On Fairness-Efficiency Tradeoffs for Multi-Resource Packet Processing

Middleboxes are widely deployed in today's data-center networks. They perform a variety of network functions, each requiring multiple hardware resources, such as CPU cycles and link bandwidth. Depending on the functions they go through, packet processing of different traffic flows may consume a vastly different amount of hardware resources. An effective algorithm is therefore highly desired to schedule packets in a way such that multiple resources are shared in a fair and efficient manner. However, we show in this paper that there exists a fairness-efficiency tradeoff when multiple resources are scheduled. Such a tradeoff has never been a problem for traditional single-resource fair queueing (e. g., GPS, WFQ, SCFQ, DRR) - as long as the queueing schemes are work conserving, both fairness and efficiency can be achieved simultaneously - and hence has received little attention. Therefore, a new and important research problem arises: given a desired fairness-efficiency tradeoff, how can we design a packet scheduling algorithm to reinforce such a tradeoff? We present our thoughts and observations in this paper.