Newsgroups: comp.parallel.pvm,comp.parallel.mpi
From: ZHGFJ <Serpent_Mage@MSN.COM>
Subject: Re: ANNOUCE: Para++ release, C++ bindings for PVM and MPI.
Organization: UBS ORKI OERU
Date: 6 Jul 1995 06:23:00 GMT
Mime-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Transfer-Encoding: 7bit
Message-ID: <3tfvg4$1df@ns2.ny.ubs.com>

>MPICH has a chameleon device driver for some systems; other
>implementations have device drivers that directly work to the
>low-level system calls (eg, T3D, SGI).  MPI is not
>inherently inefficient, as your message seems to imply.

>One should be careful to associate "heavy weight" with all abstractions.

Of course every layer that is added on causes extra time.  I have worked
at two levels high and low.

1)  Created my own message passing scheme (transputers)
2)  Used MPI.
3)  Used Microsoft code.

Guess what the standings was 1,3,2 with 1 being the fastest.  MPI is a
good idea because it does standardise the function calls.  I ended up 
writing my own MPI implementation because it turned out to be faster.

But I am still very in much in favour of MPI.  I think what could be a
really good idea is to standardise on the packet information.  This would
achieve two things.  
1)  Various MPI implementation can talk to each other.
2)  People can still if they wanted to write their own implementation.

Christian Gross
CHG Informatik


