Scheduling heavy-tailed data traffic over the wireless Internet
Ziling Shao, Upamanyu Madhow · 2003
We propose new concepts in quality of service (QoS) and scheduling for the wireless Internet that account for the following key observations: (a) Internet traffic is long range dependent, or heavy-tailed, but the conventional approach of handling this via more conservative resource provisioning is unattractive in relatively low bandwidth wireless networks; (b) flat rate charging is an attractive paradigm for data applications, in which case the network's objective becomes to satisfy as many users as possible, rather than be fair in a traditional sense. The QoS framework consists of guaranteeing a certain bandwidth, with high probability, to the transactions on the network, where a transaction might be interpreted, for example, as a TCP connection, or a group of TCP connections, for a given source-destination pair. The lengths of these transactions can vary widely, ranging from a large file transfer to a short E-mail or Web page download, and follows a heavy-tailed distribution. We demonstrate that much fewer network resources are required to satisfy a given QoS guarantee using scheduling strategies that implicitly or explicitly penalize long transactions, compared to a "fair" strategy such as round robin. For multiple users that see channels of different capacities on a shared wireless medium, we modify our QoS framework to be consistent with the proportional fairness criterion. We show that schedulers that discriminate against long transactions, while providing roughly equal "air time" to the users, work much better than conventional round robin based scheduling.