Reliable Multicore System Design Using Online Thermal- and Reliability-aware Core Clustering and DVFS Techniques

Abstract

As the advance of CMOS technology transistor geometry keeps shrinking and transistor switching speed improves Meanwhile with reduced transistor geometry and increased power density and temperature reliability has become a major concern Dynamic voltage and frequency scaling (DVFS) is a widely used technique to reduce dynamic power consumption in multicore systems and can be used to improve temperature and system reliability The chip-based DVFS technique that the voltage and frequency of all cores are the same is first proposed; however it is less effective due to less flexibility The core-based DVFS technique that each core can have different voltage and frequency is proposed to achieve greater flexibility and better improvement However since each core requires its own voltage regulator this technique will require significant hardware and design cost The cluster-based DVFS technique is recently proposed where several cores are grouped into a cluster and all the cores in the same cluster have the same voltage and frequency However this technique is done at compile time and is only targeted to reduce power consumption In this thesis we propose an online cluster-based DVFS technique targeting system temperature and reliability improvement This technique considers the temperature profile and mean time to failure (MTTF) under the influence of the Negative Bias Temperature Instability (NBTI) effect and proposes novel algorithms that periodically recluster cores according to the core states at runtime with cluster-based DVFS control for reliable multicore system design Compared to core-based DVFS our proposed technique can achieve 12 75℃ reduction in maximum temperature 8 99 ℃ reduction in average temperature 67 96% reduction in the hotspot numbers and 20 34% improvement in MTTF under the influence of the NBTI effect with 6% increase in EDP Compared to asymmetric cluster DVFS our technique can achieve 3% reduction in energy 2% reduction in EDP 14 56℃ reduction in maximum temperature 9 46℃ reduction in average temperature 69 21% reduction in hotspot numbers and 27 61% reduction in MTTF

Date of Award	2014 Sept 2
Original language	English
Supervisor	Ing-Chao Lin (Supervisor)