Processor allocation for parallel applications in computing cluster

Abstract

We study the problem of processor allocation in multi-computer for parallel processes that do I/O and IPC (Inter Process Communication). A simulator has been modified to accept synthetic parallel workload that does I/O and IPC. We study the effect of I/O and IPC on performance of parallel applications using the simulator. Also, we study a novel design approach for distributed operating system Kerrighed, which uses kernel level distributed shared memory to provide global resource management of CPU, memory and disks and efficient process migration mechanism. As kernel level distributed shared memory provides unified basis for doing IPC we study the special case of processes, which communicate using underlying DSM. We propose to use state of the DSM pages in scheduling decisions and utilize efficient process migration mechanism to reduce consistency related communication in DSM and thereby improve the response time of parallel processes.