Energy Efficient And Low Latency Interconnection Network For Multicast Invalidates In Shared Memory Systems

Abstract

Optical network-on-chip (NoC) are being investigated to reduce the latency and power consumption of networks for multicore processors. Our previous work has shown that switched optical networks can achieve lower latency for a given power consumption and component count in shared memory processors compared with arbitration-free networks such as single writer multiple reader. We have also shown the advantage of leaving optical circuits open after being generated to capture multiple memory transactions. However invalidation processes, where numerous cores are sharing a memory block, need to establish a large number of very short lived circuits and this increases the average message latency and overall on-chip contention. In this paper, a low power broadcast architecture is proposed which deals specifically with multicast messages. Separating multicast messages from unicast ones shows an improvement in average arbitration latency of up to 88.2% for the Vips benchmark while the Swaptions benchmark shows the highest improvement in average memory access time (up to 21.1%). Vips also sees an increase of 147% in the average number of messages passing through an open optical circuit. Obtaining these advantages requires an additional broadcast network which consumes only 66.1mW power.