Predictive Scheduling for Distributed Computing

Scheduling parallel jobs across distributed, possibly heterogeneous, computing resources is an increasingly important capability because of the increase of available distributed resources and of parallel programs and algorithms. A wide variety of systems, methods, and strategies have been proposed and tested to address the scheduling problem. These systems typically lack self--adaptation to new architectures and algorithms, guarantees of services, and/or any degree of proactivity in dealing with environmental changes. This pa- per outlines a strategy for predictive scheduling, in which the parallel scheduler combines sev- eral techniques in order to schedule and manage tasks intelligently, adaptively, and proactive- ly and in order to address performance--related issues before they become significant bottlenecks. Implementations of several key components are presented, along with some test results, and an architecture for a complete predictive scheduling system is outlined. Key Wo..