Newsgroups: comp.parallel.mpi
From: maggie@cray.com (Margaret Cahir)
Subject: Re: MPI on cray J90
Organization: Cray Research, Inc.
Date: 24 Jun 1996 22:47:04 GMT
Message-ID: <4qn5t8$7r5@walter.cray.com>


Hello Frans,

I tried to send email to 'hoesel@chem.rug.nl' but it bounced, with a 
complaint that it couldn't find the host. 

But to respond:

> Hi,
> I did search the net to find any info on running MPI on a cray J90 without
> any result.
> 
> Someone claimed that there is a native version available for a little less
> than $5000. (still a *LOT* of money for a program that should probably
> have been part of the supplied opereating system)
> the questions are:
> 1) does it exists, and if so how good is it?

	I am not sure what you mean by "native".  You can download a public
	domain version of MPI and install it on the machine yourself.  Cray
	offers a software package called MPT (Message Passing Toolkit).
	MPT contains a number of things:  a version of PVM that
	uses shared memory (as in memory copies, not system V calls) and
	a network-based version of MPI.  This version of MPI has had some 
	tuning done to improve performance (increased buffer sizes, etc.).

	The next release of MPT will contain a shared-memory version of MPI,
	similar to what we have done for PVM.  What this means is that 
	multiple copies of the same executable may exist in a multi-tasked 
	group.   Communciation between members of a group are quite fast
	as they are based on a memory-copy.


> 2) is $5000 money well spend on this version?

	This pricing information sounds suspect.  For starters,  I don't 
	believe that MPI is sold as a standalone product.  I suggest you 
	go directly to whoever is responsible for getting the software 
	(either on your side or the Cray salesperson or support analyst).   

	As to whether it is worth whatever price it may be... well that
	depends.  If you don't do much message-passing or synchronization
	in the code (i.e. it's extremely coarse-grained) than an optimized
	version of MPI isn't going to help you much.  However, the more 
	message-passing, reductions, syncs, etc. that you do, the bigger the 
	payoff is.

	I have ported a number of applications to the C90/J90 using the
	PVM shared-memory and the improvements have all been significant.
	Preliminary timing results on the shared-memory MPI are very 
	encouraging (one fft code ran several times faster). 

> 3) where can I get more information

	There is a little information on the Cray web page (see
	http://www.cray.com/PUBLIC/product-info/sw/PE/ and look under
	Message Passing Toolkit.

	If you can give me an email address that works, I can send you some
	more information.

Maggie

> - frans
> 
> (I'm sorry to rudely intrude your newsgroup without first reading a lot of
>  messages, but decisions have to be made pretty soon)
> 
> 
>   _______________________________________________________________
> 
>     "Hi, my name is Frans, I'm your crazy scientist for today."
> 
>     Frans van Hoesel          hoesel@chem.rug.nl
>                               http://rugmd4.chem.rug.nl/hoesel


--------------------------------------------------------------------------
Margaret A. Cahir                       Scalable Applications & Algorithms 
maggie@cray.com                         Cray Research, Inc.
                                        655E Lone Oak Rd., Eagan, MN 55121
-- 

Margaret A. Cahir
Scalable Computing and Algorithms Group
Applications Division

