Newsgroups: comp.parallel.mpi,comp.parallel.pvm
From: Clark Dorman <clark@s3i.com>
Subject: Re: PVM Or MPI, which one has a brighter future?
Organization: PSINet
Date: 27 Jan 1998 16:18:51 -0500
Message-ID: <dyb01aa4k.fsf@elmo.s3i.com>


[cc'ed to the author]

yeson@online.sh.cn writes:
>    I am new in the parallel processing area. As I know, two development
> schemes MPI and PVM are now available. 

Actually, to date I have found, as basic approaches:
	- PVM, 
	- MPI,
	- Linda and C-Linda
	- P4, 
	- TCGMSG (is it still supported?), 
	- Condor (with process migration and checkpoint restart!)
	- and there probably are several others that I just haven't
	  found yet.

In addition, there are people trying to combine approaches: 
	- Pirhana (Lions!),
	- LAM (at Ohio Supercomputing Center, MPI with PVM
	  functionality added in, I think) (and Tigers)
	- Glenda (n-tuples built on PVM) (and Bears)

and last but not least

	- PVMPI (See http://www.cs.utk.edu/~fagg/mpiglue/) (oh My!)

They are all different and I am completely confused as to what I
should use.  But I've been playing with them a lot, especially PVM
because that's what I found first.  I need to report to my superiors
about the approach that we should take and give them a system design.
And have a good justification for my choices.  Soon.

I found the following helpful for understanding PVM versus MPI: 
	<http://www.epm.ornl.gov/pvm/PVMvsMPI.ps>
I found the following helpful for understanding PVM versus Linda: 
	<http://www.ntua.gr/parallel/faqs/PVM-vs-Linda>
I found the following helpful for understanding (a little anyway)
TCGMSG, P4, PVM (though based on version 2), C-Linda, and POSYBL:
	ftp://casper.cs.yale.edu/pub/tr975.ps

> PVM seems more mature than MPI,
> but MPI has formed standard. 

MPI is a standard, it seems like mpich is the de facto implementation.
Someone feel free to correct me.

> I wonder which one has a brighter future.
> Which one will become more popular in the near future? Please give
> me some suggestions, which will help me decide to focus on MPI or PVM.
> Any suggestion will be appreciated, thanks in advance.

I am in a similar situation.  However, the fact that both MPI and PVM
have usenet groups that have non-zero traffic is reassuring with
either decision, and the best that I can determine they each have a
reasonable user-base and won't be going away soon.

[ Aside: Added to the confusion from the above, my superiors need to
know if we should be using threads (POSIX or Solaris?) or not, and
whether or not we should just build our own parallel API using sockets
(UDP or TCP?  Blocking, non-blocking, I/O multiplexing, or threads
again?)  Finally, they have been bitten by the OO bug, and want
everything to be C++.  Life was so much easier with fewer choices. ]

-- 
Clark