NFD » ndn-tools

When congestion marking is disabled and catchunks with the adaptive pipeline are only relying on timers to detect congestion, it may decrease the window more than once in response to one congestion event.

Its implementation of "conservative window adaptation (CWA)" requires it to "react to only one congestion event (timeout or congestion mark) per RTT." This option does not stop the behavior of consecutive window decrease due to an outstanding interest during one timeout event may trigger another time-out event after one RTT.

Example: segment 1 is sent at time 0s and will be lost. It triggers a timeout event at timepoint 0s+RTO. Segment 100 is sent right before 0s+RTO. Segment 100 is later lost and will trigger a timeout at 0s+2RTO. CWA requires any timeout event before 0s+RTO+RTT not to trigger a window decrease. As RTO>RTT, segment 100 will successfully trigger a second window decrease. However, segment 100 is sent out before adaptation of the previous congestion event. Thus, it should not trigger a window adaptation as the effect of the previous window adaptation is not observed yet.

Solution: Disqualifying outstanding segments during a window decrease from triggering another window decrease.