IEEE Access (Jan 2023)

MSI-A: An Energy Efficient Approximated Cache Coherence Protocol

  • Anant Saraswat,
  • Kumar Abhishek,
  • Hiteshwar Kumar Azad,
  • S. Shitharth

DOI
https://doi.org/10.1109/ACCESS.2023.3273219
Journal volume & issue
Vol. 11
pp. 48123 – 48135

Abstract

Read online

Energy consumption has become an essential factor in designing modern computer system architecture. Because of physical limits, the termination of Moore’s law and Dennard’s scaling has forced the computer design community to investigate new approaches to meet the requirements for computing resources. Approximate computing has emerged as a promising method for reducing energy consumption while trading a controllable quality loss. This paper asserts that an approximated cache coherence protocol preserves overall energy for computation. We can approximate the cache coherence protocol by adding approximated cache lines to a certain level without hindering the output. This paper introduces an enhanced approximated version of the MSI (Modified Shared Invalid) cache coherence protocol MSI-A (Modified Shared Invalid-Approx). We have verified MSI-A and MSI by employing LTL specifications in the NuSMV model checker. To illustrate the benefits of MSI-A, we have added DTMC (Discrete-Time Markov Chain) with PCTL (Probabilistic Computational Tree Logic). Although the PCTL proves the theory of approximation, we have also simulated the MSI-A in the TEJAS hardware simulator on PARSEC 3.0 to investigate the energy gains and cycle gains of MSI-A in varied applications. The cache lines considered to be approx are between 10 and 30 percent. Each application benefited from approximation according to its nature, and VIPS has indicated a total energy gain of 30.18 percent.

Keywords