For their new chip-based memory chace design: “What we do is we just assign time stamps to each operation, and we make sure that all the operations follow that time stamp order"
For their new chip-based memory chace design: “What we do is we just assign time stamps to each operation, and we make sure that all the operations follow that time stamp order"
The future of the 70s, today! https://en.wikipedia.org/wiki/Lamport_timestamps
Someone trying to flesh out the title would be interesting, if indeed you can group cache-coherence mechanisms into a few big categories all dating to >30 yrs ago, and this one can legitimately be called a new category. The paper doesn’t at all make this claim, though; its related-work section is more modest. The web writeup does, but it doesn’t elaborate in much detail.
Since I’m used to hearing of only two major approaches here, bus snooping and directories, I’d say it’s roughly right, if you assume that all the improvements on directory-based coherency since the “Illinois Protocol” in 1984 are incremental and not “new mechanisms”.
but - I dug around a little to be sure I wasn’t forgetting anything, and in this slide deck from Onur Mutlu I found a reference to an ISCA paper from 2003 “Token Coherence: Decoupling Performance and Correctness”, by Milo Martin, Mark Hill and David Wood, which certainly is neither snooping nor directory-based, and I wonder if it ever went anywhere.
Link to the actual paper here: Tardis: Time Traveling Coherence Algorithm for Distributed Shared Memory