Adaptive online scheduling of tasks with anytime property on heterogeneous resources

TL;DR

This paper presents an adaptive online scheduling algorithm for tasks with anytime properties on heterogeneous resources, balancing response time and solution quality through dynamic quality adjustments and estimations.

Contribution

It introduces a novel scheduling approach combining traditional methods with quality control heuristics and a procedure for estimating processing-time-to-quality relationships.

Findings

Effective handling of server overload in client-server applications.

Improved scheduling performance demonstrated on personnel rostering problem.

Two heuristics for setting task quality levels show promising results.

Abstract

An acceptable response time of a server is an important aspect in many client-server applications; this is evident in situations in which the server is overloaded by many computationally intensive requests. In this work, we consider that the requests, or in this case tasks, generated by the clients are instances of optimization problems solved by anytime algorithms, i.e. the quality of the solution increases with the processing time of a task. These tasks are submitted to the server which schedules them to the available computational resources where the tasks are processed. To tackle the overload problem, we propose a scheduling algorithm which combines traditional scheduling approaches with a quality control heuristic which adjusts the requested quality of the solutions and thus changes the processing time of the tasks. Two efficient quality control heuristics are introduced: the first heuristic sets a global quality for all tasks, whereas the second heuristic sets the quality for each task independently. Moreover, in practice, the relationship between the processing time and the quality is not known a priori. Because it is crucial for scheduling algorithms to know at least the estimation of these relationships, we propose a general procedure for estimating these relationships using information obtained from the already executed tasks. Finally, the performance of the proposed scheduling algorithm is demonstrated on a real-world problem from the domain of personnel rostering with very good results.