then becomes visible at the bottleneck router only after the first packet (or possibly a hypothetical packet that would have been sent otherwise but the response prevented sending it) after the response reaches the router. In total, this feedback loop takes one RTT for the effects of the congestion re- sponse to finally reach back to the signalling router. Therefore, any AQM decision based on the estimate about the “current link load” is already using a stale value for the load. In reality during an exponential load tran- sient, the load has already increased and the router is already committed beyond remorse to receiving that load increase. The increase of the load will become visible only later because the horizon problem still prevents the visibility during the congestion signalling RTT, yet the router can no longer signal in time to have any effect on that load increase.
The research question RQ5 looks for a solution to the “current link load” problem by taking the horizon problem and RTT uncertainty into account. The research questions RQ6 and RQ7 go even beyond RQ5 and attempts to rectify the RTT-long feedback latency and calculate the correct time for sending the congestion signal.
3.3
Alleviating AQM Problems with Load Tran-
sients
In order to avoid most of the problems the exponential load transients cause to AQM algorithms, we believe a paradigm shift in the AQM research is needed. To the best of our knowledge, handling the exponential load transients has never before been a focus during the design phase of the AQM algorithms. Instead, the main approach is to focus on Congestion Avoidance and if any attention is paid to the exponential load transients due to bandwidth probing, all some AQM algorithms aim to do is to not react to them (e.g., [81, 103]). Therefore, there simply are no AQM algorithms that handle exponential load transients well. Because of ignoring exponential load transients, the designed AQM algorithms have a hard time to correctly manage them as they are much more rapid than the algorithms expect by design, which results in delay spikes. Others measuring AQM performance during exponential load transients echo this observation [93].
In order to resolve the issues with the exponential load transients, we believe AQM algorithms must be designed primarily for handling exponen- tial load transients that are much more aggressive than the behavior during Congestion Avoidance. The difference between exponential load transient and Congestion Avoidance is like the difference between a nasty storm and
56 3 Active Queue Management During Flow Startup calm waters. As exponential load transients are frequent in the general Internet due to flows constantly starting up using TCP Slow Start, ignor- ing the effects of exponential load transients does not seem a wise design decision. We also believe that if an AQM algorithm is able to handle expo- nential load transients well, it will be much easier to handle also Congestion Avoidance that is much less aggressive compared with the exponential load transients.
Our first attempt to solve the issues with exponential load transients is based on the existing RED algorithm and could therefore possibly exploit the existing real hardware that is already deployed with the RED capability. We independently understood that RED is based on a sane design principle in determining the link load in transient-aware manner before finding it already mentioned [81]. Effectively, a properly configured RED offers a solution to the horizon problem. Such a solid foundation inspires us to use it as a basis for our work rather than some other AQM algorithms that do not have the same design principle. Unfortunately, the transient awareness is not reflected in the RED parametrization guidelines [77, 80].
In Pub. II we present HRED (Harsh RED) that takes exponential load transients into account by reversing the common AQM control reasoning “when at the earliest to drop” into “when at the latest must drop” to produce parametrization that puts a deadline on arresting the rapid load growth during transients. As a result, HRED achieves reasonable perfor- mance during exponential load transients stopping them in time. HRED successfully stays in the pro-active dropping mode spacing losses out in contrast to Taildrop or RED with recommended parameters that tend to drop many packets in a burst or even consecutive packets. However, there is a serious limitation with HRED as it needs known end-to-end param- eters, that is, even though HRED addresses the horizon problem during the exponential load transients, it is seriously affected by RTT uncertainty. The other limitations of HRED are its use of stale load estimate during a rapidly developing exponential load transient and its inability to adapt to the less aggressive Congestion Avoidance operating mode of TCP lead- ing easily to underutilization when no flow is undergoing an exponential load transient. The latter of them leads us to believe that also the AQM algorithm should have one operating mode responding to Slow Start and another for Congestion Avoidance.