How happy are your flows: an empirical study of packet losses in router buffers

Studies of Internet traffic have revealed that traffic is consistent with self-similar scaling, shows long-range dependence, and that flow sizes are consistent with heavytailed distributions. However, how such characteristics affect fundamental network properties such as buffer overflows and therefore the loss process and link utilization has not been explored in detail. Relying on advanced instrumentation via NetFPGA cards, we perform a sensitivity study of the packet loss process within routers for different network load levels, flow size distributions, and buffer sizes. We find that packet losses do not affect all flows similarly. Depending on the network load and the buffer sizes, some flows either suffer from significantly more drops or significantly less drops than the average loss rate. Very few flows actually observe a loss rate similar to the average loss rate. Therefore, any single flow is very unlikely to observe the global packet loss process. Furthermore, the loss process can exhibit scaling properties.