Using extended neural network map tasks

Abstract

Software distributed shared memory (DSM) systems are being used as an alternative to parallel processing. To achieve high performance in terms of completion time for an application running on a DSM, an effective task mapping or scheduling method has to be employed. The method based on the characteristics of the application and the system configuration allocates tasks and data to machines to speed up execution. Due to the unique nature of the DSM in its communication, the task mapping methods developed for message-passing systems can not be directly applied on DSM systems. In this paper, a model describing the unique nature of DSM communication and the connection between tasks is put forward first. Based on the model, two task mapping methods are developed. One of the methods uses the Hopfield neural net, the other applies simulated annealing technique. They rely on evolutionary technique to search for optimum or near optimum solutions. The proposed methods have been implemented and tested under different configurations. Their performance has been evaluated by using representative applications, The results indicate the effectiveness of the methods and reveal the accuracy of the model.