Argonne Chemistry Research Areas

Chemical Dynamics in the Gas Phase

Parallel MRSDCI Code Developments

In the past three years, the parallel version of the CI
diagonalization code has been further developed and ported
to several machines, including the 512-node Intel Touchstone
Delta, the 128-node IBM SP1 at ANL, and the group's
10-node IBM SP2. This parallel CI code is based on a
partitioning of the Hamiltonian matrix, trial vector, and
matrix-vector product and uses a distributed-memory, single
program multiple data (SPMD) programming model based on
the TCGMSG library and the global-array library developed
by R. J. Harrison (PNL). This approach leads to a portable
implementation that runs efficiently on both shared-memory
and distributed-memory computers, including heterogeneous
networks of workstations. In the past year, a new interrupt-
driven version of the global-array library has been developed
and incorporated into the IBM SP1/SP2 version of the code.
This required modification by the vendor (IBM) of the AIX
operating system and TB2 switch software. It is rare that a
chemistry group at a national laboratory has the opportunity
to impact so directly a commercial product. Within a single
hamiltonian matrix-vector product, each CPU is responsible
for only a subset of the possible combinations of segment pairs.
In collaboration with Hans Lischka and R. J. Harrison, the
bottlenecks associated with load-balancing, replicated data
distribution, and with shared file access have been largely
eliminated, and calculations have been demonstrated to scale
to over 300 nodes on the Intel Delta. This is the first successful
attempt to parallelize a production-level MRSDCI code, and
this effort represents a major step forward toward using
effectively the large-scale parallel supercomputers that are
becoming available to scientists. Generalizations of the method
are planned that will allow treatment of larger molecular systems.
Future effort will be directed also to integrate the parallel version
of the code with other parts of the COLUMBUS Program
System to allow production-level PES calculations. Ports to
other massively parallel, distributed memory machines, such as
the Cray T3D and Intel Paragon, are also underway. Figure 1
shows preliminary timings of the code on the IBM SP1, the
Cray T3D, and the Intel Paragon.

Figure 1. Performance of parallel MRSDCI code
applied to butadiene on three different massively
parallel computers.

Contacts

For more information, contact Ron Shepard

Return to Chemical Dynamics in the Gas Phase

Glassblowing

Interfacial Processes

Radiation and Photochemistry

Photosynthesis
Biological Materials Growth Facility

Atomic Physics

Coordination Chemistry

f-Electron Interactions

Actinide Facility

Computational Materials and Electrochemical Processes