Hi Andreas! On Di, 2009-05-12 at 16:15 +0100, Andreas Roussos wrote: > Can you show us the contents of /etc/exports on your head node? (rw,sync,no_root_squash)
> I've had a similar problem in the past and managed to get better > write speeds by adding 'async' to my exports options (have a look > at `man exports`). Strange, I read http://nfs.sourceforge.net/ and in B6 they stated "For the Linux implementation of NFS Version 3, using the "async" export option to allow faster writes is no longer necessary. NFS Version 3 explicitly allows a server to reply before writing data to disk, under controlled circumstances." >>but<< in fact it gives me 100MB/s if I set the async option > Hope this helps! yes of course, I still don't understand if nfs3 in async mode with a 2.6 kernel is something dangerous. But now I know where the bottlenec is. I will also try the no_wdelay option thanks and regards sven > > Regards, > Andreas Roussos > Systems Administrator > Operating Systems Group > University College London > > Sven Sternberger wrote: > > Hello! > > > > I'm looking for a way to improve our nfs performance. > > We recently updated a small cluster (16 nodes). The nodes > > mount a share via nfs. > > > > I use nfs3 over tcp. I checked the disk write speed on the server > > (450mb/s) and the network thruput (980 mb/s). I tried udp, the deadline > > scheduler and large tcp buffers. > > > > SElinux and auditd is disabled, the server (Dell2950 w 8GB memory,PERC5) > > and my testnodes (Dell 1950 w 8 GB memory) both run 64bit SL5. > > > > What I see is >> always << ~20MB/s writing speed (reading is > > about 90mb which looks ok). I can see the 20MB/s also in > > iostat. > > > > Device: rrqm/s wrqm/s r/s w/s rkB/s wkB/s avgrq-sz > > avgqu-sz await svctm %util > > sdb 0.00 6511.63 0.00 83.39 0.00 25892.36 621.00 > > 143.20 1552.83 11.96 99.70 > > > > My write test is: > > time dd if=/dev/zero of=./testme bs=1024k count=1000 > > > > I would expect that the write speed is higher. So any advice or > > explanation would be helpful > > > > best regards! > > > > sven
