Hey Samuel, That changed something, for the better. :-) Your test-version, with wbthrottle off: # ceph-osd --version ceph version 0.67.1-18-g3fe3368 (3fe3368ac7178dcd312e89d264d8d81307e582d8) # ceph --admin-daemon /var/run/ceph/ceph-osd.1.asok config show | grep wbthrottle_enable "filestore_wbthrottle_enable": "false", # rbd bench-write test --io-pattern=rand bench-write io_size 4096 io_threads 16 bytes 1073741824 pattern rand SEC OPS OPS/SEC BYTES/SEC 1 838 837.36 2283825.44 2 1679 827.66 2309860.21 3 2488 829.30 2320967.64 4 2812 655.55 1847713.58 5 3301 659.64 1849006.69 <...> 120 54113 450.83 1271297.14 121 54847 453.13 1277957.21 122 55502 454.93 1281769.31 123 55768 451.09 1271512.03 124 55853 450.42 1269452.15 126 55889 442.22 1246494.89 127 56351 443.71 1250308.66 128 57135 446.36 1257746.91 129 57445 444.03 1251566.31 130 57944 445.72 1256426.99 Now with the wbthrottle-option enabled: # ceph-osd --version ceph version 0.67.1-18-g3fe3368 (3fe3368ac7178dcd312e89d264d8d81307e582d8) # ceph --admin-daemon /var/run/ceph/ceph-osd.1.asok config show | grep wbthrottle_enable "filestore_wbthrottle_enable": "true", # rbd bench-write test --io-pattern=randbench-write io_size 4096 io_threads 16 bytes 1073741824 pattern rand SEC OPS OPS/SEC BYTES/SEC 1 822 821.23 2303875.45 2 1510 738.54 2021390.37 3 2224 740.29 1989216.54 4 2843 708.01 1912598.24 5 3294 652.35 1774048.26 <...> 120 59470 495.58 1388776.41 122 60067 489.29 1370954.96 123 60083 487.74 1366434.49 124 60449 487.49 1366477.77 125 61122 488.97 1370480.73 126 61679 489.52 1372290.03 127 62195 489.69 1372523.64 128 62608 489.02 1370226.98 129 62655 485.68 1360719.66 130 62688 482.01 1350560.76 Slightly faster than with wbthrottle disabled.. For comparison, one more time with Cuttlefish-OSDs: # ceph-osd --version ceph version 0.61.7 (8f010aff684e820ecc837c25ac77c7a05d7191ff) # rbd bench-write test --io-pattern=rand bench-write io_size 4096 io_threads 16 bytes 1073741824 pattern rand SEC OPS OPS/SEC BYTES/SEC 1 1074 1073.80 3079631.51 2 2121 1060.21 3001542.58 3 3111 845.59 2422586.94 4 3132 745.40 2132915.48 5 3336 665.33 1904205.15 <...> 120 63700 530.81 1530767.77 121 64781 535.37 1544861.37 122 64944 532.10 1535398.54 123 65312 527.70 1522971.49 124 65526 528.42 1525134.71 125 66160 529.12 1527158.94 126 67214 533.44 1539955.47 127 67855 534.29 1541965.13 128 67939 528.79 1526270.55 129 68178 528.44 1525239.57 130 68811 529.24 1527681.84 Cuttlefish still wins, but by a much smaller margin now. Looks like we're onto something. The fdatasync seems to be the key here, rather than disabling wbthrottle. Regards, Oliver On 23-08-13 19:53, Samuel Just wrote: > I pushed a branch, wip-dumpling-perf. It does two things: > 1) adds a config filestore_wbthrottle_enable (defaults to true) to > allow you to disable the wbthrottle altogether > 2) causes the wbthrottle when enabled to fdatasync rather than fsync. > > Can you rerun the random workload with that branch with > filestore_wbthrottle_enable on and then off? > -Sam > > On Fri, Aug 23, 2013 at 10:26 AM, Oliver Daudey <oliver@xxxxxxxxx> wrote: >> Hey Sage, >> >> I'm all for it and will help testing. >> >> >> Regards, >> >> Oliver _______________________________________________ ceph-users mailing list ceph-users@xxxxxxxxxxxxxx http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com