Architecting Efficient Communication to Enable Data-Intensive Computing

Seminar
Wednesday, March 21, 2018
10:30 AM to 11:30 AM
EER 3.646
Free and open to the public

Despite the end of Moore’s Law on the horizon, there is no end in sight to the rapid growth in the volume of data and applications to make use of it. To use that data, it must be stored, accessed, and moved, and this communication is often more demanding than the computation on that data. Worse yet, inefficient communication can leave a system woefully underutilized, which increases costs to the point of limiting the amount of data that can be practically processed.

In this talk, I describe techniques to improve communication efficiency, which in turn can improve performance, save energy, or even reduce manufacturing costs. First, I demonstrate how to best exploit the technology advantages of silicon photonics to design efficient interconnect architectures. Next, I show the benefit of optimizing communication with a vertically-integrated approach that considers the needs of algorithms while simultaneously appreciating the capabilities of the underlying hardware. In particular, I diagnose communication bottlenecks for graph algorithms on current hardware and develop optimizations to ameliorate those bottlenecks. Finally, I conclude by discussing the importance of novel hardware architectures to achieve efficient communication.

x x

Speaker

Scott Beamer

Lawrence Berkeley National Lab

Scott Beamer is a computer architecture postdoctoral fellow at Lawrence Berkeley National Lab (LBNL). He designs architectures, systems, and algorithms to improve communication efficiency, which in turn saves energy and reduces cost for data-intensive applications. More generally, his interests are in computer architecture, systems, and memory interconnects. He received the Kaivalya Dixit Distinguished Dissertation Award from SPEC as well best paper awards from the International Parallel & Distributed Processing Symposium (IPDPS) and the International Symposium on Workload Characterization (IISWC).