SC13 Denver, CO

The International Conference for High Performance Computing, Networking, Storage and Analysis

Power and Performance Modeling for High Performance Computing Algorithms.


Student: Jee Choi (Georgia Institute of Technology)
Supervisor:

Abstract: The overarching goal of this research is to enable algorithm designers to create more energy efficient algorithms by providing the means of analyzing the relationship between time, energy and power on real systems. Firstly, we provide a simple analytical cost model for energy and power. This model expresses energy and power for algorithms using a small set of simple parameters (e.g., memory bandwidth, number of FLOPs). We validate our model using highly optimized microbenchmarks and a fine-grained power measurement tool on state-of-the-art systems. Secondly, we iteratively refined our model and extend its validation on over a dozen "candidate compute-node building blocks" for future high performance systems, including the latest x86, GPUs, and mobile SoCs. The purpose of this study is to provide a precise analytical characterization of abstract algorithmic regimes where one building block may be preferable to others. Together, they provide the means to re-design algorithms for energy efficiency.

Poster: pdf
Two-page extended abstract: pdf


Poster Index