My research, directed towards an MS thesis, focused on cache coherence protocols. In particular, I simulated hybrid update/invalidate protocols on a COMA-F machine. COMA architectures attack the problem of capacity and conflict misses with their large attraction memories, but they suffer more for coherence misses than comparable NUMA architectures. There are two reasons for this. First, COMA machines have an extra attraction memory lookup that almost always is a miss whenever the caches suffer a coherence miss. Second there is a higher probability of a 3-hop instead of a 2-hop read miss action because the home node in COMA-F machines does not necessarily have the data for shared lines, unlike in NUMA machines. This observation motivated me to study the use of a hybrid cache coherence protocol for a COMA machine.
Here is a link to the abstract as it appeared in my thesis.
This research was based on Alain Reynaud's work on hybrid protocols while he was a student here at the University of Illinois and a member of our group.
If you have comments or suggestions, email me at mitrevsk@uiuc.edu