We present the parallelization of an elastodynamic code on a firmly coupled configuration consisting of two IBM 3090-600 VF, a total of 12 processors, joined with a connection facility. The programming environment used is Clustered FORTRAN which is a facility for writing and executing parallel programs on two coupled IBM 3090 vector multiprocessors (VMP). Clustered FORTRAN provides extensions to FORTRAN so that a single application program can execute across multiple 3090 systems as well as across the processors of a single 3090 system. To exploit parallelism on each VMP, we used the parallel DO LOOP construct and to exploit parallelism across VMP we used the SCHEDULE statement allowing concurrent execution of subprograms. It is shown that such a problem, characterized by sufficiently large amounts of arithmetic operations within the parallel tasks, can be efficiently parallelized obtaining average speed-ups of more than five on one 3090-600 VF and nearly ten on a two 3090-600 VF cluster with a small programming effort.

A two-level parallel implementation of a large-scale geophysical simulation code is presented. The software/hardware environment utilizes IBM Clustered Fortran, allowing a single application program to execute concurrently on two IBM 3090 computers (first level of parallelism) while exploiting the multiple vector processors of each 3090 system (second level of parallelism). The experiments reported show that the problem is characterized by large task granularity and small communication/computation ratio, thus leading to sustained parallel speed-ups of nearly ten on a cluster of two 3090 computers, totalling twelve vector processors.>

An instruction-level simulator for IBM 3090 with VF (vector facility) has been developed for studying the performance of vector processors and their memory hierarchies. Results of a study of the locality of several large scientific applications are presented. The cache miss ratios of vectorized applications are found to be almost equal to those of their original scalar executions. Moreover, both the spatial and temporal locality of these applications (in scalar and vector executions) are strong enough to show a sufficiently high hit ratio on conventional cache structures.>