Integrated CPU and l2 cache voltage scaling using machine learning

  • Authors:

  • Nevine AbouGhazaleh;Alexandre Ferreira;Cosmin Rusu;Ruibin Xu;Frank Liberato;Bruce Childers;Daniel Mosse;Rami Melhem

  • Affiliations:

  • University of Pittsburgh, Pittsburgh, PA;University of Pittsburgh, Pittsburgh, PA;University of Pittsburgh, Pittsburgh, PA;University of Pittsburgh, Pittsburgh, PA;University of Pittsburgh, Pittsburgh, PA;University of Pittsburgh, Pittsburgh, PA;University of Pittsburgh, Pittsburgh, PA;University of Pittsburgh, Pittsburgh, PA

  • Venue:
  • Proceedings of the 2007 ACM SIGPLAN/SIGBED conference on Languages, compilers, and tools for embedded systems
  • Year:
  • 2007

Quantified Score

Hi-index 0.00

Visualization

Abstract

Embedded systems serve an emerging and diverse set of applications. As a result, more computational and storage capabilities are added to accommodate ever more demanding applications. Unfortunately, adding more resources typically comes on the expense of higher energy costs. New chip design with Multiple Clock Domains (MCD) opens the opportunity for fine-grain power management within theprocessor chip. When used with dynamic voltage scaling (DVS), we can control the voltage and power of each domain independently. A significant power and energy improvement has been shown when using MCD design in comparison to managing a single voltage domain for the whole chip, as in traditional chips with global DVS. In this paper, we propose PACSL a Power-Aware Compiler-based approach using Supervised Learning. PACSL automatically derives an integrated CPU-core and on-chip L2 cache DVS policy tailored to a specific system and workload. Our approach uses supervised machine learning to discover a policy, which relies on monitoring a few performance counters. We present our approach detailing the role of a compiler in constructing a custom power management policy. We also discuss some implementation issues associated with our technique. We show that PACSL improves on traditional power management techniques that are used in general MCD chips. Our technique saves 22% on average (up to 46%) in energy-delay product over a DVS technique that applies independent DVS decisions in each domain. Compared to no-power management, our technique improves energy-delay product by 26% on average (up to 64%).