T. V. Lakshman, Partho P. Mishra and K. K. Ramakrishnan Bell Labs AT&T Research 101 Crawfords Corner Road 600 Mountain Avenue Holmdel, NJ 07733& Murray Hill, NJ 07974 firstname.lastname@example.org partho, email@example.comAbstract:
Rate based congestion management has been proposed for feedback control of Asynchronous Transfer Mode (ATM) networks to carry bursty data. The ``Explicit rate'' feedback control schemes suggested for ATM's Available Bit Rate (ABR) service attempt to manage congestion by having cooperating sources maintain a smooth output rate, without exceeding the capacity of the network. This results in small queueing delays at the switches and a low cell loss rate. Secondly, they attempt to achieve fair service with a wide range of queueing and service disciplines in the network, including FIFO queues.
Compressed video is inherently bursty with rate fluctuations happening over both short (less than typical round trip times) and long time scales. We suggest that short time-scale fluctuations can be absorbed by buffering whereas the long time-scale fluctuations need increased or decreased network bandwidth. Compressed video sources are also rate adaptive. We believe that schemes most suitable for transporting compressed video should be based on cooperation between the network and the traffic sources.
The mechanism proposed for the ABR service promises to have the appropriate features for negotiation of the rate between the source and the network for transporting compressed video as well. A Resource Management (RM) cell transmitted by the source carries the desired rate for a source. The rate provided by the network in the returning RM cell gives the appropriate rate information back to the source which may be used to adapt the bit-rate of a video encoder. The request made by the source has to be a suitably predicted rate to allow encoding the frame at the time feedback is received >from the network.
Maintaining low overall delay is critical especially for interactive video. The explicit rate scheme (with an appropriate switch rate allocation mechanism) ensures that the aggregate rate of all the sources sharing a resource remains below the resource capacity (viz., the link bandwidth). This in conjunction with ABR sources maintaining a smooth flow of cells at the current allowed transmission rate, to minimize burstiness, keeps queueing delays small.
We believe the minimum bandwidth assurance of the ABR service in conjunction with rate feedback can provide the appropriate service for a large class of video applications. Predicting the appropriate minimum rate to request for the VC, and predicting the source's demand dynamically are critical for the success of the technique espoused here. The primary requirement we impose on the network is to properly separate those flows that are admission controlled and require low delay from others (e.g., bursty data) that may not be admission controlled.
Features of our Proposed Scheme
The key features of our scheme are: 1) We use the inherent negotiation in scheme like ABR, via RM cells to allow sources to indicate their needed rate over very short intervals. This needed rate is generated using forecasts which exploit the high short-term correlation in video. The source adapts its rate to the rate communicated back by the network, whenever necessary. The source enhances its adaptation by using information about the source buffer occupancy and a recent set of allowed rates.
2) We use the minimum cell rate guarantee of ABR to ensure that the transported video streams get an acceptable service quality. Each source at call set-up time may decide for itself what this minimum acceptable level is. This use of the minimum cell rate distinguishes our scheme from completely rate-adaptive video such as those used in the Internet video tools (NV, VIC).
3) We expect a separation of the video flows, that are admission-controlled, from bursty data that may not be admission-controlled. As a result,we can achieve acceptably low delays for the video flows, since explicit-rate-based ABR maintains small network queues by minimizing burstiness (at the cell-level) and ensures that the capacity of the links in the network are not over-allocated (at least for the admission-controlled flows).
4) We propose a new rate allocation mechanism in the network based on a weighted max-min fairness criterion. This enhanced rate allocation mechanism allows the network to treat flows unequally (in proportion to their weights), with regard to the rate allocations at their bottleneck. With this mechanism, higher-rate sources whose quality is more likely to be affected are treated preferentially (instead of all sources experiencing the same rate reduction).
5) If sources are willing to adapt only minimally, they can operate in VBR-like mode by setting their MCR values to an effective bandwidth calculated using the buffering at the sources and acceptable loss at that buffer. We envisage the possibility of ``renegotiating MCR'' if the need arises. A specific scheme and associated analysis is the subject of further work, however.
6) Our method is similar to RCBR in that admission control is simple. However, our renegotiation is done without signaling and over short intervals. Consequently, for real time video, we neither need to detect changes in long-term (hundreds of milliseconds or seconds) bandwidth requirements nor make long-term bandwidth predictions. Our predictions are limited to a few frames (approximately 3-5 frames at most). For stored video, our scheme is simple and works very well since our forecasts may be perfect, with a simple look-ahead. No complex calculation of the optimal renegotiation schedule is needed.
We expect to present simulation results showing the efficacy of our proposed scheme, with a long video-teleconferencing trace. We'll show that the weighted max-min fair allocation enhances fairness in an application sense, since the eventual degradation in quality is more sensibly managed. We also show that even when the bottleneck link is being utilized near saturation, the total end-end frame delay is within acceptable levels (less than 300 milliseconds) over a WAN with a propagation delay of 80 milliseconds.
We believe that transporting video using explicit rate based feedback control, as we propose, has the potential to combine the best features of VBR, CBR and RCBR video without some of their primary drawbacks.