MM5

The Distributed Memory (DM) Parallel Option


Introduction


Beginning with Version 2.8, in March, 1998 and continuing through the current Version 3, MM5 has been extended for use on distributed memory (DM) parallel computers such as the IBM SP, Cray T3E, Fujitsu VPP, Beowulf clusters of PCs and workstations, and distributed memory clusters of multi-processor machines. It also provides an alternative to shared-memory parallel execution on  distributed shared memory (DSM) machines such as as the Silicon Graphics Origin 2000, the Hewlett Packard SPP, and others.   Increased memory and performance from scalable parallel machines will support higher resolution studies over larger domains.  The option is employed only when the model is compiled with the command `make mpp'; non-distributed memory parallel users (C90, J90, workstation, etc.) will continue to compile with `make' alone, and will notice no difference in usage, results, or performance.

The distributed memory option to MM5 is implemented using MPI message-passing provided by RSL, a parallel runtime system library.  The changes to MM5 code for distributed memory parallelism are automated and to a large extent hidden from the user, using  FLIC, a Fortran translator/preprocessor that is applied to the model when it is compiled. Papers and other information describing the "same-source" approach to parallelizing MM5 are available.

The distributed-memory parallel option complements the OpenMP shared-memory parallel directives in the MM5 allowing mixed shared-/distributed-memory parallel execution on distributed memory clusters of multi-processor shared-memory nodes.  Examples of this architecture are the newer generation IBM SP computers with 2-way and 4-way SMP nodes and Compaq SMP cluster systems.

Some 1999/2000 performance data for the distributed-memory option to MM5 running on a number of parallel platforms is available by clicking here.

Additional information is available on-line through the following links:


Please direct questions and/or feedback to mesouser@ucar.edu


The distributed-memory parallel option to MM5 was developed in collaboration with the Mathematics and Computer Science Division  of  Argonne National Laboratory, under funding by International Business Machines Corporation. It is supported as part of the official MM5 by the Mesoscale Prediction Group of MMM.

Credits:


Copyright © UCAR 1997 - Disclaimer - mesouser@ncar.ucar.edu

Last Modified: Feb 1, 1999