Distributed Probabilistic Network Traffic Measurements

Abstract

Measuring the per-flow traffic in large networks is very challenging due to the high performance requirements on the one hand, and due to the necessity to merge locally recorded data from multiple routers in order to obtain network-wide statistics on the other hand. The latter is nontrivial because traffic that traversed more than one measurement point must only be counted once, which requires duplicate-insensitive distributed counting mechanisms. Sampling-based traffic accounting as implemented in today’s routers results in large approximation errors, and does not allow for merging information from multiple points in the network into network-wide total traffic statistics. Here, we present Distributed Probabilistic Counting (DPC), an algorithm to obtain duplicate-insensitive distributed per-flow traffic statistics based on a probabilistic counting technique. DPC is structurally simple, very fast, and highly parallelizable, and therefore allows for efficient implementations in software and hardware. At the same time it provides very accurate traffic statistics, as we demonstrate based on both artificial and real-world traffic data.