Ok -- I _do_ understand that lustre is a work in progress, and no I am not expecting things to perform or work jsut perfectly, but I am seeing things that do not make sense... I have read through the Kernel24Perf page, as well as the note at the end of the striping manula in lustre.pdf -- both seem to indicate that O_DIRECT should give better numbers than without -- also, I would expect that running locally, the performance would be close to that of regular ext3 speed. Cluster setup: Head node - Dual P3 550, 1G ram compute nodes - dual P2 450, 512M ram All have e100 Fast E cards. Right now, I am just testing on compute node. I am serving a 15G IDE drive, and get iozone numbers of ~20MB/s on ext3 on a local disk. I get ~12MB/s on lustre, either locally or from the remote node. When run remotely with ''-I'' to iozone, I get ~5MB/s.I have done an ''echo 0 > proc/sys/portals/debug'' to help speed things up, but nothing other than that. The network looks to be fully loaded from the client node to the head node when benchmarking. So here go the questions ( I seem to have lots of these :) ) 1) Why would O_DIRECT perform worse? 2) Why are there around 15 or 20 ll_ost_XX processes on the head node? When running iozone, it seems they all want cpu time, and it drives up the load pretty good. 3) Is there a way to specify mount options to ext3 for use in lustre? I noticed on the kernel24Perf page that it was using writeback and ordered mode. Nic -- Nicholas Henke Penguin Herder & Linux Cluster System Programmer Liniac Project - Univ. of Pennsylvania