Enhancing the delay performance of dynamic backpressure algorithms

The backpressure algorithm for dynamic network resource allocation achieves throughput optimality by making use of one-hop queue length differences. This elegant algorithm, on the other hand, does not yield good delay performance in general. We introduce a new class of enhanced dynamic backpressure algorithms which make use of a general queue-dependent bias function to exploit queue state information beyond one hop. We prove the throughput optimality of the enhanced algorithms. We further elaborate on two specific algorithms within this class, which have demonstrably improved delay performance while maintaining acceptable implementation complexity.