Abstract
An essential component of effective use of distributed systems is proper task placement, or scheduling. To produce high-quality schedules, scheduling algorithms require underlying support mechanisms that provide information describing the distributed system. The work presented here makes a clear distinction between scheduling policies and the underlying mechanism, and focuses on the problem of providing general purpose mechanisms that facilitate a broad spectrum of task placement algorithms.
This dissertation proposes a model for distributed scheduling support mechanisms.
This model includes scalable and extensible mechanisms that support the efficient implementation of scheduling policies on distributed systems, while preserving the autonomy of the component systems. The mechanisms include provably correct information
exchange protocols for system state dissemination in distributed systems.
MESSIAHS is a prototype implementation of these mechanisms, including a scheduling module that implements the basic mechanism, as well as a library of function calls and a specialized programming language for writing distributed schedulers. As a demonstration of the utility of the prototype, several algorithms from the literature are implemented and their performance is analyzed. The experimental results show average overhead of approximately 10% using MESSIAHS, measured against a theoretical ideal running time. The results indicate that it is possible to build scalable, general-purpose mechanisms that support a variety of task placement algorithms while preserving autonomy.
