Newsgroups: comp.parallel.pvm
From: john@elroy.jpl.nasa.gov (John Wright)
Subject: Re: PVM on shared memory computers ?
Organization: Image Analysis Systems Group, JPL
Date: 22 May 1996 14:41:22 GMT
Message-ID: <4nv92i$ndc@elroy.jpl.nasa.gov>

In article <4ntu7d$o8g@fido.asd.sgi.com>,
Steve Simonds  <simonds@sgi.com> wrote:
>Bela Ban <bba@zurich.ibm.com> wrote:
>>Hi !
>>
>>I read in the PVM's Users's Guide that PVM is only available on
>>distributed-memory computers. Can someone tell me what would prevent PVM
>>from being run on shared-memory computers ? I even think there is the
>>advantage of sharing data structures rather than having to send data
>>(including marshalling / unmarshalling ) between processes !
>>
>>Is there a PVM implementation for shared-memory computers available ?
>
>PVM is implemented on the SGI Challenge/Power Challenge systems quite well.
>And the performance is outstanding I might add.
>
>Look for SGIMP64 and SGIMP
>

The last time I tried to use SGIMP my software kept pulling a bunch
of end of buffer, out of data errors.  I modified the pvmgetarch
command to return SGI5, even for our multiprocessor Onyx and the
system worked correctly every time.  I believe that the problem was
in the implementation of the message passing of shared memory and
using SGI5 forced the messages through the TCP/IP or whatever port.
I have seen others with this problem but no fix reported to date.
Does anyone know if this problem has been solved?  I would not
consider it solved if changes need to be made to the code depending
on whether it runs on a shared memory system or a distributed system.

John


-- 
john.r.wright@jpl.nasa.gov
http://www-dial.jpl.nasa.gov/~john

