Optimal Task Scheduling for Parallel Systems Using State-space Search

Download or read book Optimal Task Scheduling for Parallel Systems Using State-space Search written by Michael Orr and published by . This book was released on 2020 with total page 161 pages. Available in PDF, EPUB and Kindle. Book excerpt: It is of ever-increasing importance that programs are able to take full advantage of the parallel systems on which they are run. Task scheduling is the problem of producing a schedule for a program, such that the tasks which make up the program are each allocated to a specific processor and in a specific order which minimises the overall run-time. This problem is NP-hard, so that the amount of work required grows exponentially as the number of tasks is increased. Although the NP-hardness of the problem usually discourages optimal solving, an optimal schedule can give a significant advantage in time critical systems or applications where a single schedule is reused many times. Previous research with branch-and-bound for optimal task scheduling has shown promise with small task graphs, being competitive with other methods. The state-space model used in that work has an obvious drawback of allowing many duplicate states to occur in the state-space, which theoretically causes a large amount of additional time and memory to be required. This thesis proposes a new state-space model called Allocation-Ordering (AO), which improves on older models through its carefully designed lack of duplicate states. AO divides the task scheduling problem into two distinct sub-problems (allocation and ordering) which are handled in sequence within the state-space. Experimental evaluation confirms the benefits of the model. The benefits of AO’s lack of duplicate states for other branch and bound algorithms are then explored, specifically variants with interesting properties such as parallelisation and low memory requirements. We then investigate its applicability to more complex task scheduling models: the model is first adapted to allow optimal task scheduling with related heterogeneous processors, and then to allow optimal task scheduling with task duplication. The success of the adaptation of AO shows its flexibility, and suggests it may have wide applicability to variants of the task scheduling problem, and potentially other problems.