Why were these hypercube systems used in the first place? What kind of application would this kind of topology be useful for over an easier-to-visualize system?
c0d3r
Hypercube systems improve spacial locality. In the diagram above (with 8 processors per level), every processor is only 3 steps away from the farthest processor in its hierarchy. Compare this to a linear system, in which a processor could be 7 steps away from the farthest processor.
stl
What does it mean that the "root of the network can become a bottleneck" and why does that happen with hierarchical snooping?
c0d3r
I believe that means that the traffic to/from a level can become extremely congested because all processors in that level are communicating using the same network. This only happens when processors are grouped this way.
ote
Why are there larger latencies than direct communication? Shouldn't the latencies be the same considering it is the same snooping happening, just at every level?
enuitt
I presume the latencies are due to the bottleneck from being bandwidth limited when multiple processors try to access the same memory
eknight7
@ote, @enuitt: The latencies can also be because now each processor has to send messages to multiple interconnects, one on which its connected with other processors, and then the next higher level and so on.
Why were these hypercube systems used in the first place? What kind of application would this kind of topology be useful for over an easier-to-visualize system?
Hypercube systems improve spacial locality. In the diagram above (with 8 processors per level), every processor is only 3 steps away from the farthest processor in its hierarchy. Compare this to a linear system, in which a processor could be 7 steps away from the farthest processor.
What does it mean that the "root of the network can become a bottleneck" and why does that happen with hierarchical snooping?
I believe that means that the traffic to/from a level can become extremely congested because all processors in that level are communicating using the same network. This only happens when processors are grouped this way.
Why are there larger latencies than direct communication? Shouldn't the latencies be the same considering it is the same snooping happening, just at every level?
I presume the latencies are due to the bottleneck from being bandwidth limited when multiple processors try to access the same memory
@ote, @enuitt: The latencies can also be because now each processor has to send messages to multiple interconnects, one on which its connected with other processors, and then the next higher level and so on.