Flow Aware Traffic Management

When the TCP and IP protocols were first proposed the emphasis was on a simple best-effort service. Later on TCP was modified in order to address the issue of congestion control, that is to make sure that the senders would not overwhelm the network by sending data too fast. Over the years these protocols have shown an excellent scalability and have allowed the Internet to grow from a network connecting a few research centers to one that connects a very large number of computers all over the world. In its simplest form, which is still by far the most widely used, TCP congestion control works by using drop tail queues that simply drop packets whenever there is no more space in the router’s buffers. TCP senders react to these losses by reducing their sending window and, hence, their rate. This mechanism, while simple and scalable, has several well known limitations: 1) often different flows experience synchronized losses, leading to lower link utilization [5], 2) when flows with different Round Trip Times (RTT) share the same bottleneck link, the flows with a smaller RTT will receive a larger share of the capacity [1]. Not only this allocation of capacity is non-optimal in general, but it cannot be modified as long as only drop tail queues and TCP are used. Over the years several solutions have been proposed to address these limitations, especially the first one. Random Early Detection (RED) [5] was the first, and by far the most widely known and implemented; Active Queue Management (AQM) algorithm aims at improving link utilization by desynchronizing packet losses for different flows, albeit without addressing the fairness issue. Furthermore, while RED can indeed improve performances in some cases, this is not always true [9] and tuning its parameters is non-trivial, hampering its widespread use. Many other AQM algorithms have been proposed (like REM [8], Blue [4], Stochastic fair blue [3], just to name a few), but, to the best of our knowledge, none of them is capable of addressing both issues while being easy to configure and supporting different fairness criteria. Another advantage of using AQM algorithms is that routers can use Explicit Congestion Notification, in order to inform a sender that it needs to reduce its sending rate (that is to cut its sending window). This reduces the number of dropped packets and can improve performances as the sender does not need to rely on timeouts and/or duplicate acknowledgments in order to infer congestion. We propose a new flow-aware traffic management mechanism that aims at addressing the two aforementioned limitations of TCP, while being self-configuring and supporting different fairness criteria. Clearly the fairness criteria has to be explicitly selected (configured) by the user; this corresponds to selecting a specific objective, leaving the optimal tuning of parameters to the system. This mechanism is part of AlcatelLucent’s “Semantic Networking” paradigm [12]. Given that several studies indicate that 20% of the flows are responsible for 80% of the traffic, and that the overwhelming majority of these are TCP flows, we propose to concentrate on controlling these long-lived TCP flows, often called “elephants.” In so doing it is possible to improve the Quality of Service/Quality of Experience offered to short-lived and streaming flows. By controlling each elephant individually it is also possible to better control the service received by each one and prevent flows from harming each other.