An Adaptive Multipath Routing Algorithm for Maximizing Flow Throughputs

In datacenter networks, as the communication intensive middleware such as Hadoop being popular, the traffic among servers is heavily increasing and its traffic patterns vary from time to time. Conventional tree network topologies and routing mechanisms with single path routing will cause congestion along oversubscribed links. Thus, load- balancing path selections are needed to alleviate congestion and improve application performance. Multipath routing algorithms can distribute traffic over diverse paths optimally than simple solutions like ECMP. They, however, request considerable computations, i.e. frequently updating the path cost tree and dynamically optimizing path selections, thus they do not adapt large scale datacenter networks. Multinomial Logit Based (MLB) routing, a lightweight multipath routing algorithm, has been proposed as a countermeasure. However, it is not always efficient against various network topologies and varying traffic patterns unless its control parameter is adaptively chosen. In this paper, we present an automatic parameter tuning method, which estimates the path cost of individual flows in the network and dynamically tunes the traffic dispersion parameter to reduce the total path cost of all flows. In our experiments with real OpenFlow switches, we exhibit that our method optimally tunes the parameter and maximize the total throughput of the flows.