On 06/29/2010 08:25 PM, Ying Han wrote:
apologies if you got this email twice, the first emails seems not getting through :( This is the discussion we would like to have on the upcoming Linux VM summit. Problem: Google runs large scale of machines and each machine runs Linux. We try to achieve higher utilization by better bin-packing of jobs on existing systems and for this we depend on having accurate resource usage estimation. Linux VM subsystem is designed in a way that it tries to allocate every single page available by filling up page cache pages. Some of the pages might be touched once and never touched again. Pageout deamon(kswapd) only evicts pages under memory pressure, so pages which are actually stale will end up taking memory space. It would be nice to have a way to measure the portion of working set for each process periodically. A user-land resource management program can trigger reclaim of the stale pages making room for packing more jobs any time.
Something like this functionality could also be useful for virtualization, kicking off garbage collection in JVMs and other runtimes, as well as resizing other workloads that cache data... I would like to discuss this topic so we can figure out the kind of functionality needed to achieve what everybody wants. -- All rights reversed -- 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>