Hi huangql, which lustre version you using?
On Mon, 2009-08-17 at 15:23 +0800, huangql wrote: > Hi, all > > Our system run well past two weeks, However, we found there are some > computing nodes which has so many threads named "ll_imp_inval", and > the load average of the clients(computing nodes) is up to 28. As a > results, Users can't submit jobs to the clients. I read the source > file(import.c) and In my opinion, when each ptlrpc-connect-import or > ptlrpc-import-recovery, the ll_imp_inval thread is triggered. So if > the server or clients have something wrong, the thread will not exit. > Is it right? > don't. ll_imp_inval is evictor thread - which started if client isn't connected to server (MDS or OST) until recovery is finished, and server ask client to flush own staled data. > > we run 'ps -aux | grep ll_imp_inval' ,the results as follows: > > root 22568 0.0 0.0 0 0 ? D Aug13 0:00 [ll_imp_inval] > root 22569 0.0 0.0 0 0 ? D Aug13 0:00 [ll_imp_inval] > root 22570 0.0 0.0 0 0 ? D Aug13 0:00 [ll_imp_inval] > root 22571 0.0 0.0 0 0 ? D Aug13 0:00 [ll_imp_inval] > ... is possible to see output from sysrq-t (echo t > /proc/sysrq-trigger) ? > > We had check out the log, and found the main messages as follows,and > in other nodes we can get the client evicted messages: > > > > Thank you for your help in advance and I hope receive your letter as > soon as possible. > > > Best wishes, > Sarea > > > 2009-08-17 > > ______________________________________________________________________ > huangql > _______________________________________________ > Lustre-discuss mailing list > [email protected] > http://lists.lustre.org/mailman/listinfo/lustre-discuss _______________________________________________ Lustre-discuss mailing list [email protected] http://lists.lustre.org/mailman/listinfo/lustre-discuss
