On Wednesday 09 May 2012 11:01:00 Jens Nauber | tyclipso. net wrote: > we have set up a replicated gluster volume on 3 glusterservers, 2 of them > act as well as gluster clients using NFS-Mode. > > It seems we have a performance problem especially in read access, compared > by Iozone-Test for GlusterFS and local filesystem (testfiles attached) Comparing a network-filesystem to a local filesystem is a bit unfair, isn't it? You "think" reading happens via local disk, but actually reading happens across the network when you use nfs-mount. Even when you use the fuse-mount locally, you still have a userspace-component (which has only a very little influence). And while reading of big contents all happens locally and is thus fast, accessing files (small or big, read or write) has an overhead as the attributes of the remote bricks are checked too to see if the volume is consistent and trigger a self-heal if needed. So while reading big files from local fuse-mount is fast, reading many small files still has the network-penalty of any distributed and/or replicated system. A better comparison is kernel-nfs against gluster with its userspace-nfs. And maybe also compare gluster to moosefs or ceph or lustre or any fs above drbd... Have fun, Arnold -------------- next part -------------- A non-text attachment was scrubbed... Name: not available Type: application/pgp-signature Size: 198 bytes Desc: This is a digitally signed message part. URL: <http://gluster.org/pipermail/gluster-users/attachments/20120509/1d26a628/attachment.pgp>
