An adaptive load balancing method for multiple paths using TCP flow statistics

摘要

We propose an adaptive load balancing method for multiple paths that makes it possible to achieve high TCP performance in each path. In conventional load balancing methods, link utilization is the main parameter to be balanced among multiple paths. However; when we take into account TCP-level performance, balancing traffic in terms of only link utilization may not always result in balancing TCP performance on each path. Our method utilizes flow statistics such as the number of active flows in each path, which is easy to measure at ingress nodes. By adaptively equalizing the average bandwidth used per active flow in each path, which is calculated by dividing the input rate to the path by the mean number of active flows, our method achieves fair and high TCP performance for each path. Unlike other methods, intermediate nodes between an ingress-egress pair are not required to perform traffic controls or measurements besides normal packet forwarding. We describe a load balancing method for adaptively equalizing the average bandwidth used per active flow on each path and show its effectiveness under any heterogeneous condition through simulation analysis.