Newsgroups: comp.parallel.mpi,comp.parallel.pvm
From: tony@aurora.cs.msstate.edu (Tony Skjellum)
Subject: Re: MPICH on T3D Repost
Organization: Mississippi State University
Date: 24 Apr 1995 22:26:06 GMT
Message-ID: <3nh8hu$3pb@NNTP.MsState.Edu>

Dear Cray T3D MPI Users (or potential):

Follow up message.  Bug reports and comments on T3D MPICH to:
     crayt3d-mpi@erc.msstate.edu
This will be superceded by standard MPICH mechanism when we
reintegrate for 1.0.9 version.

Here is some initial performance info.  Clearly, we need to
do more work on our short message latency, though the fit
shown below does not definitively measure that.  As we said in
previous posting, this is the *first* release (should perhaps be
*zeroth*) and we expect to make improvements.  Our technology
is put-based, and that should be a big help in performance
for the machine.    Showing you this run is meant, in part,
to demonstrate that the MPICH code runs, and that it at
least gets high bandwidth.

The following run was done by Ron Brightwell.  The entire T3D port was
a term project for my graduate Systems Programming class at MSU.  

We will be supporting this port within the MPICH framework
very soon, and a T3D-specific publication will
follow that explains the architecture and optimizations we have
planned.

All of this early release effort is aimed at ending the long
wait of T3D users to get MPI.  Please send your comments to
the above alias or to me.

MPICH is a joint project between Argonne National Laboratory
(Rusty Lusk, Bill Gropp), and MSU (Tony Skjellum, Nathan Doss)
and my research group (including Ron Brightwell).

Regards,
Tony Skjellum

cf,


T3D:-) ./mpptest -npes 2 -size 1024 20480 1024
set default
set font variable
set curve window y 0.15 0.90
set order d d d x y d
title left 'time (us)', bottom 'Size (bytes)',
      top 'Comm Perf for MPI (h4p)',
 'type = blocking'

#p0     p1      dist    len     ave time (us)   rate
0       1       1       1024    329.443884      3108268.36
0       1       1       2048    108.538056      18868957.82
0       1       1       3072    116.943552      26269084.08
0       1       1       4096    123.622686      33133077.21
0       1       1       5120    134.356266      38107638.39
0       1       1       6144    158.230710      38829377.69
0       1       1       7168    168.864696      42448185.85
0       1       1       8192    176.486310      46417198.03
0       1       1       9216    182.893194      50390065.36
0       1       1       10240   191.864970      53370868.06
0       1       1       11264   208.209276      54099414.86
0       1       1       12288   217.772874      56425760.35
0       1       1       13312   222.848802      59735569.05
0       1       1       14336   239.621778      59827617.17
0       1       1       15360   246.635928      62278031.12
0       1       1       16384   254.153790      64464905.28
0       1       1       17408   260.348418      66864243.44
0       1       1       18432   271.590066      67866988.92
0       1       1       19456   285.982884      68032043.48
0       1       1       20480   294.947136      69436171.78
# Model complexity is (1.353011e-04 + n * 6.916538e-09)
# startup = 135.30 usec and transfer rate = 144.58 Mbytes/sec
plot square
join
set function x 1024 20480 '135.301149+0.006917*x'
join dots
wait
new page
-----------------------------------------------------------



