On Mon, 14 Jun 2010 12:17:41 +0100 Mel Gorman <mel@xxxxxxxxx> wrote: > SysBench > ======== > traceonly-v2r5 stackreduce-v2r5 nodirect-v2r5 > 1 11025.01 ( 0.00%) 10249.52 (-7.57%) 10430.57 (-5.70%) > 2 3844.63 ( 0.00%) 4988.95 (22.94%) 4038.95 ( 4.81%) > 3 3210.23 ( 0.00%) 2918.52 (-9.99%) 3113.38 (-3.11%) > 4 1958.91 ( 0.00%) 1987.69 ( 1.45%) 1808.37 (-8.32%) > 5 2864.92 ( 0.00%) 3126.13 ( 8.36%) 2355.70 (-21.62%) > 6 4831.63 ( 0.00%) 3815.67 (-26.63%) 4164.09 (-16.03%) > 7 3788.37 ( 0.00%) 3140.39 (-20.63%) 3471.36 (-9.13%) > 8 2293.61 ( 0.00%) 1636.87 (-40.12%) 1754.25 (-30.75%) > FTrace Reclaim Statistics > traceonly-v2r5 stackreduce-v2r5 nodirect-v2r5 > Direct reclaims 9843 13398 51651 > Direct reclaim pages scanned 871367 1008709 3080593 > Direct reclaim write async I/O 24883 30699 0 > Direct reclaim write sync I/O 0 0 0 Hmm, page-scan and reclaims jumps up but... > User/Sys Time Running Test (seconds) 734.52 712.39 703.9 > Percentage Time Spent Direct Reclaim 0.00% 0.00% 0.00% > Total Elapsed Time (seconds) 9710.02 9589.20 9334.45 > Percentage Time kswapd Awake 0.06% 0.00% 0.00% > Execution time is reduced. Does this shows removing "I/O noise" by direct reclaim makes the system happy ? or writeback in direct reclaim give us too much costs ? It seems I'll have to consider about avoiding direct-reciam in memcg, later. BTW, I think we'll have to add wait-for-pages-to-be-cleaned trick in direct reclaim if we want to avoid too much scanning, later. Thank you for interesting test. Regards, -Kame -- 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>