On finding optimal clusterings of task graphs

Abstract

Currently, many parallel algorithms are defined for shared memory architectures. The preferred machine model is the PRAM, but this model does not take into account properties of existing architectures that have a distributed memory and an asynchronous execution model. A transformation of PRAM programs into distributed, asynchronous ones is known. In order to produce not only correct but also efficient code the tasks have to be clustered. We introduce a parallel algorithm producing an optimal clustering for coarse grained task graphs with respect to the execution time on an asynchronous distributed random access machine, the A-DRAM. This machine model assumes distributed memory, asynchronous execution of tasks, computation costs, and communication delay.<>