If we group multiple processors into in directory nodes, we decrease the number of bits needed to represent the directory to manageable levels, but transmission of a message will be node-to-node, instead of processor-to-processor. This is a similar trade-off as the one made in assignment 2.
rav
A longer cache line will in general incur more false sharing, however true sharing would decrease. I think this is application/program specific, in some cases, increasing the cache line size would improve performance.
koala
@rav I agree. False sharing can definitely happen if data needed crosses the cache line, but we are also caching more data each time we load. Whether increasing cache line size would improve the performance is dependent on the program.
Longer cache line may incur more false sharing.
If we group multiple processors into in directory nodes, we decrease the number of bits needed to represent the directory to manageable levels, but transmission of a message will be node-to-node, instead of processor-to-processor. This is a similar trade-off as the one made in assignment 2.
A longer cache line will in general incur more false sharing, however true sharing would decrease. I think this is application/program specific, in some cases, increasing the cache line size would improve performance.
@rav I agree. False sharing can definitely happen if data needed crosses the cache line, but we are also caching more data each time we load. Whether increasing cache line size would improve the performance is dependent on the program.