[Lustre-discuss] Slow NFS/CIFS when exporting Lustre
Oleg Drokin
green at whamcloud.com
Mon Sep 26 05:06:58 PDT 2011
Hello!
On Sep 22, 2011, at 11:31 AM, Dan wrote:
> After researching this in the archives it seems 85% of Lustre's native performance is reasonable via NFS. Over GigE I'm seeing 114 MB/sec read/write via Lustre native. NFS writes at 44 MB/sec via NFS (about 38%) but reads at 95 MB/sec. CIFS is lucky to see 18 MB/sec writes after tuning Samba (12 MB/sec prior). I've tried tuning the Lustre client and NFS client and server but nothing impacts the performance except setting noatime on NFS, which gets another 1.5 MB/sec on writes. What are some of you seeing in this configuration? Suggestions?
One issue is likely frequent attempts to use xattrs (there is a patch to speed this up in recent Lustre, LU-144 I think it is) and flocks.
Also I heard to cifs only does small IO - 64kb packets over the wire (and corresponding small reads), if you have high latency link, that would
probably have pretty negative impact too.
Bye,
Oleg
--
Oleg Drokin
Senior Software Engineer
Whamcloud, Inc.
More information about the lustre-discuss
mailing list