We first discuss mechanisms that identify producer-consumer sharing in applications and respond by migrating responsibility for tracking the relevant cache lines to the producer node ("directory delegation") and performing "speculative updates" to push modified data to consumers prior to them requesting the data. The mechanisms require little additional hardware and are verified to not violate sequential consistency. On a mix of benchmarks, we find that directory delegation and speculative updates reduce the average remote miss rate by 40%, reduce network traffic by 15%, and improve runtime performance by 21%.
We then discuss how to exploit a "heterogeneous interconnect" comprised of wires with varying latency, bandwidth, and energy characteristics. Intelligently mapping critical coherence messages to wires optimized for delay and non-critical messages to wires optimized for low power can improve performance while simultaneously reducing energy consumption. We found that doing so improves the performance of a mix of scientific and commercial benchmarks by 13.3%, while reducing energy consumption by 14% on a simulated 16-core CMP system.
Bio: Professor Carter is the Associate Director of the School of Computing at the University of Utah. He received his PhD in 1993 from Rice University for his work on the Munin distributed shared memory system. He is the author of over 50 technical papers and holder of half a dozen patents. His research interests include computer architecture, distributed systems, operating systems, and networking. His recent research projects include the Impulse "smart" memory system, the Ultraviolet petaflop/petabyte multiprocessor project (with SGI and MIT), and the SWARM wide-area data management middleware system. He served as the CTO of MangoSoft, serves on the technical advisory boards of several other startups, and remains interested in high-tech entrepreneurship.