Task mapping on distributed shared memory systems using hopfield neural network

摘要

In order to reduce the execution time of a parallel program, the tasks/threads of the program have to be carefull mapped onto the processors of a system. Most mapping methods used on current Multithreaded Distributed Shared Memory (DSM) systems only consider the workload balance. Due to the ignorance of the communicatio between tasks/threads, these methods may lead to such mappings have excessive cross-proessor communication, which degrades the performance. In this paper, we present a static method to map user programs onto a multithreaded DSM system. In contrast with the prievious work, this method takes into account both load balance and communication. It applies the Hopfield neural network on the mapping problem of the multithreaded DSM system to find a near-optimum apping for a program. We have implemented this method on Cohesion which is a DSM system supporting multithreading. Two programs, Successive-Over-Relaxation (SOR) and Vector Quantization (VQ), are used to test the effectiveness of this mapping method. The result shows that our method indeed can find the optimal mapping for these programs.