Ordered Mesh Network Interconnect (OMNI) : design and implementation of in-network coherence

Abstract

CMOS technology scaling has enabled increasing transistor density on chip. At the same time, multi-core processors that provide increased performance, vis-a'-vis power efficiency, have become prevalent in a power constrained environment. The shared memory model is a predominant paradigm in such systems, easing programmability and increasing portability. However with memory being shared by an increasing number of cores, a scalable coherence mechanism is imperative for these systems. Snoopy coherence has been a favored coherence scheme owing to its high performance and simplicity. However there are few viable proposals to extend snoopy coherence to unordered interconnects - specifically, modular packet-switched interconnects that have emerged as a scalable solution to the communication challenges in the CMP era. This thesis proposes a distributed in-network global ordering scheme that enables snoopy coherence on unordered interconnects. The proposed scheme is realized on a two-dimensional mesh interconnection network, referred to as OMNI (Ordered Mesh Network Interconnect). OMNI is an enabling solution for the SCORPIO processor prototype developed at MIT - a 36-core chip multi-processor supporting snoopy coherence, and fabricated in a commercial 45nm technology. OMNI is shown to be effective, reducing runtime by 36% in comparison to directory and Hammer coherence protocol implementations. The OMNI network achieves an operating frequency of 833 MHz post-layout, occupies 10% of the chip area, and consumes less than 100mW of power.