To go from "I" to "M", we "shout" on the bus to go to "S", and then "shout" on the bus to go to "M".
WITH E: One Bus "shout"
To go from "I" to "M", when no one else is has cache, we "shout" on the bus to go to "E", and we can go to "M".
This comment was marked helpful 0 times.
bwasti
E has identical reactions to shared, a RdX and Rd go to Invalid and Shared respectively. Couldn't the E be eliminated if every time the S wrote it asked if anyone else has that cache line and if not simply move directly to M without waiting for an update?
This comment was marked helpful 0 times.
wanghp18
E eliminates the situation of "false shared". Without E, PrRd/BusRd always tranfers I to S, even if local processor is the only one holds the cache line, i.e. the cache line is not shared.
If there is no remote processor does BusRd(X) and the local processor does PrWr in E state, there is no need to do BusRdX thus save one broadcast. And I think this is the benefit of including E.
This comment was marked helpful 0 times.
aew
I agree with wanghp18 and smklein, adding E allows us to go from E to M without any bus shouts. It improves efficiency by allowing us to shout only once when we go from I to M, and zero bus transactions when going from E to M. This improves efficiency when only one processor is writing to its exclusive cache line, because it doesn't have to shout to the other processors unnecessarily.
This comment was marked helpful 0 times.
kkz
@bwasti I think your idea is valid, but may incur excessive flushing (as M state is implicitly "dirty").
Quick summary of why we have "E":
WITHOUT E:
Two Bus "shouts"To go from "I" to "M", we "shout" on the bus to go to "S", and then "shout" on the bus to go to "M".
WITH E:
One Bus "shout"To go from "I" to "M", when no one else is has cache, we "shout" on the bus to go to "E", and we can go to "M".
This comment was marked helpful 0 times.
E has identical reactions to shared, a RdX and Rd go to Invalid and Shared respectively. Couldn't the E be eliminated if every time the S wrote it asked if anyone else has that cache line and if not simply move directly to M without waiting for an update?
This comment was marked helpful 0 times.
E eliminates the situation of "false shared". Without E, PrRd/BusRd always tranfers I to S, even if local processor is the only one holds the cache line, i.e. the cache line is not shared.
If there is no remote processor does BusRd(X) and the local processor does PrWr in E state, there is no need to do BusRdX thus save one broadcast. And I think this is the benefit of including E.
This comment was marked helpful 0 times.
I agree with wanghp18 and smklein, adding E allows us to go from E to M without any bus shouts. It improves efficiency by allowing us to shout only once when we go from I to M, and zero bus transactions when going from E to M. This improves efficiency when only one processor is writing to its exclusive cache line, because it doesn't have to shout to the other processors unnecessarily.
This comment was marked helpful 0 times.
@bwasti I think your idea is valid, but may incur excessive flushing (as M state is implicitly "dirty").
This comment was marked helpful 0 times.
@bwasti then the protocol is MSI.
This comment was marked helpful 0 times.