{Def,Max}MemPer{CPU,Node} seem used to avoid over-subscribing memory and
causing paging. Although I read slurm.conf, I have further questions about them.
1. MaxPerCPU and MaxPerNode look for setting maximum available memory
sizes. What exactly DefPerCPU and DefPerNode mean?
2. What happens if a single process requests more memory larger than the
specification? How would it work differently for each specified values?
3. If each node has a different memory size, how can I set different
DefMemPerNode or MaxMemPerNode value for each node?
4. Why would {Def,Max}MemPerCPU generally be used with select/cons_res and
{def,Max}MemPerNode generally be used with select/linear? Would it be okay to
use {def,Max}MemPerNode with select/cons_res?
Thank you.
Joo-Kyung
From: David Bigagli
[http://lists.schedmd.com/cgi-bin/dada/mail.cgi/r/slurmdev/577441466023/]
Sent: Tuesday, March 12, 2013 12:03 PM
To: slurm-dev
Subject: [slurm-dev] Re: Memory swapping, and transition delay issues.
Hi, the problem of memory over-subscription is discusses in 'man slurm.conf'.
Have a look at
DefMemPerCPU, DefMemPerNode and the suggested configuration when using
CR_CPU_Memory.
/David
On Tue, Mar 12, 2013 at 3:15 PM, Joo-Kyung Kim <[email protected]
<http://lists.schedmd.com/cgi-bin/dada/mail.cgi/r/slurmdev/690317113748/> >
wrote:
Hi,
I am using SLURM 2.4.0.
When I run many memory intensive processes at the same time using a slurm job,
there are a lot of memory swappings by kswapd0.
I changed SelectTypeParameters from CR_CPU to CR_CPU_MEMORY (SchedulerType =
sched/backfill and SelectType = select/cons_res aren’t changed), I am not sure
there would be some significant reduction of memory swapping.
Would there be a good way to automatically reduce the high memory swapping?
Another question: It seems that it takes more than 15 seconds for completing
jobs to be completed jobs so that the next jobs can be runned on a cluster with
8 nodes.
I wonder it’s normal that the transition takes some time or would there be any
problem?
Thank you.
Joo-Kyung
<http://lists.schedmd.com/cgi-bin/dada/mail.cgi/r/slurmdev/449583192704/>
<http://lists.schedmd.com/cgi-bin/dada/mail.cgi/r/slurmdev/858117913508/>
