I guess we could. I put it last so that you would always defer to an entity that knew better (eg, the RM).
From: Ethan Mallove [mailto:email@example.com]
Sent: Thursday, June 28, 2007 11:21 AM Eastern Standard Time
To: General user list for the MPI Testing Tool
Subject: Re: [MTT users] max_np on multi-processor nodes
On Thu, Jun/28/2007 08:07:11AM, Jeff Squyres wrote:
> On Jun 28, 2007, at 12:42 AM, Ethan Mallove wrote:
> > env_max_procs uses this order of precedence to get a
> > "max_np".
> > 1. Various resource managers
> > 2. If a global hostfile was specified
> > 3. If a global hostlist was specified
> > 4. If a global max_np was specified
> > Shouldn't 4 be at the top of the list?
> > I'm doing single node testing, but the fact that I have only
> > one node listed in hostlist causes env_max_procs to return
> > "1". I'd like to override the 1 value with the number of
> > CPU's on the single node I'm running on.
> Is the :N notation not sufficient?
> # OMPI Core: if you would rather list the hosts individually on the
> # mpirun command line, list hosts here delimited by whitespace (if you
> # have a hostfile listed above, this value will be ignored!). Hosts
> # can optionally be suffixed with ":num", where "num" is an integer
> # indicating how many processes may be started on that machine (if not
> # specified, ":1" is assumed). The sum of all of these values is used
> # for &env_max_np() at run time. Example (4 uniprocessors):
> # hostlist = node1 node2 node3 node4
> # Another example (4 2-way SMPs):
> # hostlist = node1:2 node2:2 node3:2 node4:2
> # The "hostlist" value is ignored if you are running in a scheduled
> # environment or if you have specified a hostfile.
Oops. That's what I need.
But re the above precedence list, if I wanted to
oversubscribe using a hostlist and a manually overriden
max_np param, wouldn't we still need to put item #4 at the
top of the list?
> Jeff Squyres
> Cisco Systems
> mtt-users mailing list
mtt-users mailing list