On Fri, Aug 09, 2013 at 11:32:34PM +0200, Valerio Pachera wrote: > 2013/8/9 Liu Yuan <namei.unix at gmail.com>: > > The long wait time doesn't attribute to how sheepdog get the disk usage, but how > > busy sheep daemon is. In the recovery, sheep main loop is saturated by recovery > > requests and the 'md info' get kind of starved. > > Ok, clear. > During recovery sheep was using no more than 20% of cpu and disks are > busy between 30% and 60% (on the weak server). > With no recovery and no guests running, it takes 30 seconds. > > root at sheepdog001:~# collie node recovery > Nodes In Recovery: > Id Host:Port V-Nodes Zone Progress > > root at sheepdog001:~# time collie node md info --all > Id Size Used Avail Use% Path > Node 0: > 0 166 GB 154 GB 12 GB 92% /mnt/sheep/dsk01/obj > 1 465 GB 319 GB 147 GB 68% /mnt/sheep/dsk02 > 2 1.8 TB 1.1 TB 716 GB 61% /mnt/sheep/dsk03 > Node 1: > 0 166 GB 84 GB 82 GB 50% /mnt/sheep/dsk01/obj > 1 465 GB 299 GB 166 GB 64% /mnt/sheep/dsk02 > 2 1.8 TB 1.2 TB 667 GB 64% /mnt/sheep/dsk03 > Node 2: > 0 2.7 TB 1.9 TB 873 GB 68% /mnt/sheep/dsk02 > Node 3: > 0 465 GB 215 GB 250 GB 46% /mnt/sheep/dsk03 > 1 1.8 TB 1.4 TB 449 GB 75% /mnt/sheep/dsk04 > > real 0m30.504s > user 0m0.000s > sys 0m0.012s > > Correct me if I'm wrong but md info is calculated one after the other, > not in parallel right? No, not in parallel as you guessed. Collie's performance will be enhanced in the later version. Thanks Yuan |