ECPs, transition metals, and parallel computing
Hi,
I hope the netters won't think I'm "hyping" GAMESS, but
Theresa's
e-mail has sparked me to inquire about something that has been percolating
in my mind since Doug Smith's e-mail a while back on parallel computing.
I vaguely recall last year we had some discussion on parallel
computational chemistry. If not, this could be the time. We have been lucky
to have access to the iPSC/860 at Oak Ridge through a collaboration between the
Computational Chemistry Group at MSU and the Joint Institute of Computational
Sciences located at U. T.-Knoxville. The speed of the machine is enough to
keep even impatient, untenured assistant professors from complaining.
We have looked at both transition metal and lanthanide catalyst systems
and compared identical jobs on the iPSC/860 versus the Cray Y-MP at San Diego
Supercomputer Center. I don't have an example handy which entails using ECPs
to calculate 2nd dervis numerically, but the rough conclusion is the same -
"4 to 8 nodes give Cray-like speed!"
The table below shows some very promising timings for two
sample calculations. One is a 44 basis set RHF calculations of
the nonlinear optical properties of water and the other is geometry
optimization of LuCl2Hs, a catalyst model, using effective core potential.
Timings in Seconds:
Water LuCl2H
iPSC/860 1 node 608.12
2 node 322.48
4 node 1777.33 179.21
8 node 958.37
112.76<<<<<<<<<
16 node 516.19 79.77
32 node 291.31 59.26
64 node 197.82 51.41
Cray Y-MP8/864
144.22<<<<<<<<
DECstation 3100 7743.02
For the two comparisons the times given are total CPU times. Of particular
interest to us, is the comparison between the Cray Y-MP8/864 at the San
Diego Supercomputer Center and the iPSC/860.
Several questions for discussion.
1) What other parallel "goodies" are out there? I thought I
read
somewhere that Gaussian has a parallel version out or soon to be realized.
Ditto for HONDO. What sorts of machines are these ported to? I can imagine
that a "hot" field like this would have new options almost daily.
2) Are there parallel programs that can do MP2 and/or MCSCF?
Perhaps other correlated wavefunctions? Are folks working on these?
3) Do any of these iPSC/860 "hypercubes" or related machines
come with
more than 8Mb/node? From my point of view, this would be the advance that
would make it feasible to look at realistic models of experimental systems.
Tom Cundari and Henry Kurtz
Computational Chemistry Group
Department of Chemistry
Memphis State University
Memphis, TN 38152