Performance of work conserving schedulers and scheduling of some synchronous dataflow graphs

Abstract

We know a lot about competitive or approximation ratios of scheduling algorithms. This, though, cannot be translated into direct bounds on the schedule produced by a scheduling algorithm, because often the optimal solution is intractable. We derive a methodology to find absolute bounds on the scheduling of jobs with precedence constraints on parallel identical machines. Our bounds hold for a large class of online and offline scheduling algorithms: the "work conserving" scheduling algorithms. We apply this methodology to prove that an important class of synchronous dataflow graphs -the parallelized pipelines -has very good performance characteristics when scheduled by a work conserving scheduler. Real time guarantees and granularity design for these dataflow graphs are discussed. We argue that parallelized pipelines should be dynamically scheduled on multiprocessor architectures.Keywords: Parallel job scheduling, online scheduling algorithms, synchronous dataflow graphs, real time systems, performance analysis of parallel systems. Performance of Work Conserving Schedulers and Scheduling of Some Synchronous Dataflow GraphsUdayan Kanade Codito Technologies Pvt. Ltd.udayan@codito.com AbstractWe know a lot about competitive or approximation ratios of scheduling algorithms. This, though, cannot be translated into direct bounds on the schedule produced by a scheduling algorithm, because often the optimal solution is intractable. We derive a methodology to find absolute bounds on the scheduling of jobs with precedence constraints on parallel identical machines. Our bounds hold for a large class of online and offline scheduling algorithms: the "work conserving" scheduling algorithms. We apply this methodology to prove that an important class of synchronous dataflow graphs -the parallelized pipelineshas very good performance characteristics when scheduled by a work conserving scheduler. Real time guarantees and granularity design for these dataflow graphs are discussed. We argue that parallelized pipelines should be dynamically scheduled on multiprocessor architectures.