|
||||||
|
||||||
SGI Altix Debugging, Performance, and Timing FAQ |
||||||
|
User Information Home
 NCSA's Help Desk is available 24 hours a day, seven days a week, 365 days a year:
How do I find out how much memory my program used?See the documentation on Checking memory use. How can I tell if my threaded or OpenMP program used multiple cpus?Run the program with /usr/bin/time and compare the user and elapsed time reported by /usr/bin/time and note also the %CPU [comparing cputime and wall clock time from "qhist jobid" would also work]. Here's sample output from /usr/bin/time for a job that requested 4 cpus and ran with 4 threads. The batch script contained a line like this: /usr/bin/time a.outThe stderr file contained this output from /usr/bin/time: 176.04user 0.28system 0:44.93elapsed 392%CPU (0avgtext+0avgdata 0maxresident)k 0inputs+0outputs (0major+4569minor)pagefaults 0swapsNote that 176.04user/44.93elapsed=3.918 which corresponds to the 392 %CPU. The sample program used 392/400 or about 98% of its 4 cpus. Qhist for the job showed:
...
Usage:
wall clock: 00:00:56
cputime: 00:02:59
That works out to be (2min*60sec+59sec)/56sec = 3.20 . The wall clock time for a short job contains some batch system overhead, so the ratio is a little smaller than it should be for the test job, but the qhist data confirm that >3 cpus were utilized.
I just received an email stating that my job started. However, when I checked its status, it was still in the queue. I have not had problems like this before!This can happen when, at times, the scheduler attempts to start the job even though sufficient resources to run the job are unavailable. The job should start as soon as the resources are available. I am trying to debug a simulation on cobalt. It seems the debugger of choice is totalview even though that is quite painful via X11 forwarding. Is it possible to run the debugger in a text console? |
||||||
|
|
||||||
