Shopping for job takers [distributed scheduler agent

Abstract

Workstations interconnected by a high-speed network can extend the processing power available to an individual far beyond a single workstation. However, a distributed scheduler agent is required to exploit these computational resources. The distributed scheduler provides task placement advice to distributed and parallel applications. This enables the job at hand to be dispatched to a remote host for execution. We describe three versions of a distributed scheduler agent based on different scheduling algorithms and compare their performance. The three algorithms are based on the communication architecture, the transfer policy, information policy and the placement policy adopted. The transfer policy decides when an attempt should be made to transfer an arriving job to another host. The information policy decides when information about the states of other hosts in the system should be collected, where it should be collected from, and what information should be collected. The placement policy determines the host to which a job should be transferred for execution. These characteristics thus serve as a framework for comparing these scheduling algorithms.