Abstract—Multi-field packet classification is a fundamental function that enables routers to support a variety of network services. Most of the existing multi-field packet classification algorithms can be divided into two classes: cutting-based and merging-based solutions. However, neither of them is scalable with respect to memory requirement for all rule sets with various characteristics. This paper makes several observations on reallife rule sets and proposes a novel hybrid scheme to leverage the desirable features of the two classes of algorithms. We propose a SRAM-based parallel multi-pipeline architecture to achieve high throughput. Several challenges in mapping the hybrid algorithm onto the architecture are addressed. Extensive simulations and FPGA implementation results show that the proposed scheme sustains 80 Gbps throughput for minimum size (40 bytes) packets while consuming a small amount of on-chip resources for large rule sets consisting of up to 10K unique entries. I.

Abstract — A Bloom filter is a simple space-efficient randomized data structure for representing a set in order to support membership queries. In recent years, Bloom filters have increased in popularity in database and networking applications. In this paper, we introduce a k-stage pipelined Bloom filter architecture to decrease power consumption. In the bit-array of a Bloom filter, bits corresponding to the index pointed to by hashing functions are checked and a “match”/“mismatch” is determined. The match/mismatch determination process can be organized in a k-stage pipelined Bloom filter architecture. We present a k-stage pipelined Bloom filter, the power consumption analysis and utilize a software packet classifier to customize the k-stage pipelined Bloom filter architecture in packet classification. The results of the software packet classifier with real packet traces show that more than 75 % of mismatched packets can be detected by the first three stages of the pipelined Bloom filter architecture (the remaining 25 % comprises 17 % matched and 8 % mismatched packets). Therefore, a 4-stage pipelined Bloom filter architecture with one hashing function in the first three stages and k − 3 parallel hashing functions in the last stage is more appropriate for power consumption optimization in packet classification.