RA_Book - page 391 of 391

First PagePrevious PageNext PageLast PageTable of ContentsPrintSearch

Date and Author(s)

Change for running multiple jobs on a node


Recently there have been some changes to the scheduler configuration
on Mio. To address some performance issues the default environment 
is now such that when a job starts it is sure to grab all of the memory 
the node for running your script.  

This has a side effect that if you try to run multiple jobs on a node
your first job grabs all of the memory and the scheduler thinks there
is not enough memory for additional jobs.  So it waits.

If you would like to run multiple jobs on a node you need to specify the
memory allocated to each job.   his can be done by adding an additional
SBATCH line to your script.  For example the line:  

#SBATCH --mem=8000

This says each job gets 8Gb of memory.

This will allow multiple jobs to be placed on a node.  For example:

[tkaiser@mio001 asubdir]$ sbatch --nodelist=compute063 array
Submitted batch job 3129073
[tkaiser@mio001 asubdir]$ sbatch --nodelist=compute063 array
Submitted batch job 3129074
[tkaiser@mio001 asubdir]$ squeue -u tkaiser
             JOBID PARTITION     NAME     USER ST       TIME  NODES NODELIST(REASON)
           3129074   compute ridgeapp  tkaiser  R       0:07      1 compute063
           3129073   compute ridgeapp  tkaiser  R       0:12      1 compute063
[tkaiser@mio001 asubdir]$ 

The web page http://geco.mines.edu/prototype/Who_owns_nodes_on_Mio/index.shtml 
and the command 

/opt/utility/slurmnodes -fRealMemory

will show how much memory is available on nodes