Traffic Analysis of a Congested Link

While networks capacities are usually planned to cope with the offered traffic, transient link overloads may occur due to link failures or traffic peaks. According to previous analytical and simulation studies, performance degradation is relatively slow at moderate overloads. The reason is the high variance of the current flow size distribution. We check here these results on a traffic capture. We note a degradation still slower than predicted by the model. We indeed observe stable throughputs for users and connections. The processor sharing model with persistent flows may explain this good performance. Persistent flows impact continuously the performance but they have no influence on the stability of the system and possible overloads. Overloads are only due to dynamic customers. This model distinguishes congestion, saturation of a link, and overload, when the arrival rate exceeds the service rate. Although peer-to-peer traffic imposes an increase of link capacities on public networks, it makes them more robust to congestion