Here if we had a buffer per possible output, then wouldn't the blue flit that is currently blocked by the green flit be able to proceed, thereby allowing more blue to flow through?
This comment was marked helpful 0 times.
sbly
But here we only have two buffers/ lanes per router, and there are three types of packets (purple, blue, and green) at the router you're talking about.
This comment was marked helpful 0 times.
sluck
I think also that virtual channel flow would not be able to have a buffer at each router for every possible output because that would require predicting the maximum number of packets that would be buffered at a router at any one time (and it might also have high space utilization, e.g. use up a fair amount of memory, while not even using all of the buffers available at router at a time; or alternatively, if the buffers are split as on the previous slide, the buffer sizes might be too small). The idea behind virtual channel flow is to attempt reduce the amount of head of line blocking, but as depicted in this diagram, it cannot completely eliminate it.
This comment was marked helpful 0 times.
adsmith
I think sss was referring to a buffer per possible output as in output link from the switch, not a buffer per possible destination. This would only require a number of buffer channels equal to the radix of the switch, which is a fixed constant.
This could still cause blocking (say if a very large packet was filling up the buffer and making a smaller packet going along the same output link wait), but I'm not sure if this would be considered HOL blocking per se.
Here if we had a buffer per possible output, then wouldn't the blue flit that is currently blocked by the green flit be able to proceed, thereby allowing more blue to flow through?
This comment was marked helpful 0 times.
But here we only have two buffers/ lanes per router, and there are three types of packets (purple, blue, and green) at the router you're talking about.
This comment was marked helpful 0 times.
I think also that virtual channel flow would not be able to have a buffer at each router for every possible output because that would require predicting the maximum number of packets that would be buffered at a router at any one time (and it might also have high space utilization, e.g. use up a fair amount of memory, while not even using all of the buffers available at router at a time; or alternatively, if the buffers are split as on the previous slide, the buffer sizes might be too small). The idea behind virtual channel flow is to attempt reduce the amount of head of line blocking, but as depicted in this diagram, it cannot completely eliminate it.
This comment was marked helpful 0 times.
I think sss was referring to a buffer per possible output as in output link from the switch, not a buffer per possible destination. This would only require a number of buffer channels equal to the radix of the switch, which is a fixed constant.
This could still cause blocking (say if a very large packet was filling up the buffer and making a smaller packet going along the same output link wait), but I'm not sure if this would be considered HOL blocking per se.
This comment was marked helpful 0 times.