Tuesday February 7, 2006
Hamerschlag Hall D-210
Carnegie Mellon University
Prior research indicates that there is much spatial variation in an
application's memory access patterns. Modern memory systems, however, use
small fixed-size cache blocks and as such cannot exploit the variation,
because increasing the block size would prohibitively increase bandwidth
demands. In this talk, I will show that memory access patterns in
commercial workloads are spatially correlated over large memory regions
(e.g., several kB), and that such patterns are repetitive and predictable
through code-based correlation. I propose a hardware mechanism, called
Spatial Memory Streaming, that learns repetitive patterns and streams
cache blocks ahead of demand misses. I evaluate this design on commercial
and scientific workloads, demonstrating speedups of 19% on average and
282% at best.
Stephen Somogyi is a fourth year graduate student in the Computer
Architecture Lab at Carnegie Mellon, working with Prof. Babak Falsafi.
Stephen's research involves techniques to accelerate multiprocessor
commercial servers through improvements to the memory system.