Newsgroups: comp.parallel.mpi
From: tqian@vt.edu (tqian)
Reply-To: Ted Qian
Subject: Re: PVMe/MPI comparison on SP2
Organization: Virginia Tech/Blacksburg Electronic Village
Date: 13 Apr 1996 23:20:33 GMT
Message-ID: <4kpcs1$pm9@solaris.cc.vt.edu>

In article <DpswIw.1H7@apollo.hp.com>, wjiang@apollo.hp.com (Weicheng Jiang) says:
>
>In article <4kn227$k0f@server.cs.vt.edu> heagy@cs.vt.edu (Winfield S. Heagy) writes:
>>Hi.  We recently completed an experiment to compare PVMe and MPI
>>on the SP2.  The performace of MPI was far superior to PVMe on
>>this platform.  These results confirm results we have seen in other
>>papers.  Can anyone explain why PVMe has overall slower message
>>passing times than MPI on the SP2?
>
>When you pack your messages into the PVM buffer, you incur an extra memory-
>to-memory copy and jeopardize the performance. To get the same performance
>as MPI, you should use pvm_psend/pvm_precv, which is supported in the ORNL/UT
>version of pvm3.3.10. Please refer to the following paper for a detailed
>discussion of PVM performance on the SP2:
>
>        http://www.netlib.org/utk/papers/pvmmpp.ps
>        
>Hope this information helps,
>
>Weicheng Jiang
>PVM Specialist, Hewlett Packard Co.
>
>
>
We also noticed PVMe on IBM SP2's performance worsen dramatically when a
non blocking pvm_nrecv is used by polling continuously. Does any PVM and/
or SP2 expert notice that phenomenon? 

