I would like to sollicit comments on the following patches. In order to optimize job placement accross many machines, we collect memory utilization statistics for each cgroup on each machine. The statistics are intended to be compared accross many machines - we don't just want to know which cgroup to reclaim from on an individual machine, we also need to know which machine is best to target a job onto within a large cluster. Also, we try to have a low impact on the normal MM algorithms - we think they already do a fine job balancing resources on individual machines, so we are not trying to mess up with that here. Patch 1 introduces no functionality; it modifies the page_referenced API so that it can be more easily extended in patch 2. Patch 2 introduces page_referenced_kstaled(), which is similar to page_referenced() but is used for idle page tracking rather than for memory reclaimation. Since both functions clear the pte_young bits and we don't want them to interfere with each other, two new page flags are introduced that track when young pte references have been cleared by each of the page_referenced variants. The page_referenced functions are also extended to return the dirty status of any pte references encountered. Patch 3 introduces the 'kstaled' thread that handles idle page tracking. The thread starts disabled; one enables it by setting a scanning interval in /sys/kernel/mm/kstaled/scan_seconds. It then scans all physical memory pages, looking for idle pages - pages that have not been touched since the previous scan interval. These pages are further classified into idle_clean (which are immediately reclaimable), idle_dirty_swap (which are reclaimable if swap is enabled on the system), and idle_dirty_file (which are reclaimable after writeback occurs). These statistics are published for each cgroup in a new /dev/cgroup/*/memory.idle_page_stats file. We did not use the memory.stat file there because we thought these stats are different - first, they are meaningless until one sets the scan_seconds value, and then they are only updated once per scan interval where the memory.stat values are continually updated. Patch 4 is a small optimization skipping over memory holes Patch 5 rate limits the idle page scanning so that it occurs in small chunks over the length of the scan interval, rather than all at once. Please note that there are known problems in these changes. In particular, kstaled_scan_page gets page references in a way that is known to be unsafe when THP is enabled. I'm still figuring out how to address this, but thought it may be useful to send the current patches for discussion first. Michel Lespinasse (5): page_referenced: replace vm_flags parameter with struct pr_info kstaled: page_referenced_kstaled() and supporting infrastructure. kstaled: minimalistic implementation. kstaled: skip non-RAM regions. kstaled: rate limit pages scanned per second. arch/x86/include/asm/page_types.h | 8 + arch/x86/kernel/e820.c | 45 +++++ include/linux/ksm.h | 9 +- include/linux/mmzone.h | 7 + include/linux/page-flags.h | 25 +++ include/linux/rmap.h | 78 ++++++++- mm/ksm.c | 15 +- mm/memcontrol.c | 339 +++++++++++++++++++++++++++++++++++++ mm/memory.c | 14 ++ mm/rmap.c | 136 ++++++++------- mm/swap.c | 1 + mm/vmscan.c | 18 +- 12 files changed, 604 insertions(+), 91 deletions(-) -- 1.7.3.1 -- 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/ . Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/ Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>