I see a potential benefit in the first one, in that each time a processor broadcasts, it doesn't have to send it to so many processors. This allows for scaling of the broadcasts. However, I imagine that if suddenly a processor on another interconnect wanted to do something with the cache line that a processor on another interconnect already modified, then there will be overhead in getting visibility of that modification to the new processor. Thus, I would believe that thrashing becomes even worse in this case. In the second example, having different memory banks could be weird if some memory needs to be shared between processors on different interconnects. In that case, I suppose they could be duplicated in memory, and this in itself has a lot of overhead.
I see a potential benefit in the first one, in that each time a processor broadcasts, it doesn't have to send it to so many processors. This allows for scaling of the broadcasts. However, I imagine that if suddenly a processor on another interconnect wanted to do something with the cache line that a processor on another interconnect already modified, then there will be overhead in getting visibility of that modification to the new processor. Thus, I would believe that thrashing becomes even worse in this case. In the second example, having different memory banks could be weird if some memory needs to be shared between processors on different interconnects. In that case, I suppose they could be duplicated in memory, and this in itself has a lot of overhead.
This comment was marked helpful 0 times.