A bunch of people represented "processors", which requested data from memory, labelled either X or Y. Whenever someone wanted a request, they "shouted" at the interconnect. In parallel, this represented a lot of people yelling at Kayvon.
When a processor wanted to "write" to X, they had to yell at ALL the other processors and tell them to invalidate X from their cache. This was a mess, and it didn't scale well.
To resolve this problem, we added some people, who represented memory "directories". The "X directory" kept track of all the people sharing X. Thus, when someone wanted to write to X, only the "sharers" of X were notified. This scaled in a much better way than everyone shouting over the interconnect!
A demo was shown in class:
A bunch of people represented "processors", which requested data from memory, labelled either X or Y. Whenever someone wanted a request, they "shouted" at the interconnect. In parallel, this represented a lot of people yelling at Kayvon.
When a processor wanted to "write" to X, they had to yell at ALL the other processors and tell them to invalidate X from their cache. This was a mess, and it didn't scale well.
To resolve this problem, we added some people, who represented memory "directories". The "X directory" kept track of all the people sharing X. Thus, when someone wanted to write to X, only the "sharers" of X were notified. This scaled in a much better way than everyone shouting over the interconnect!
This comment was marked helpful 0 times.