Autotasking on the CRAY Y-MP: a performance evaluation on seismic codes

Parallelism is now universally acccptcd as the only practical way of substantially increasing the performance of today’s high-speed processing systems. Nevertheless, programming parallel systems is s[ill a poorly understood art. With the availability of traditional vector processors such as Cray and Convex systems in multi-headed versions, toots for exploiting the available tightly coupled MIMD parallelism become very important. Frequently, the effectiveness of tools depends greatly on the application area; as a general rule, automatic parallelization tools may do well in certain situations but badly in others. This paper reports on a performance susdy of Cray’s Autotasking for the Y-M P/4 for certain seismic codes; specifically the algorithm involved was 3D DMO. The experimental set up was as follows: Code that previously had been cwefully hartdcoded for a four-processor X-MP was stripped of all psuallclization directives (micro and macro tasking) and the resulting uniprocessor code was submitted to Autotasking. The results arc encouraging in that the parallelism achieved by the autotasked programs compares very favorably with that of the originaf handcoded programs.