The International Conference for High Performance Computing, Networking, Storage and Analysis
Machine Learning Algorithms for the Performance and Energy-Aware Characterization of Linear Algebra Kernels on Multithreaded Architectures.
Authors: A. Cristiano I. Malossi (IBM Corporation), Yves Ineichen (IBM Corporation), Costas Bekas (IBM Corporation), Alessandro Curioni (IBM Corporation), Enrique S. Quintana-Ortí (James I University)
Abstract: The performance and energy optimization of the 13 ``dwarfs'', proposed by UC-Berkeley in 2006, can have a tremendous impact on a vast number of scientific applications and existing computational libraries. To achieve this goal, scientists and software engineers need tools for
analyzing and modeling the performance-power-energy interactions of their kernels on real HPC systems.
In this poster we present systematic methods to derive reliable time-power-energy models for dense and sparse linear algebra operations. Our strategy is based on decomposing the kernels into sub-components (e.g., arithmetics and memory accesses) and identifying the critical features
that drive their performance, power, and energy consumption. The proposed techniques provide tools for analyzing and reengineering algorithms for the desired power- and energy-efficiency as well as to reduce operational costs of HPC-supercomputers and cloud-systems with thousands of