부산대학교 도서관

Congestion control in intermittently connected lossy networks (ICLNs) is critical to enable efficient use of network resources, and mitigate congestion-induced packet losses and delays. Unfortunately, well-known protocols for end-to-end congestion control, such as the celebrated Transmission Control Protocol (TCP), are not well-suited for use in ICLNs. This work builds on the idea of hop-by-hop flow control and dynamic buffer-space allocation to develop a congestion-control approach able to react quickly to the onset of congestion with local congestion-control actions, and propagate direct and indirect congestion-control indicators toward the traffic sources. The local flow-control policy is developed using a reinforcement learning framework that captures queue-occupancy and sojourn-time indicators, and is coupled with a dynamic buffer-space allocation policy developed based on the Markowitz portfolio selection framework. The buffer-space allocation policy is responsible for allocating unused buffer space across all active flows in the node based on the relative traffic loads. The performance of the proposed framework when handling congestion locally is explored via numerical tests on a simulated environment.