Store-Ordered Streaming: Cooperative Mechanisms to Eliminate Coherent Read Misses in a DSM

Tuesday April 1, 2003
Hamerschlag Hall D-210
4:30pm

Stephen Somogyi
Carnegie Mellon University

Modern shared-memory mulitprocessor systems employ many techniques to improve the performance of memory coherence activities. Much as branch predictors in modern processors rely on repetitive branch outcomes in programs to speculatively execute instructions past branches, coherence predictors rely on repetitive sharing patterns in applications to predict coherence activity. Using accurate and timely predictors, a DSM can speculatively trigger protocol operations in advance to hide coherence overhead and communication latency.

Store-Ordered Streaming relies on the key observation that cache blocks are often consumed in the same order they are produced in distributed shared-memory systems. SORDS comprises a set of four co-operative mechanisms that target repetitive and predictable shared misses in such a DSM: Downgrade Predictors, Memory Read Predictors, Forward Streams, and Forward Queues. Together, these allow for accurate prediction of "when" a cache block has been produced, "who" will subsequently consume it, and a method for efficiently streaming blocks in correct order from producer to consumer. In this presentation, I will review Babak Falsafi and Chris Gniady's paper of the same name.

Stephen Somogyi is a first year graduate student in Electrical & Computer Engineering at Carnegie Mellon University, advised by Babak Falsafi. His research interests revolve around computer architecture, covering both uni- and multiprocessor systems, and techniques for overcoming the limitations imposed by traditional memory hierarchies.