Design And Implementation Of Scheduling Algorithm Of Heterogeneous Cluster Based On Torque

With the booming of the high performance computing industry, more and moreresearchers and business companies are investing much manpower and resources inthis field. As a main kind of high performance computing, cluster occupies a more andmore important position because of its high expandability and well performance. Withthe high development of GPU, it is becoming one of the main components of cluster.Torque is a very common open source cluster management system and has beenapplied high performance computing environment of many colleges and researchinstitutes. Torque is very powerful in managing resource which is capable ofsupporting over1500nodes and updated for all versions of UNIX/Linux operatingsystem. However, in terms of scheduling strategy, Torque functions much weakerbecause it adopts simple FirstFit scheduling policy without considering cluster loadbalance and cannot schedule GPU jobs. This article has proposed an extension Torquescheduling strategy, which is a dynamic load balance scheduling algorithm based onself-learning for heterogeneous cluster. Then, two experiments relating to GPU areproposed, which are convolution algorithm and X.264, to testify the extension Torquescheduling strategies. The results show that the GPU jobs are well scheduled on therefined Torque platform and the performance has been improved greatly.