One thing interesting is that in a mostly write environment, the sharer count can be very low (only the writer share the cache line). It can benefit more from the directory, because in snooping it needs to visit all cores while in the directory it only needs to visit one.
Firephinx
Also, since the sharer count is low in many cases, it is not necessary to have a bit for each processor and instead we can use a doubly linked list to indicate which processors are currently sharing the data, which is in the later slides. Although, the directory that has a bit for each processor allows the notifications to drop the line to be sent in parallel, since there will most likely be few sharers, iteratively sending the message to each of the sharers will on average not take much longer especially for frequently read/written objects.
One thing interesting is that in a mostly write environment, the sharer count can be very low (only the writer share the cache line). It can benefit more from the directory, because in snooping it needs to visit all cores while in the directory it only needs to visit one.
Also, since the sharer count is low in many cases, it is not necessary to have a bit for each processor and instead we can use a doubly linked list to indicate which processors are currently sharing the data, which is in the later slides. Although, the directory that has a bit for each processor allows the notifications to drop the line to be sent in parallel, since there will most likely be few sharers, iteratively sending the message to each of the sharers will on average not take much longer especially for frequently read/written objects.