From: Maciej Golebiewski <s30125@hotbot.com>
Newsgroups: comp.parallel.mpi
Subject: Re: Building MPICH on SMP: ch_shmem or ch_p4 (?)
Date: Thu, 29 Apr 1999 15:19:13 +0200
Organization: Organization? Me???
Message-Id: <37285C51.71E7C9CF@hotbot.com>
References: <7g9jq9$trq$1@vvv.srcc.msu.su>
Mime-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Transfer-Encoding: 7bit
Xref: ukc comp.parallel.mpi:5010


Alexander N. Andreyev wrote:
> 
> I am building MPICH on an SMP machine.There are two methods of configuring
> MPICH on multiprocessors, proposed in documentation:
> 
> 1) device=ch_shmem
> and 2) device=ch_p4, comm=shared
> 

I have never tried ch_p4 with shmem, but on our machines ch_lfshmem is
much faster (at least the latency) than ch_shmem.

> What are the differences in MPI communications between these two modes?
> Which one should perform better?

If you are using a group of SMPs, go for ch_p4 with shared, to be able
to do also inter-node transfers. If all you have is only one SMP box
and all the processes will run only on that one machine, use ch_lfshmem
as it will be probably faster than ch_p4 (because it's much simpler
and smaller code).