Get all the updates for this publication
This work presents visualization of performance analysis in divisible computation. Multi-core processor units are used as a hardware platform. To assess the performance of parallel algorithms, we commonly use runtime as a common metric. Metrics such as asymptotic analysis, speedup, reliability, efficiency, total cost and scalability are also considered for improved runtime. As the data size scales up exponentially energy-efficiency and power-efficiency is gradually becoming equally important as that of performance. Algorithm with improved performance results in preserving of energy as well. The structure of the parallel algorithm precisely determines the trade-off between performance and energy consumption. In addition we perceive major energy savings may be achieved, when task granularity is properly selected. Granularity has great impact on performance in parallel algorithm. A small task provides high speed-up and coarse grain results in load imbalance; hence optimal performance can be achieved between fine grain and coarse grain. In extorting the relation between metrics we propose a method of vitalizing these interrelations as two-dimensional maps. It is a line that joins points on the map with same or equal energy at a particular point of time. As an instance for the approach we present map for Amdahl's law of parallel computation. Map for this model will furnish relative and comparative analysis into the insights of performance. © 2018 American Scientific Publishers. All rights reserved.
Journal | Data powered by TypesetJournal of Computational and Theoretical Nanoscience |
---|---|
Publisher | Data powered by TypesetAmerican Scientific Publishers |
ISSN | 1546-1955 |
Open Access | 0 |