Conference msbcs::hpc

507.0. "Don't use -pref_comm" by NNTPD::"[email protected]" (Richard Warren) Thu Jun 05 1997 09:56

Normally PSE will choose the method of communication based
on its database for the particular partition.  In the absence
of a real database, i.e. "simple farms", PSE uses the
following prioritized list:  shm,mc,atm,fddi,ethernet

Users are given the ability to override this automatic selection
as you did in your example by using the -pref_comm switch: 
      appl3.hrz.uni-siegen.de> pi_example -pref_comm shm

which tells PSE to utilize *ONLY* shared memory (shm) as a means
of communicating.  The errors occur because PSE will load-balance
the execution over the set of available machines, and it is very likely 
that some number of HPF peers will be on a "remote" machine unless there's
only
a single SMP box in the partition.  To be more precise, if your application
is compiled for 4 peers (f90 -wsf 4) and you have two machines, each with
2 processors, and interconnected via memory channel, then PSE will "allocate"
2 peers on each machine.  The "messaging" between peers on the same "host"
will be SHM and between hosts will be MC.   In a different scenario, if
you choose to run on a single host with "at least" 4 cpus, then SHM will be
used between all peers.  To see this explicitly, run the application with
the "-connections" switch.   Note that SHM does NOT work between cluster
nodes!

Hope this helps,
Richard
[Posted by WWW Notes gateway]