Abstract
To avoid a congestion collapse, network flows should adjust their sending rates. Adaptive flows adjust the rate, while unresponsive flows do not respond to congestion and keep sending packets. Unresponsive flows waste resources by taking their share of the upstream links of a domain and dropping packets later when the downstream links are congested We use network tomography-an edge-to-edge mechanism to infer per-link internal characteristics of a domain-to identify unresponsive flows that cause packet drops in other flows. We have designed an algorithm to dynamically regulate unresponsive flows. The congestion control algorithm is evaluated using both adaptive and unresponsive flows, with sending rates as high as four times of the bottleneck bandwidth, and in presence of short and long-lived background traffic.
