On Thursday, March 10, 2011 at 7:45 AM, Sylar Shen wrote: > Hi Gregory, > Here are my conditions: > 1. I use in-kernel client(my OS is Fedora 14) > 2. Replication level is 1, so there are 2 identical files(the original > and the copy). > 3. The network card on each server and the switch are 1Gb/s. > > You said you'd expect mush faster results on the buffered write test, > maybe approaching the network interface limits. > I think you may misunderstand what I mean. > I used the in-kernel client to mount Ceph(mount -t ceph 192.168.1.11:/ > /mnt/ceph) . > I re-exported NFS on the client using "exportfs client:/mnt/ceph -o > fsid=1234,rw,no_squash_root" > Besides, I used another server to connect to the client by NFS > protocol and did the write test using dd command(with buffer). > And I got Ceph=46.6MB/s while Gluster=39.3MB/s, the speed looked similar. > So this confuses me. Because I think even if Ceph is re-exported by > NFS protocol, it should be a lot faster than Gluster. > Did I do something wrong or was it really influenced by the speed of > the switch or the router? > Thanks in advance! Hmm. We haven't run a lot of tests via NFS re-export, but Sage doesn't think it should impact performance too much. Could you try running those dd tests on the Ceph mount directly? We generally see results of 90MB/s to 110MB/s when running simple sequential write tests like that. It's possible that the machine you're using to export NFS from doesn't have a full-duplex card and so you're saturating its network interface, which would yield about 45MB/s writes like you're getting. -Greg -- To unsubscribe from this list: send the line "unsubscribe ceph-devel" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
