Energy Bounded Scalability Analysis of Parallel Algorithms
Vijay Anand Korthikanti, Gul Agha · Illinois Digital Environment for Access to Learning and Scholarship (University of Illinois at Urbana-Champaign) · 2009
The amount of energy available in some contexts is strictly limited. For example, in mobile computing, available energy is constrained by battery capacity. As multicore processors with a large number of processors, it will be possible to significantly vary the number and frequency of cores used in order to manage the performance and energy consumption of an algorithm. We develop a method to analyze the scalability of an algorithm given an energy budget. The resulting energy-bounded scalability analysis can be used to optimize performance of a parallel algorithm executed on a scalable multicore architecture given an energy budget. We illustrate our methodology by analyzing the behavior of four parallel algorithms on scalable multicore architectures: namely, parallel addition, two versions of parallel quicksort, and a parallel version of Prim’s Minimum Spanning Tree algorithm. We study the sensitivity of energy-bounded scalability to changes in parameters such as the ratio of the energy required for a computational operation versus the energy required for communicating a unit message. Our results shows that changing the number and frequency of cores used in a multicore architecture could significantly improve performance under fixed energy budgets. I.