Implementing the Perfect ARC2D Benchmark on the BBN TC2000 Parallel Supercomputer

The TC.2000 is a MIMD parallel processor wi,th memory that is physically distributed memory, but logically shared. Interprocessor covnmunication, and therefore access to shared memory, is sufficiently fast that most applications can be ported to the TC.2000 without rewriting the code from scratch. This paper shows how this was done for the Perfect ARC'2D benchmark. The code was first restructured by changing the order of subroutine calls so that interprocessor communication would be reduced to the equivalent of three full transposes ofthe data per iteration. The parallel implementation was then completed by inserting shared data declarations and parallel extensions provided by the TC.2000 Fortran language. Thi:F approach was easier to implement than a domain decomposition technique, but requires more interprocessor communication. It is feasible only (because of the TC.2000'~ highspeed interprocessor communications network. References to shared memory take about 25% of the totai execution time for the parallel version of ARC2D. an acceptable amount considering the code did not have to be completely rewritten. High parallel efficiency was obtained using up