Algorithmic Energy Saving for Parallel Cholesky, LU, and QR Factorizations

TL;DR

This paper introduces TX, a library-level DVFS scheduling method that analyzes task dependencies in matrix factorizations to significantly improve energy efficiency with minimal performance loss.

Contribution

The paper presents a novel, application-aware, library-level DVFS scheduling approach for matrix factorizations that outperforms OS-level solutions in energy savings.

Findings

TX saves up to 17.8% more energy than OS solutions.

Performance loss is negligible at 3.5% on average.

Applicable to Cholesky, LU, and QR factorizations.

Abstract

The pressing demands of improving energy efficiency for high performance scientific computing have motivated a large body of software-controlled hard- ware solutions using Dynamic Voltage and Frequency Scaling (DVFS) that strategically switch processors to low-power states, when the peak processor performance is not necessary. Although OS level solutions have demonstrated the effectiveness of saving energy in a black-box fashion, for applications with variable execution characteristics, the optimal energy efficiency can be blundered away due to defective prediction mechanism and untapped load imbalance. In this paper, we propose TX, a library level race-to-halt DVFS scheduling approach that analyzes Task Dependency Set of each task in parallel Cholesky, LU, and QR factorizations to achieve substantial energy savings OS level solutions cannot fulfill. Partially giving up the generality of OS level solutions per requiring library level source modification, TX lever- ages algorithmic characteristics of the applications to gain greater energy savings. Experimental results on two power-aware clusters indicate that TX can save up to 17.8% more energy than state-of-the-art OS level solutions with negligible 3.5% on average performance loss.