Resource management in message passing environments

This paper discusses the need for resource management support for parallel applications running on workstation clusters and communicating by message passing among tasks. Many resource management systems are only able to start a message passing runtime environment and parallel applications, but dynamic reconfiguration fails because of the missing cooperation between the resource manager and the runtime environment. In order to utilize computational resources in message passing environments efficiently, to control execution of parallel applications by rescheduling tasks at runtime, and to minimize their execution time, a resource management system has been developed and preliminary tests results have been carried out. Most of our efforts in this regard have been to design an efficient approach to load measurement and process scheduling and implement the resource management system in a manner such that it can easily be adapted to any message passing framework. Although our first version is based on the PVM system, we also intend to implement an MPI - based resource management system.