Energy-aware Self-Adaptive Middleware for Heterogeneous Parallel Architectures

Abstract

Hardware in HPC environments in recent years has become ever more heterogeneous in order to improve computational performance and as an aspect of managing power and energy constraints. This increase in heterogeneity requires middleware abstractions to eliminate additional complexities that it brings. In this paper we present a self-adaptation framework which includes aspects such as automated configuration, deployment and redeployment of applications to different heterogeneous infrastructure. This therefore not only mitigates complexity but aims to take advantage of the existing heterogeneity. The overall result of this paper is a generic event driven self-adaptive system that manages application QoS at runtime, which includes the automatic migration of applications between different accelerated infrastructures. Discussion covers when this migration is appropriate and quantifies the likely benefits.

The power impact of hardware and software actuators on self-adaptable many-core systems

“…This section presents the actuation methods adopted in this work. In the literature, several specific actuators are presented, such as pause/unpause app, kill app, increase/reduce power cap , and set power cap [10] . However, our choice was based on their relevance and coverage.…”

“…The ODA paradigm requires self-awareness and self-configuration from the system to enable the adaptation policies [9,10] . In this work, self-awareness means that the system can observe itself through virtual or physical sensors, and self-configuration is the ability to adjust the available system actuators to update system settings.…”

“…The biggest challenge remains in how to control the actuators set of a many-core system in a coordinated way to provide the required adaptability to improve its energy efficiency. Note that energy-efficient self-adaptive techniques are also a concern in other domains, such as High-Performance Computing (HPC) [10,11] .…”