DIAL: Distributed Elephant Flow Counting on SDN

In network measurement, many statistics counters need maintaining on the switch and thus consuming a lot of very precious on-chip memory. In fact, most network flows in the Internet are small flows that don't need large-width counters. In this paper, we present DIAL, a distributed counting approach, which duplicates the counting rules to leverage global memory resources, maximizing the counting efficiency, which is complementary to existing counter solutions. We pose and formulate the problem of finding the optimal placement for duplicated counting rules. After proving its NP hardness, we give some heuristics to fast generate a near-optimal placement. After describing the feasibility of implementation, we carry out some evaluation for DIAL. Our simulated results with Internet traffic and topologies show that DIAL can significantly decrease the memory cost and increase the memory efficiency for both fixed-width and variable-width counter architecture, with acceptable extra overheads, which is a great save of the precious high-speed memory in the switch.