainly going to take a hit to do the flush but we *might* > > gain slightly if an allocation succeeds because a watermark check passed > > when the counters were updated. It's a definite hit for a possible gain > > though which is not a great trade-off. Would need some performance testing. > > > > I still think my patch on adjusting thresholds is our best proposal so > > far on how to reduce Shaohua's performance problems while still being > > safer from livelocks due to memory exhaustion. > > OK, I will try to explain a detai of my worry. > > Initial variable ZVC commit (df9ecaba3f1) says > > > [PATCH] ZVC: Scale thresholds depending on the size of the system > > > > The ZVC counter update threshold is currently set to a fixed value of 32. > > This patch sets up the threshold depending on the number of processors and > > the sizes of the zones in the system. > > > > With the current threshold of 32, I was able to observe slight contention > > when more than 130-140 processors concurrently updated the counters. The > > contention vanished when I either increased the threshold to 64 or used > > Andrew's idea of overstepping the interval (see ZVC overstep patch). > > > > However, we saw contention again at 220-230 processors. So we need higher > > values for larger systems. > > So, I'm worry about your patch reintroduce old cache contention issue that Christoph > observed when run 128-256cpus system. May I ask how do you think this issue? > It only reintroduces the overhead while kswapd is awake and the system is in danger of accidentally allocating all of its pages. Yes, it's slower but it's less risky. -- Mel Gorman Part-time Phd Student Linux Technology Center University of Limerick IBM Dublin Software Lab -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to majordomo@xxxxxxxxxx For more info on Linux MM, see: http://www.linux-mm.org/ . Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>