On Sat, May 17, 2014 at 5:40 PM, Michal Pazdera <michal.pazdera at gmail.com> wrote: > Hi everyone, > > I wonder if CephFS is able to read from all replicas simontaniously, that > will result in doubled read performance if > a replica of 2 has been used. Nope, definitely not. > I have done some humble testing on 5PCs (2OSD > (each on 1PC) + 1MON/MDS PC + 2 CephFS clients). > I have did two configurations: > > 1) default from ceph-deploy, replication 2, 4MB objects with 1 stripe count > and 4MB stripe size. > 2) custom conf without replication, 4MB object with 4 stripe count and 1MB > stripe size. > > If replication is used write performance is around 30MB/s and read 100MB/s. > If not write is around 55MB/s and read is suprisingly lower 95MB/s. I would > expect to be around 160 or 180 since i have used two hard drives. > > journal was on the same OSD > > So i think with replication the data was served from both replicas, but im > not sure. I have tried to find some more information about > this but couldnt find any. I think you're just seeing random variation based on on-disk layouts or something. 95 and 100 are not far off for tests run over multiple hard drives. And while there are a number of other things it could be (lots of stuff bottlenecks around 100MB/s), it might just be the combination of your network throughput (which in practice can get ~115MB/s, if it's gigabit) and occasional latencies. > I have one last question. Does journals impact read performance? I mean if > journal is on the same OSD, the write performance is > two times slower because write is done to journal and then onto OSD. Does > something simalar happen for reads to? I think not > coz it seems silly to me. Nope, the journals have no direct impact on reads. -Greg Software Engineer #42 @ http://inktank.com | http://ceph.com
 |