EPC: An ensemble packet classification framework for efficient and stable performance

The increasing demands of emerging network applications have compelled routers to offer enhanced functions, such as traffic accounting and quality of service (QoS). These functions rely heavily on packet classification. With network transmission speeds reaching unprecedented levels, the optimization of throughput has become a common practice. One such method is the deployment of multiple algorithm replicas with the best parameter configuration for parallel packet classification. However, this solution fails to address the issue of performance fluctuations in individual specific instance of algorithm (SIA). This is because most algorithms prioritize optimization of average lookup speed, often neglecting overall performance stability. Our evaluation of state-of-the-art algorithms has revealed that these algorithms commonly suffer from performance fluctuations due to data skewness. To address this issue, this work proposes a novel solution called Ensemble Packet Classification (EPC) that aims to achieve efficient and stable performance. EPC leverages the principles ensemble learning to generate an optimal combination scheme of diverse SIAs that exhibit similar performance but possess complementary characteristics. To evaluate the effectiveness of EPC, we select five state-of-theart algorithms as baselines. The experiment results show that when augmented with EPC, the throughput parallel solutions based on these algorithms increases by 12.07%-19.26%. Additionally, the 95th percentile lookup time is reduced by 14.78%-26.77%. By fully harnessing the complementarity of SIAs, EPC effectively addresses the issue of long-tail while increasing throughput.