MPMM: A Massively Parallel Mesoscale Model

Static domain decomposition is a technique that provides a quick path to porting atmospheric models on distributed memory parallel computers. However, parallel inefficiencies in the form of load imbalances and ill-tuned communication are difficult to correct without complicated and explicit recoding. Reconfiguring the code to run on larger or smaller numbers of processors may require recompiling. Modularity and machine independence may also suffer. If full advantage is to be taken of Massively Parallel Processing (MPP) technology, tools and techniques that allow for dynamic performance tuning and reconfiguration are required. Program Composition Notation (PCN) is a language and run-time system for expressing parallel programs developed at Argonne and at the California Institute of Technology. It provides an intermediate layer between the application program and the physical processors of a computer. It allows the model to be statically decomposed over a virtual machine, but this virtual machine can be mapped and remapped dynamically over the physical computer. Programs are portable to as many machines as PCN itself, modularity is easily preserved, and communication tuning for a particular computer is encapsulated within the PCN run-time system. In this paper we report on a project at Argonne National Laboratory to parallelize the Penn State/NCARmore » Mesoscale Model version 5 using a fine grain decomposition dynamically mapped and managed under PCN.« less