Worth mentioning, the H-Tree is still a blocking network. If every node from 0-7 talks with their counterpart on the right, only one communication will be fulfilled. This is the reason that we optimize the design and come up with the fat tree, to add more bandwidth to the upper level, thus provide close-to non-blocking communication with lowest cost
regi
Is the idea that the thicker, "fatter" links can support simultaneous messages, or just that they can service them faster? Would a fat tree still be considered blocking?
muyangya
@regi: I'm not sure if I'm getting this correct. Here's my thought:
According to Charles's paper proposing the fat tree: http://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=6312192. The "fatter" links simply implies that this link have a higher capacity(# of wires probably). I think it is most likely that it can support simultaneous messages. Even if it does not support simultaneous message, the ability to service the messages faster implies that multiple messages can be serviced in unit time frame.
@regi @muyangya In the lecture, the guest lecturer mentioned that the benefit of the fat link is so that multiple communications can occur at once. Technically, as long as all the links are fat enough, H-Trees would not be blocking.
BryceToTheCore
@regi I think the fat links just act as larger pipes that can handle more water.
ekr
@BryceToTheCore That analogy works, but misses the subtlety of this question. Higher bandwidth will allow for more messages to be sent per unit time, but we could still only be allowed to send one message at a time. In this case, the network is technically still "blocking", though it's much less likely to encounter a situation where it actually does end up blocking a communication.
Worth mentioning, the H-Tree is still a blocking network. If every node from 0-7 talks with their counterpart on the right, only one communication will be fulfilled. This is the reason that we optimize the design and come up with the fat tree, to add more bandwidth to the upper level, thus provide close-to non-blocking communication with lowest cost
Is the idea that the thicker, "fatter" links can support simultaneous messages, or just that they can service them faster? Would a fat tree still be considered blocking?
@regi: I'm not sure if I'm getting this correct. Here's my thought:
According to Charles's paper proposing the fat tree: http://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=6312192. The "fatter" links simply implies that this link have a higher capacity(# of wires probably). I think it is most likely that it can support simultaneous messages. Even if it does not support simultaneous message, the ability to service the messages faster implies that multiple messages can be serviced in unit time frame.
Also, according to http://ccr.sigcomm.org/online/files/p63-alfares.pdf. It is mentioned fat trees are rearrangeably non-blocking. So I thinks it should be considered non-blocking in a routable scenario.
@regi @muyangya In the lecture, the guest lecturer mentioned that the benefit of the fat link is so that multiple communications can occur at once. Technically, as long as all the links are fat enough, H-Trees would not be blocking.
@regi I think the fat links just act as larger pipes that can handle more water.
@BryceToTheCore That analogy works, but misses the subtlety of this question. Higher bandwidth will allow for more messages to be sent per unit time, but we could still only be allowed to send one message at a time. In this case, the network is technically still "blocking", though it's much less likely to encounter a situation where it actually does end up blocking a communication.