email: nmis@di.uoa.gr phone: (+301) 727 5103, fax: (+301) 723 1569
Abstract
In this paper we present a parallel implementation of the numerical weather prediction Eta model on distributed memory MIMD processors. The required interprocessor communication was implemented and validated on the message passing library Message Passing Interface (MPI). Domain decomposition techniques are used for the horizontal discretization, whereas in the vertical direction each column per grid point is computed.
Four basic design principles were followed:
- Retain the original sequential code, as much as possible, for a number of reasons: the parallel code is recognisable by the sequential code developers; a smaller development effort is required as modifications are kept to a minimum; the parallel code performance may be directly compared with the sequential code. We used the Single Program Multiple Data (SPMD) and the message-passing paradigms
- Minimize communication between processes
- Develop portable parallel code for maintaining a single code for any execution environment.
- Develop scalable code for the speed-up to be proportional to the number of available processors.
It is found that for a mesh network topology the communication to computation ratio is minimized if the computation domain of each processor is a square of side sqrt(N/p) for a sqrt(N) x sqrt(N) mesh of grid points, where p is the number of processors. Moreover, we show that the number of grid points per processor plays an important role in the communication complexity and when ts < tw sqrt(N/p), where ts is the start-up time and tw is the per word transfer time, then the communication to computation ratio is improved by an order of magnitude. All data communication for finite differencing and I/O is encapsulated within a parallel library. Finally, we present performance results on the Convex Exemplar SPP-2000.