johnousterhout
commented
1 year ago I just had another thought about this anomaly. In looking at the other PowerTCP graphs, it looks like the point where bandwidth sharing begins (about t=.001s) is almost exactly the time when the first (highest priority) message completes. Everything looks fine, but somehow things are getting confused in the transition to the next message. I wonder if the simulator is moving a new message into the second priority level, but forgets to promote the existing message to the highest priority level, leaving two messages executing in the same priority level? That would explain the bandwidth sharing.
serhatarslan-hub
The figure above illustrates an anomaly that was observed during measurements for the PowerTCP paper. In this experiment, 11 senders all tried to send long messages to a single receiver with a 25 Gbps downlink. The graph shows the bandwidth observed for each of the messages during the incast, with an overcommitment degree of 2. For the first 1 ms, Homa behaves as expected, medicating all of its bandwidth to the red message. However, between 1 ms and 2 ms bandwidth is shared equally between 2 messages; this should not happen with Homa. Then, after 2 ms, things get even stranger, with wild oscillations in the bandwidth allocated to various messages.
The expected behavior is for Homa to dedicate all of its bandwidth to a single message at a time until it completes, then dedicating all of its bandwidth to another message until it completes, and so on.