Enhancing Rate Adaptive IP Streaming Media Performance with the use of Active Queue Management

Abstract

The Internet is today a world wide packet switching arena constituting enormous possibilities of new services and business creation. E.g., there is a clear tendency that more and more real-time services are making the jump from dedicated circuit-switched and broadcasting networks into packet switching. Examples are telephony, videoconferencing, and television. The Internet today is thus hosting a large set of different services, including the delay tolerant Web-surfing traffic, but also the non-delay tolerant real-time services. An additional challenge with most real-time traffic is that its traffic pattern do not adapt to the varying traffic load as Web-traffic do. Still, these new services work well, as long as the packet switching capacity is sufficient. Problems arise when the growth of real-time service usage is larger than the capacity increase. During peak hours, users will then start to experience media services fall-out and excessive communication delay.

The reason is that the Internet as we know it today was not built to handle such services at all. In motor traffic, as a comparison, queues build up when the traffic load is larger than the road and crossover capacity. The Internet behaves in a similar fashion: information is sent in packets that can be compared to cars. If too many packets are heading the same direction, queues of packets build up in the Internet routers, causing extra delay during such peak hours. In one way the Internet is more fearful than motor traffic: if queues get too long, new arriving packets are simply dropped, i.e. they just vanish. Luckily, there is no direct parallel to this phenomenon in the motor traffic comparison realm! To assist the queuing problems in motor traffic, special traffic lanes can be defined to allow e.g. only buses, taxis, and cars where the driver has at least one passenger, to drive in that lane.

Thus, these road-users will experience less delay in peak hours than the rest of the population. The Internet is tried “healed” with some comparable means. E.g. with the use of IntServ or DiffServ Quality of Service, packets belonging to high priority applications are treated in a preferential fashion. But what happens if too many applications start to use these “special-lanes”? What if the total capacity is over-loaded over a significant time period? The answer to fix the problem is simple: the aggregate traffic generation must slow down! In motor traffic, this means that each car carries more people (i.e. fewer cars in total), or, equivalently, that big cars are exchanged by smaller cars, thus producing smaller queues. In the multimedia real-time packet switching realm, the equivalent solution is that the same content must be compressed more efficiently, thus producing fewer and/or smaller packets.

This thesis proposes a solution for live interactive real-time streaming media where a tight interaction between the media sources and the network is very essential. A novel router architecture, “P-AQM”, for packet switched networks is its core component. The primary P-AQM design objective is native support for rate adaptive real-time multimedia flows, addressing low queuing delay and low packet losses even at high traffic load to assist conversational media flows. The second objective is bandwidth fairness among the media flows, but also fairness to elastic (TCP) flows. These two design objectives are achieved due to the aforementioned interaction between the network routers and the traffic sources: the routers signal the traffic congestion level, while the media and TCP sources apply rate adaption. TCP has built-in congestion control mechanism (e.g. Tahoe or Reno) that reacts on packet drops or packet ECN tags performed by the router. Real-time media using the UDP protocol has no standardized congestion control mechanism. While DCCP/TFRC has become a compelling IETF standard during the last years, the work of this thesis has chosen another solution for media rate control that bypasses the TFRC performance. Using the traffic congestion level signals from P-AQM routers, the media rate control can be done much more precise, react faster to traffic load changes, and obtain intra-flow global max-min fairness. The cost of these improvements is gradual deployment of the new P-AQM packet switching routers, and some added signaling traffic.

The P-AQM design is following classical control theory principles, and has been developed and improved using a combination of analytical and simulation tools. As a side effect to the need for true decodable rate adaptive video traffic, a simulation framework and tool-set, “Evalvid-RA”, was developed to generate such traffic. Evalvid-RA can also assist other researchers in improving their own work, e.g. applying rate adaptive video codecs over the DCCP/TFRC protocol.