Throughput is better on a large buffer size because the switch is able to clear its buffer while waiting for a response to a credit update (a new flit to arrive).
This comment was marked helpful 0 times.
benchoi
Elaboration on point 2: if there are few/small buffers, when sending a credit back to tell the source node to send more data, the buffers may drain, leaving the path idle and thereby degrading throughput. If there are many large buffers, the delay while waiting for the source node to receive the credit and send more data is not a problem since the big buffer at the switch that sends the credit is still being drained.
Throughput is better on a large buffer size because the switch is able to clear its buffer while waiting for a response to a credit update (a new flit to arrive).
This comment was marked helpful 0 times.
Elaboration on point 2: if there are few/small buffers, when sending a credit back to tell the source node to send more data, the buffers may drain, leaving the path idle and thereby degrading throughput. If there are many large buffers, the delay while waiting for the source node to receive the credit and send more data is not a problem since the big buffer at the switch that sends the credit is still being drained.
This comment was marked helpful 0 times.