Competitive router scheduling with structured data

We consider the task of transmitting structured information over bounded-capacity links. Our information model is a stream of basic units called superpackets that are broken into k packets each. To model the possible structure and redundancy of the superpackets, we assume that for each superpacket there is a collection of minimal subsets of packets whose delivery makes the superpacket useful. This very general model encompasses, for example, MPEG streams, where one can think of a group of pictures (GoP) as a superpacket. The fundamental difficulty is that networks can forward only the primitive packets, but applications can use only superpackets, and thus if no minimal subset is delivered, the whole superpacket becomes useless. Our aim is to maximize goodput (number of useful superpackets) in the face of overloaded communication links, where we are forced to drop some packets.Specifically, we assume that an arbitrary stream of packets arrives at a router with multiple bounded-capacity outgoing links. An online algorithm needs to decide, for each superpacket, which outgoing link to use (all packets of the same superpacket must use the same link) and, in case of an overload at a link, which packets to drop and which to transmit so as to maximize goodput. We analyze a simple randomized competitive algorithm for the general case and provide a nearly matching lower bound on the competitive ratio of any randomized online algorithm.