 |
 |
Hi FSL folks, We're having a couple of competing issues that result in prematurely killed jobs in our SGE queues. In these analyses, we are utilizing FIR basis sets and creating contrasts with real EVs, resulting in a somewhat complex design (120 PEs, 420 COPEs). In several places, feat code estimates the duration of the job. However, incorrect estimation is causing problems on our cluster, either not allowing them to be submitted or killing them before they finish. Is there a way to make this estimation more accurate? Trying to run lower-level analyses, feat self-submits to the SGE, but somewhere the -T flag is set, giving an estimated time for completion. However, the analyses take less time than the estimated duration. When the estimated duration is too high, our queues reject the job, resulting in the following message: "Unable to run job: error: no suitable queues. Exiting." Increasing the maximum duration for the queue allows the job to run, but it completes in substantially less time than estimated. On the flip-side, trying to run higher-level analyses, featregapply is called to register the native space stats into MNI space. The amount of time that featregapply takes is about 3.5 hours when the registered images don't exist. But it still takes 1.5 hours even when the registered images already exist. Unfortunately, there is a flag estimating the duration of the feat2_pre call at 3600 seconds (1 hour). If I run the higher-level repeatedly when no one else is using the cluster, I can sometimes get the higher-level to complete (see https://www.jiscmail.ac.uk/cgi-bin/webadmin?A2=FSL;d7cb9097.1012 for details). We can't figure out a global fix / queue setting, and our sysadmin has been changing the queues around depending on whether we plan to run lower-level or higher-level analyses. Can you offer a more-permanent fix? Thanks, Greg ____________________________________________________________________ Greg Burgess, Ph.D. Research Associate, Institute of Cognitive Science University of Colorado - Boulder Email: [log in to unmask] University of Colorado - Boulder UCB 594 Boulder, CO 80309-0594