A Technique for Multiprocessor Memory Resource Estimation

An approach for estimating the required memory resources to execute a program on a multiprocessor system is presented. The technique relies on the information contained in a data dependency graph representing a program to be executed. Data dependency graphs can be generated by a compiler as an intermediate representation of the program without accounting for any specific details of the target machine architecture. Therefore, this approach can be used in the early design phases of multiprocessor architectures for performance analysis of targeted or benchmark applications. Memory resource estimates are separated into two categories; algorithmic and run-time requirements. Experimental results of this technique are presented for several scientific benchmark code fragments.