CCL: computer for high-level QC calculations
- From: "Jeff Hammond"
- Subject: CCL: computer for high-level QC calculations
- Date: Wed, 4 Jun 2008 15:28:51 -0400
Sent to CCL by: "Jeff Hammond" [jeff.science-*-gmail.com]
An SGI Altix 4700 is a great machine to run NWChem on, as the code runs
extremely well on Itanium processors and will utilize every last drop of shared
memory if you ask it to.
GAMESS, PQS and Aces II (MAB) utilize disk for large CC calculations and were
designed to run on machines with more modest interconnects. That is not to say
that they won't benefit from a higher network bandwidth and lower latency, but I
don't think you'll get your money's worth. For these codes, you're probably
better off buying a machine like the ones the developers demonstrate those codes
on, which is a Beowulf cluster with striped SCSI disk arrays and channel-bonded
Aces 3 is quite new, but should run well on an SGI Altix, assuming it is
supported. The Aces 3 design paradigm permits platform-specific optimization
which might be able to exploit the special features of the Altix.
You can refer to the following articles which discuss the relative performance
of various codes in parallel. Each paper notes the hardware that the code was
tested and timed on.
Aces II (MAB) - J. Chem. Theory Comput., 4 (1), 64 -74, 2008 -
GAMESS - J. Chem. Theory Comput., 3 (4), 1312 -1328, 2007 -
PQS - J. Chem. Theory Comput., 3 (4), 1368 -1377, 2007 -
Aces 3 (QTP) J. Chem. Phys. 128, 194104 (2008) - http://link.aip.org/link/?JCPSA6/128/194104/1
NWChem (old CC code) J. Phys. Chem. A, 109 (31), 6934 -6938, 2005 -
NWChem (TCE) J. Chem. Phys. 127, 144105 (2007) - http://link.aip.org/link/?JCP/127/144105
(The timing data in this paper is already obsolete, but the scalability is
I am not aware of any reason why Itanium would be an issue. The Intel compilers
do a fine job of generating fast code and this chip is the best available for
DGEMM, which is the fundamental kernel for coupled-cluster codes.