< previous page	page_542	next page >
Page 542 Table 8.3 Bus-based or snoopy cache coherence protocols. Source of a new cache line Number of States 3 4 4 5 Invalid Shared-Clean Private-Dirty Invalid Shared-Clean Private-Dirty Shared-Dirty Invalid Private-Clean Shared-Clean Private-Dirty Invalid Private-Clean Shared-Clean Private-Dirty Shared-Dirty Invalidate Memory Write-Inval. MESI Cache-to-cache data movement Synapse Berkeley Illinois Update Cache-to-cache data movement Firefly Dragon Table 8.4 Summary of snoopy protocols. Name Category Memory Copy Policy Unique Feature Invalidate Wr. invalid. Copyback Simplicity Berkeley Wr. invalid. Copyback Explicit memory ownership Illinois Wr. invalid. Copyback Clean private state; can supply data from any cache with a clean copy Firefly Wr. updatea Copyback for Private Write through for Shared Memory updated on broadcast Dragon Wr. update Copyback for Private Write through for Shared Memory not updated on broadcast aOr, Write broadcast in Eggers [83]. memory or by the cache that makes the most recent write to the shared data line (and now has an exclusive copy). Table 8.3 places seven widely discussed protocols into the various categories. Another important distinction among snoopy protocols is the complexity of the protocol itself. This is largely determined by the number of states that must be maintained by the individual cache lines for shared data. The simplest protocols use only three states to ensure coherency. More complex protocols use as many as five or more states for coherence. In general, the larger the number of states, the more complex the protocol, but the less traffic on the bus.
< previous page	page_542	next page >