Work

A fast, effective scheduling framework for parallel computing systems

Year: 1990
Type: article
Abstract: We would like fast, effective schedulers to be available for all classes of concurrent applications and parallel computer architectures. Task assignment and scheduling for arbitrarily large multicomputers cannot be performed efficiently with a single, sequential scheduler. Also, dynamic applications for parallel systems cannot be scheduled optimally under reasonable assumptions. Therefore, efficient schedulers for such applications must be distributed and use heuristics. Building schedulers for parallel systems involves addressing two problems simultaneously--(1) developing efficient and effective heuristics while (2) allowing for various scheduler to machine mappings. We present a framework for the construction of highly flexible parallel, real-time schedulers. The framework's flexibility is demonstrated by mapping it to both a multiprocessor and a multi-computer architecture. Flexibility is further displayed by using different message passing and process migration facilities with each architecture. We also display flexibility by building scheduling systems for applications with two vastly different scheduling objectives: (1) upholding the deadlines of the application's tasks and (2) minimizing the completion time of the entire application. Finally, flexibility is demonstrated by adding scheduling constraints (i.e., scheduling resources other than CPU time, precedence constraints, group scheduling, task criticalities and hardware constraints) to the basic schedulers for both classes of applications. These facts verify that our framework encapsulates the essential features for developing parallel schedulers. Furthermore, implementations of on-line schedulers are shown efficient and effective by experimentation with workload generators and sample real-time and non-real-time applications. First, limitations are shown to exist regarding the benefits of dynamic scheduling due to the cost of task migration and scheduling overhead. Next, the implemented schedulers' performances are measured. Efficiency is measured as both latency (i.e., the time between when a task seeks scheduling and when the task is scheduled) and overhead (i.e., the processing time used to make scheduling decisions). Measurements indicate our implemented schedulers are highly efficient and create little scheduling overhead. The effectiveness of each scheduler is displayed by favorably comparing its computed schedules to optimal schedules. All results indicate that efficient, effective heuristic schedulers exist for scheduling dynamic applications.
Author: Ben Blake
Language: English
Open Access status: green
OpenAlex ID: W100566344
FWCI: 0.514
Cites: 44
Cited by: 3
Related to: 10