Latency signals

It is clear that to provide for low latency transport we can no longer rely on loss as the primary signal for use in congestion control. However, it is still an important signal to indicate that the network path has been overloaded. We can no longer rely on loss as the main signal since it is only elicited once the buffers on the congested link have filled. The size of such buffers is unknown and if Bufferbloated can lead to excessive delay. By low latency we mean that the end-to-end delay of the flow is maintained close to the minimum RTT of the path, whilst achieving reasonable path utilisation. To provide for low latency transport between end systems, over the existing infrastructure, we need to develop techniques that allow us to have a good notion of the latency incurred by a flow on the network, and how it changes over time. This latency state information can then allow a congestion control algorithm to set an appropriate send rate. To allow for some measure of fairness between flows they each need to have a common procurement and interpretation of this state information. Thus each flow should be able to compete in an equal manner with other such flows by using this common interpretation by its congestion control system (as is the case now for loss). Furthermore to provide for adequate competition with loss-based approaches, such as TCP, it should be possible to relate this state information to TCP’s state. Thus the congestion control algorithm needs to assimilate both latency and loss state information so it may be possible to relate it to TCP’s behaviour. In general an adequate response in the face of a loss-based flow means that it needs maintain a reasonable share of the capacity of the path. To achieve this end the congestion control algorithm needs to be able to interpret loss signals in a similar fashion to the loss-based flow and choose a suitably competitive rate. However a key aspect of this functionality is being able to fall back to delay-based operation once a competing loss-based flow has subsided. Another aspect of this dual behaviour is that it is not generally possible for the congestion control system to maintain low delay whilst in direct competition with a loss-based flow due to the fact that loss-based flow will fill the buffers. Typically such state information may be obtained through measurement of the send and receive times of the actual packets in the flow, though it is also possible