MAPS : Adaptive Path Selection for Multipath Transport Protocols in the Internet TR-2011-09

As the access links of the Internet get faster, there is an incentive for the end hosts to demand higher throughputs. To meet their demands with minimum change at the Internet core, it is necessary to explore the existing core capacity efficiently. Multipath transport protocols try to address this problem by allocating traffic efficiently and fairly over some preselected paths, thus utilizing the path diversity of the network. However, because the Internet is heterogeneous, i.e, different parts of the Internet have different capacities, it is critical to select those paths with great care. Without efficient path selection, some of the flows might have to pass the more congested part of the network. As a result, their throughput might be significantly lower than using another set of paths. Adaptive path selection for multipath transport layer (MAPS) solves the problem by actively probing the residue capacity of unused paths for each flow. Moreover, MAPS only initiates the probes sparingly, depending on the realtime quality of current subflows. Our simulation shows that with the efficient path selection scheme by MAPS, the underlying transport layer is able to improve the average throughput of bulk transfer by as much as 40%; moreover, it is also able to improve the fairness index significantly when the traffic demand is high.