A programming environment for dynamic resource allocation and data distribution

Abstract

Dynamic resource allocation, particularly in the form of processorpartitioning and scheduling, is an important factor in achieving good utilization and high performance in multiprocessor systems. We are developing a Distributed Resource Management System (DRMS) for dynamic allocation of resources during the execution of applications. DRMS provides means for applications to specify their resource requirements and for the system to manipulate resources allocated to competing jobs. DRMS also provides the programming support necessary to create reconfigurable applications, that can execute on time-variant processor partitions. Run-time program migration and data redistribution are performed automatically. We present performance results using application benchmarks on the first implementation of DRMS for the IBM SP2. Our performance results show that (i) the support for dynamic resource control has minimal side effects on the application performance, and (ii) the costs associated with the data redistributions and program restructuring to support changes in the processor partitions are relatively small.