Adaptive estimation and prediction of power and performance in high performance computing

Abstract

Power consumption has become an increasingly important constraint in high-performance computing systems, shifting the focus from peak performance towards improving power efficiency. This has resulted in significant research on reducing and managing power consumption. To have an effective power management system in place, it is essential to model and estimate the runtime power of a computing system. Performance monitoring counters (PMCs) along with regression methods are commonly used in this regard to model and estimate the runtime power. However, architectural intuitions remain fundamental with regards to the current models that relate a computing system’s power to its PMCs.

By employing an orthogonal approach, we examine the relationship between power and PMCs from a stochastic perspective. In this paper, we argue that autoregressive moving average (ARMA) models are excellent candidates for modeling various trends in performance and power. ARMA models focus on a time series perspective of events, and we adaptively update them through algorithms such as recursive-least-squares (RLS) filter, Kalman filter (KF), or multivariate normal regression (MVNR). We extend the notion of our model to predict near future power and PMC values. Our empirical results show that the system-level dynamic power is estimated with an average error of 8%, and dynamic runtime power and instructions per cycle can be predicted (65 time steps ahead) with an average error of less than 11.1% and 7%, respectively.