Felip,
Slurm provides a lot of control over the allocation of CPU resources. It's 
certainly possible to run two or more jobs on the same node at the same 
time with each job allocated a different set of CPUs.  For more 
information, see the CPU Management Guide:
http://www.schedmd.com/slurmdocs/cpu_management.html
Martin Perry
Bull Phoenix




From:   Felip Moll <[email protected]>
To:     "slurm-dev" <[email protected]>, 
Date:   02/12/2013 12:33 PM
Subject:        [slurm-dev] Control task distribution across nodes



Hello SLURM list!

This is my second question about job binding. As I said to another thread 
I set up CR_Core_Memory and Slurm 2.4.3 on a small cluster of 15 compute 
nodes, 2 quad-cores each.


Suppose that the cluster is empty of jobs. When a user sends one serial 
job to run it goes to node0, processor0, core 0. All right.

If the same or another user sends another serial job, Slurm sends it to 
node1, processor0, core0. In terms of performance it is optimal because it 
separates the tasks as it can.

In terms of energy for example, it is not optimal because node1 could be 
suspended while both 2 tasks gone to node 0.

An other problem that I found the other day is that one single user sent 
15 jobs to the cluster and this jobs were spreaded into 15 nodes thus 
making impossible to send then a job asking for 4 cores.

Is there any solution here? Is it possible to control how Slurm distribute 
tasks?


Thank you all!,
Felip


Reply via email to