Thursday, 2 August 2007

Fun at Manchester SL4, lcg_util and pbs

In the midst of getting a successful upgrade-to-SL4 profile working, we upgraded our release of lcg_util from 1.3.7-5 to 1.5.1-1 this prooved to be unsuccessful, SAM test failures galore. After looking around for a solution on the internet I settled for rolling back to the previous version, thanks to the wonders of cfengine this didn't take long, and happily cfengine should be forcing that version on all nodes.

This Morning I came in to find we were again failing the SAM tests, this time the ever-so-helpful
"Cannot plan: BrokerHelper: no compatible resources"

Pointing to a problem deep in the depths of the batch system. Looking at our queues (via showq), there were a lot of Idle jobs yet more than enough CPUs. The PBS logs revealed a new error message,
Cannot execute at specified host because of checkpoint or
stagein files
for two of the jobs, eventually I managed to track it down to a node. Seeing as there wasn't any sign of the job file anymore, and pbs was refusing to re-run the job on another node, I had to resort to the trusty `qdel`, after thinking about it for the barest of moments, all of the Idle jobs woke up and started running.

Just for some gratuitous cross-linking, Steve Traylen appears to have provided a solution Over at the Scotgrid blog.

No comments: