On Mon, 21 Feb 2011, Joe Landman wrote: > On 02/21/2011 09:53 AM, David Lloyd wrote: >> Likewise, we're aware of the very poor performance of gluster with small >> files. We serve a lot of large files, and we're now moved most of the small >> files off to a normal nfs server. Again small files aren't known to break >> gluster are they? > > Small files are the bane of every cluster file system. We recommend using > NFS client with GlusterFS for smaller files, simply due to the additional > caching you can get out of the NFS system. > Those who are looking for better metadata performance might want to see if MooseFS fits their needs. It uses a metadata server which caches the entire filesystem metadata in RAM, so it seems to be very responsive. I have my home directory on GlusterFS and I did a quick trial on MooseFS. du on GlusterFS (native FUSE, not NFS) was 1 minute 8 seconds, du on MooseFS was 2 seconds. Good metadata performance should equate to good small-file performance (but you'll want to run your own tests to be sure). Note that MooseFS stores data with a fixed ~64KB block size, so it will be somewhat wasteful of space on very small files. Thanks, Brent Nelson Director of Computing Dept. of Physics University of Florida
