Newsgroups: comp.parallel.mpi,comp.parallel.pvm
From: kohr@donner.mcs.anl.gov (Dave Kohr)
Subject: Re: Memory Channel: Message Passing Latency 5.8 usecs, ~60 MB/sec @ 4KB messages
Organization: MCS, Argonne National Laboratory
Date: Wed, 14 Aug 1996 4:58:3 GMT
Message-ID: <8399986834943@donner.mcs.anl.gov>

In article <4urfml$cnr@bach.convex.com>, Paco Romero <romero@convex.com>
wrote:
>In article <01bb8966$3b025a20$c4400410@richard.pa.dec.com>,
>Richard Kaufmann <richard@ilo.dec.com> wrote:
>>The bottom line is that Memory Channel has achieved 5.8 usecs
>>one-way latency for HPF, 6.9 usecs for MPI and 8.0 usecs for PVM.  Achieved
>>"ping pong" bandwidth is >60 MB/sec for both HPF and MPI.
>
>The HP PVM and MPI implementations on the HP/Convex Exemplar SPP1200 achieved
>a 5 usec latency back in 1995.  With additional tuning we  achieve an 
>impressive 3.99 usec latency and 135 MB/s bandwidth on the SPP1600

What are the minimum latencies and peak theoretical bandwidths of the two
interconnects?

Dave Kohr    Hacker/Researcher    Argonne National Laboratory, MCS Division
Building 203, Room C-246   Phone: (708) 252-4243   E-mail: kohr@mcs.anl.gov
See also my WWW Home Page:      http://www.mcs.anl.gov/home/kohr/index.html
                      "Surfing the Silicon Prairie."

