Previous: Benchmark_Queue
Next: User_Assistance
View by Date; View by Name
Batch Memory Limits

                       Setting Batch Memory Limits 
                             May 18, 2005

NCSA has recently experienced problems with our SGI Altix system which
have been traced to batch jobs which used more memory than requested.

To specify a memory limit in your batch script, please add a line like
the following near the beginning of your batch script.

  #PBS -l mem=5000mb

The memory specified is the total memory that will be used by all the
processors assigned to the job, not the memory per processor. It can be
given in units of Kbytes (kb), Mbytes (mb) or Gbytes (gb).

For more information about writing batch scripts on the SGI Altix, see:

  http://www.ncsa.uiuc.edu/UserInfo/Resources/Hardware/SGIAltix/Doc/Jobs.html

and see the sample scripts in /usr/local/doc/pbs/samples.

Please estimate your memory requirements carefully before you run a job
the first time with a new problem size.  The "size" command reports the
static memory needed but you may also need look through the source code (if
available) to estimate the amount of memory that is dynamically allocated.
If you are running a third party code, you may need help from the developer
or vendor to estimate memory needs.

We recommend that you add about 20% to your estimate for the first run,
then measure the actual memory usage to help give a more accurate limit
for succeeding runs.

To see a list of jobs that you have run, use the "qhist" command.  To
determine the amount of memory that was used by a specific job, use
the "qhist jobid" command. For example:

  % qhist 6006

  Scanning PBS raw accounting records: 03/28/2005 - 05/18/2005

  Compute Host:       co-compute2/0*8
  JobId:              6006
  JobName:            script8
  User:               davem
  Project:            aed
  Queue:              standard

  Job limits:
    wall clock:       18:00:00
    Requested CPUs:   8
    Available CPUs:   8
    Requested Memory: 3800mb
    Available Memory: 32gb

  Queued:             05/18/05 13:42
  Started:            05/18/05 13:43
  Ended:              05/18/05 13:49

  Usage:
    wall clock:       00:05:17
       cputime:       00:04:34
           SUs:       0.08
        memory:       707.02M

On the last line, the qhist command reports the maximum value for RSS
(memory actually used). You should add 5-10% to the value as your
estimate for future runs.

If you have any questions, please contact NCSA's Consulting Office at
consult@ncsa.uiuc.edu or (217)244-0640.
SGI Altix 3700:usr/news/Batch_Memory_Limits
Last Modified: May 18, 2005