[torqueusers] Re: [Mauiusers] resource contention on multi core
James A. Peltier
jpeltier at cs.sfu.ca
Thu Oct 11 10:47:51 MDT 2007
Kevin Van Workum wrote:
> I've seen some discussion on this topic, but would like to hear
> people's current opinions and solutions. When using multi core nodes,
> what is the best way to prevent one user from using more than his
> portion of the node's resources. For example, what stops him from
> requesting a single processor on a quad core node, and then using all
> four processors? Or if I'd like to restrict him to 1/4 of the
> available physical memory, how can I do that. Other local resources
> for which contention may exist are the network and disk i/o. How do
> you deal with these issues?
> mauiusers mailing list
> mauiusers at supercluster.org
This can easily be done with Maui assuming the using something like
and various other options combined. In this case, allow only 1
processor equivilent unless there is no one else queued. There are
various others and this is just pointed out as an example.
Resources like memory and I/O contention however are more difficult.
For example a particular job could write out GigaBytes or TerraBytes of
data tieing up I/O quite easily, but it still needs to be done. In this
case adding HBAs and spindles is the only way to ease the problem.
Additionally, a person could request 1 PE and all of the memory in the
system. Is that wasteful, well somewhat, 3/4 of the machines compute
isn't used, but that's the way things work.
Remember, the idea is to keep everyone as close to equally unhappy as
possible. You will rarely find a grounds onto which everyone is happy,
unless you only have one user of the cluster :)
James A. Peltier
Technical Director, RHCE
SCIRF | GrUVi @ Simon Fraser University - Burnaby Campus
Phone : 778-782-3610
Fax : 778-782-3045
Mobile : 778-840-6434
E-Mail : jpeltier at cs.sfu.ca
Website : http://gruvi.cs.sfu.ca | http://scirf.cs.sfu.ca
MSN : subatomic_spam at hotmail.com
More information about the torqueusers