While testing my 1.6 beta5 setup using writemany from /lustre/tests, I saw some interesting numbers that might be related to directory traversal, but can''t fully understand it. All I did was to run "writemany <dir> 1 1". By giving different <dir>, I got the number of files per sec for local file system, NFS, Lustre loal (server and client runnig on the same node) and Lustre remote (server and client running on different nodes). I also used different directory traversal level. Here are the numbers: traversal level Local NFS Lustre(local) Lustre(remote) ----------------------------------------------------------------------------------------------------- 0 ~20,000 ~20,000 ~20,000 ~20,000 1 ~20,000 ~140 ~900 ~200 2 ~20,000 ~140 ~800 ~200 3 ~20,000 ~140 ~700 ~200 4 ~20,000 ~140 ~700 ~200 Questions: 1. Why the huge drop from 0 to 1 level? 2. Why the numbers level out for further directory traversal? 3. Has any formal comparasion tests been done between Lustre and other districuted file Systems (i.e., NFS, GFS etc)? 4. Are there any test results available for running Lustre again a benchmark suck as iozone, bonnie etc? Thanks Lin shen -------------- next part -------------- An HTML attachment was scrubbed... URL: http://mail.clusterfs.com/pipermail/lustre-discuss/attachments/20061120/c291353f/attachment.html