Heterogeneity as Key Feature of High Performance Computing: the PQE1 Prototype
An operational forecast system is being developed at ENEA, which consists of a high resolution meteorological limited area model coupled with ocean models for the prediction of the state of the Mediterranean Sea and of high water events in the Venice Lagoon. In order to satisfy the operational requirement (3-days forecasts in a few hours) a very e ective two level parallelization strategy was adopted: the parallel implementation of the most time-consuming codes and the simultaneous execution of di erent models. Following this approach, the integrated system has been implemented on the hybrid machine PQE1, which integrates the Meiko CS-2 MIMD architecture and the APE100/Quadrics SIMD architectures by means of HiPPI channels. The rst level of parallelism allowed to obtain a highly performing meteorological code QBOLAM, running at more than 2 GFlops sustained on a 128 nodes Quadrics machine. The second level of parallelism allowed to gain more than a factor 2 with respect to the sequential execution of the di erent processes.