Minimizing Packet Delay via Load Balancing in Clos Switching Networks for Datacenters
We consider traffic scheduling in an <i>N</i> × <i>N</i> packet switch with an optical switch fabric, where the fabric requires a reconfiguration overhead to change its switch configurations. To provide 100% throughput with bounded packet delay, a speedup in the switch fabric is necessary to compensate for both the reconfiguration overhead and the inefficiency of the scheduling algorithm. In order to reduce the implementation cost of the switch, we aim at minimizing the required speedup for a given packet delay bound. Conventional Birkhoff-von Neumann traffic matrix decomposition requires <i>N</i><sup>2</sup> - 2<i>N</i> + 2 configurations in the schedule, which lead to a very large packet delay bound. The existing DOUBLE algorithm requires a fixed number of only 2<i>N</i> configurations, but it cannot adjust its schedule according to different switch parameters. In this paper, we first design a generic approach to decompose a traffic matrix into an arbitrary number of <i>N<inf>s</inf></i> (<i>N</i><sup>2</sup> - 2<i>N</i> + 2 > <i>N<inf>s</inf></i> > <i>N</i> configurations. Then, by taking the reconfiguration overhead into account, we formulate a speedup function. Minimizing the speedup function results in an efficient scheduling algorithm ADAPT. We further observe that the algorithmic efficiency of ADAPT can be improved by better utilizing the switch bandwidth. This leads to a more efficient algorithm SRF (Scheduling Residue First). ADAPT and SRF can automatically adjust the number of configurations in a schedule according to different switch parameters. We show that both algorithms outperform the existing DOUBLE algorithm.