It seems like you might want to try optimizing algorithm so that a node only accesses memory which maps to a directory either within that node, or close to it, in order to reduce latency.
In fact, could there be a dynamic scheme where the location of the directory of each cache line is adjusted during run-time based on access patterns to reduce latency?
whitelez
@crow I think if there were a dynamic scheme, the communication overhead for synchronizing the directory setting would be high. The static assignment is a lot easier to implement and does not require any further communication for the mutual agreement on directory assignment.
It seems like you might want to try optimizing algorithm so that a node only accesses memory which maps to a directory either within that node, or close to it, in order to reduce latency.
In fact, could there be a dynamic scheme where the location of the directory of each cache line is adjusted during run-time based on access patterns to reduce latency?
@crow I think if there were a dynamic scheme, the communication overhead for synchronizing the directory setting would be high. The static assignment is a lot easier to implement and does not require any further communication for the mutual agreement on directory assignment.