Year: 1988
Type: article
Abstract: Parallel computers seem to offer the greatest promise in realizing the orders of magnitude increase in computational power required by many scientific applications. One of the crucial problems to be addressed in this context is to reduce the overheads due to interprocessor communications/synchronization, and to provide load-balancing in the distribution of computational load.
The Process to Processor Mapping problem is that of assigning a set of parallel tasks to a set of processors in such a way that some performance criterion is optimized. Usually, speedup is used as the performance criterion to be maximized. This requires the allocation of the parallel components of the program (processes) among the processors in such a way that: (a) the total computational load is distributed as evenly as possible, and (b) the time required to perform necessary interprocess communication is minimized. Satisfying both of these objectives at the same time is a difficult problem, and thus, heuristics are needed to obtain satisfactory sub-optimal mappings.
In this dissertation, three distinct approaches to the task allocation problem are pursued. In the first of three, a heuristic two-step, graph-based mapping scheme (nearest-neighbor mapping) with polynomial-time complexity is developed to map finite element graphs onto hypercubes, and the effectiveness of the approach is measured both by estimation and experimental measurement on a 16 node hypercube parallel computer.
The second approach is based on a minicut graph partitioning heuristic, and two schemes, hierarchical and direct, are developed. The hierarchical approach uses two phases, cluster formation and processor allocation. The direct approach is found to be very effective when compared to the other mapping heuristics discussed in the dissertation.
Finally, the use of simulated annealing for task allocation is evaluated. A special problem is identified that arises in the context of constrained optimization in the form of local-optima traps to the standard simulated annealing method. An adaptive simulated annealing technique that uses a scaled cost function is proposed to overcome the problem and is shown to be effective with task interaction graphs of widely varying characteristics.
Author: Fikret Erçal
Institution: The Ohio State University
Language: English
Subfield: Hardware and Architecture
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W1560068970
