Stochastic Networks with Multipath Flow Control: Impact of Resource Pools on Flow-Level Performance and Network Congestion

Abstract

Multipath flow control has been proposed as a key way to improve the Internet’s performance, reliability, and flexibility in supporting changing loads. Yet, at this point, there are very few tools to quantify the performance benefits; particularly in the context of a stochastic network supporting best effort flows, e.g., file transfers and web browsing sessions, where the metric of interest is transfer delay. This paper’s focus is on developing analysis tools to evaluate flow-level performance and to support network design when multipath bandwidth allocation is based on proportional fairness. To overcome the analytical intractability of such systems we study closely related multipath approximations based on insensitive allocations such as balanced fairness. We obtain flow-level performance bounds on the mean per bit delay, exhibiting the role of resource pooling in the network, and use these to explore scenarios where increased path diversity need not result in high gains. While insightful these results are difficult to use to drive network design and capacity allocation. To that end, we study the large deviations for congestion events, i.e., accumulation of flows, in networks supporting multipath flow control. We show that such asymptotics are determined by certain critical resource pools, and study the sensitivity of congestion asymptotics to the pool’s capacity and traffic loads. This suggests a disciplined approach to a capacity allocation problem in multipath networks based on a linear optimization problem.