Scalable and Efficient Multi-rate Data Distribution (Extended Summary)

Data networks, such as wireless networks and the Internet, in addition to offering reliable point-to-point unicast communication, also support a multicast facility by which a sender can disseminate data to multiple receivers (clients) without incurring a per-receiver cost. One of the shortcomings of traditional multicast protocols is their inability to deliver data reliably to all receivers in the presence of packet loss in the network. This is primarily because the error recovery mechanism of unicast– acknowledgments from receivers–does not scale well under multicast: many simultaneous acknowledgments may overwhelm the sender’s ability to respond – a process known as ACK implosion. Significant progress towards reliable multicast communication was made with the introduction of Forward Error Correction (FEC) code [1], which uses error (more accurately erasure) correcting codes to enable receivers to recover source data in the presence of packet loss (erasures). This has since been generalized as the digital fountain approach [2], in which source symbols are encoded into a large number of encoded symbols using error correcting codes before being distributed through multicast. Extending the analogy of a real fountain, in a digital fountain one encoded symbol is as good as another – to a receiver, it only matters how many distinct encoded symbols it gets. The entire original source data can be recovered from enough encoded data symbols. An ideal digital fountain should have the following desirable properties: WORK CONSERVATION: A receiver gets no duplicate symbols and is able to recover source data immediately after receiving an equal amount of encoded data. ADDITIVITY: The rate of an aggregation of streams is effectively the sum of the rates of the individual streams. RATE EQUALITY: A receiver gets data at exactly its preferred rate, irrespective of other receivers of the same fountain. Although the digital fountain approach is an attractive choice for use in scalable bulk data distribution, significant issues remain. The most important comes from receiver heterogeneity: different receivers wish to download data at different rates, but the server (sender) must transmit data at a fixed rate through a fountain. This is clearly unacceptable for time-critical applications like video streaming, or for receivers with strict limits on their reception network bandwidth. An approximate solution to this [2, 3] is layered multicast: data is multicast at different rates over multiple channels; receivers subscribe to a subset of these channels according to their preferred rates. Usually, the subscription is cumulative: subscribing at layer i implies subscription to layers 0 through i − 1. Due to practical limits on the number of channels, layered digital fountains achieve only approximate rate equality. Using multiple channels can potentially violate the other two desirable properties of ideal digital fountains too: duplicate symbols transmitted across multiple channels lead to a non-work conserving layered fountain even if each channel (layer) separately is. Unwise choice of rates of different streams over multiple channels may cause their effective aggregate throughput to be much less than the sum of the individual rates. It is hence important to study how to assign sending rates to multiple channels to form a scalable and efficient data distribution fountain to accommodate heterogeneous receivers meeting their goal of downloading the original data with minimum cost. To our knowledge, there has been no prior work that attempts to quantify the performance implications of assigning rates in cumulative layered multicast to receivers with arbitrary rate utilities. Note that in our model each multicast channel allows packet loss but a receiver’s goal is to perfectly recover the entire original data. This is different from transmission using MDC (multiple description codes) where lossy recovery of the original data up to certain quality is allowed for an individual receiver based on its available network resource [4]. In this paper, we address the rate assignment issue by introducing rate-controlled streams (RCS). A RCS is a data stream encoded from an original source data, with a specific sending rate. We show a general framework where multiple RCSs can be combined to form a new RCS with higher rate, deriving conditions for achieving optimal stream-combining in terms of aggregated rate. We then study how to optimally assign rates to multiple streams which are sent over different channels with possible packet loss to meet various conditions.