Bio-Inspired and Heuristic Methods Applied to a Benchmark of the Task Scheduling Problem

Abstract

Task scheduling seeks for the time-efficient allocation of the tasks of a parallel program to a multiprocessor system. Being intractable, heuristic methods have been developed to solve this problem. Among the more traditional approaches, approximate techniques as constructive list-based heuristics or simply random schedulers have been extensively employed. On the other hand, bio-inspired models, such as cellular automata (CA) and evolutionary-based schedulers, have been recently investigated as alternative approaches. However, the comparative analysis of the experimental results is primarily limited by the capacity of benchmarks to represent the problem in a full range of difficulty. Aiming to investigate the usage of a more comprehensive benchmark on comparative experiments, we have developed a set of scheduling instances based on real-world programs by applying variations in their features, including number of tasks, number of available processors and communication costs. We have applied two simple heuristics to serve both as baselines for performance and to evaluate the complexity of each problem instance as basis for comparison. Moreover, we investigate here three bio-inspired schedulers applied to the same instances. Two of them are genetic algorithm (GA) approaches while the third employs a GA to find good CA rules able to schedule unseen instances of a parallel program in a very fast operation. Our results show that the CA-based scheduler outperforms the other methods significantly on mosts instances while, on certain instances of the problem, a good solution can be produced consistently by a heuristic based on random allocations. We conclude that these instances are unfit for benchmark purposes and that there is a necessity of careful analysis and selection of problem instances for performance evaluation in this field of research.