High-Speed Packet Classification: A Case for Approximate Sorting

Abstract

Buffer capacities at routers are ever-increasing to accommodate the extreme-scale increase in the volume of incoming traffic. Sophisticated packet classification is being adopted to address challenges in meeting application demands. With the number of rules for classification and packets arriving at routers per second reaching 100's of thousands to millions, special hardware is also being built for packet classification to increase throughput at routers. Sorting packets (based on various conditions) in the buffers offers a significant advantage for the packet classification process. The sorting step is a time-consuming step given the large volume of packets in the incoming buffers. We propose a technique to approximately sort the packets by capping the maximum number of comparisons that can be made. We show that the performance of well-known decision tree-based packet classification methods such as HyperCuts, EffiCuts, and SmartSplits can be improved by more than 32% with approximate sorting for when the number of packets is less than 10K at any given time. If the number of packets in the arriving buffer is greater than 10K, we can divide them into smaller portions to achieve similar gains.