I meet problem of lustre MDS. When I launch MPI job on lustre clients, all of ll_mdt_xxx threads on MDS go into "D" state for a while.During that time, there is no response when I do some operattion on lustre, such as "df","ls", or some other commands. After some seconds, ll_mdt threads come back then everying is OK. More clients involed, more time we have to wait, even hundreds of seconds. I have 16 ll_mdt threads, and I have try to tune it to 32 or 64. It seems not work. Is there a bottleneck on MDS when lots of client read same one binary to exec? and Is there someone can give me some advice ? thanks -sky -------------- next part -------------- An HTML attachment was scrubbed... URL: http://mail.clusterfs.com/pipermail/lustre-devel/attachments/20070313/a949e32a/attachment.html