Dolphin Wulfkit

[Joachim Worringen]

Greg Lindahl wrote:
> 
> On Sat, May 04, 2002 at 02:43:52PM +0200, Steffen Persvold wrote:
> 
> > Well ScaMPI does "grouping" too I guess. Examples :
> >
> > # /opt/scali/bin/mpimon someapp -- node1 node2 nnode1 node2
> >
> > Here rank 0 and 2 would run on node1 but when they are communicating
> > with eachother shm is used, same with rank 1 and 3 on node2.
> 
> This is "round robin", not grouping. The question was about your MPI's
> default policy. The correct default policy is to put ranks 0 and 1 on
> the same node, not 0 and 2, because that keeps nearest neighbor codes
> happy even if they don't call the MPI cartesian routines...

Yes, this is round robin. And the example which followed in Steffen's
posting (the one you didn't quote) was grouping. I assume Steffen gave
both examples to illustrate that it's up the command line to decide how
to map processes to nodes with ScaMPI, and he left it as an exercise for
the reader to assign "round robin" and "grouping" to these examples. 

So, as Patrick already pointed out, it is up to the scheduling system
(or the user at the command line) to specify a mapping. Usually, the
scheduler will use grouping mapping. But you can do as you like which
may be useful in performance comparisons.

> > AFAIK the PMB benchmark uses rank 0 and 1 when doing ping-ping and
> > ping-pong tests,
> 
> So a "round robin" default breaks PMB if you aren't careful. Which is
> unfortunate.

Do you have a definition of "breaking PMB"? 

 Joachim

-- 
|  _  RWTH|  Joachim Worringen
|_|_`_    |  Lehrstuhl fuer Betriebssysteme, RWTH Aachen
  | |_)(_`|  http://www.lfbs.rwth-aachen.de/~joachim
    |_)._)|  fon: ++49-241-80.27609 fax: ++49-241-80.22339