Locality Wall: Hardware Limitations and Software Opportunities

After decades of rapid improvement, the customary advance in hardware technology symbolized by Moore’s law is clearly decelerating. Continued performance scaling will depend less on new material and more on new organization. In particular, it requires new solutions to better manage and further reduce data movement, which has the largest impact on computing speed and system power. This problem, however, is highly complex as both the processor and memory on modern systems are increasingly parallel, distributed and heterogeneous.

This talk will discuss locality as the primary objective in the organization of both software and hardware. For software, locality is a quality of its dynamic data usage. This quality can be quantified scientifically using the concepts of access locality and timescale locality. The relation between these quantities and those of cache misses is given by the recent Higher Order Theory of Locality. This talk will cover the theoretical foundation of locality and discuss new opportunities in performance modeling and optimization.

Chen Ding is a Professor of Computer Science at University of Rochester. He received Ph.D. from Rice University, M.S. from Michigan Tech, and B.S. from Peking University. His research seeks to understand the composite and emergent behavior in computer systems especially its dynamic parallelism and active data usage and develop software techniques for locality optimization, data management, and program parallelization and optimization. His work received the young investigator awards from NSF and DOE. He co-founded the ACM SIGPLAN Workshop on Memory System Performance and Correctness (MSPC) and was a visiting researcher at Microsoft Research, a visiting associate professor at MIT, and has been a faculty fellow at IBM Center for Advanced Studies. More information about his work can be found at http://www.cs.rochester.edu/~cding/.

Host: Tim Nelson