Isolation of Physical and Logical Views of Dark-Silicon Many-Core Systems for Reliability and Performance Co-Optimization

Abstract

A fraction of a many-core chip has to become powered off in order to maintain allowable power budget and safe temperature in the dark silicon era. Techniques have been developed to selectively activate cores in distributed physical locations to avoid temperature hotspot. It results in the unexpected increase of communication overhead due to longer average distance between active cores on Network-on-Chip (NoC). We propose a physical and logical isolated framework based on Folded Torus-like NoC for heterogeneous many-core systems to achieve the guaranteed temperature reliability and satisfied application performance requirement. Physically distributed cores are interconnected through folded torus-like NoC and organized in clusters to enable logically condensed intercommunications within it. Compared to traditional mesh-like systems, the proposed folded torus-like organization can achieve on average 39.44 % application performance improvement and decrease average 9.3\(^{\circ }\)C of the chip.