Coherence Based Message Prediction for Optically Interconnected Chip Multiprocessors
Anouk Van Laer1,a, Chamath Ellawala1,b, Muhammad Ridwan Madarbux1,c, Philip M. Watts1,d and Timothy M. Jones2
1Department of Electronic and Electrical Engineering, University College London, London, United Kingdom.
2Computer Laboratory, University of Cambridge, Cambridge, United Kingdom.
Photonic networks on chip have been proposed to reduce latency and power consumption of on-chip communication in chip multiprocessors. However, in switched photonic networks, the path setup latency can create a high overhead, particularly for the short messages generated by shared memory chip multiprocessors (CMP). This has led to proposals for networks which avoid switching using all-to-all or single writer multiple reader (SWMR) networks which dramatically increase optical component counts and hence power consumption. In this work we propose a predictor which uses information from the coherence protocol and previously transmitted messages to predict future messages and hence hide the path setup latency by speculatively setup photonic paths. We show that a directly mapped predictor can achieve prediction hit rates of up to 85% for PARSEC benchmarks in a 16-core x86 system using the MESI coherence protocol whereas a more resource efficient set associative predictor can still achieve prediction rates up to 75%.
Full Text (PDF)