Addressing Issues in Research on Packet Classification in Core Networks

dissertation

English

Although the Internet has changed significantly since the beginning of the 21st century, packet classification is still one of the most common operations implemented in networking devices. Nevertheless, the requirements on its performance are continuously increasing, especially in core networks. Currently, packet classification algorithms have to support 100 Gbps throughput. In addition, classification rule sets are becoming larger and the number of bits involved in the classification decision is growing due to 128-bit IPv6 addresses and classification according to more than 5 header fields in the OpenFlow protocol. Therefore, the majority of contemporary research on packet classification in core networks address the performace of packet classification algorithms, which has to keep pace with continuously increasing requirements. However, the researchers also focus on benchmarking newly developed algorithms because they have to be benchmarked using real rule sets, but such data are not available for most of the packet classification use cases. This thesis deals with both of these issues because it is important not only to design packet classification algorithms having high performance but also to assess their parameters by benchmarking based on proper data sets. Regarding the performace of packet classification algorithms, this thesis focuses on improving prefix matching, which is used in the majority of 1-dimensional and also multi-dimensional algorithms. Since a software implementation of prefix matching cannot fulfill the requirements imposed on packet classification in core networks, the thesis proposes a novel pipelined prefix matching architecture that targets Xilinx FPGA chips and utilizes their distributed on-chip memory. To fit the whole prefix matching data structure into FPGA's on-chip memory, this thesis also proposes a memory-efficient trie-based representation of a prefix set. The proposed representation is more memory efficient than well-known multibit tries Tree Bitmap and Shape Shifting Trie and for IPv4 prefix sets it also significantly overcomes the Prefix Partitioning lookup algorithm. The architecture then comprises two independent processing pipelines (to utilize both ports of on-chip memory blocks) that are together able to perform almost 255 million lookups per second, which translates into throughput of 170 Gbps for the shortest Ethernet frames. To allow realistic packet classification algorithms benchmarking, the thesis introduces a new open source synthetic rule set generator called ClassBench-ng, which integrates the generation of IPv4, IPv6, and OpenFlow 1.0.0 classification rule sets following the statistical properties specified in an input seed. Apart from the rule set generation, ClassBench-ng also supports an analysis of a real rule set in the ovs-ofctl format producing a corresponding seed that may be used for the generation of a similar synthetic rule set later on. Therefore, researchers having access to real classification rule sets can share their benchmarking data with other members of the community via statistical-based (thus anonymous) seeds produced by ClassBench-ng. With respect to the precision of the rule set generation process, ClassBench-ng is better than original ClassBench and FRuG in case of IPv4 prefixes and than Non-random Generator in case of IPv6 prefixes, when considering an average score for all IP prefix-related parameters. Moreover, it also clearly outperforms FRuG in the precision of OpenFlow rule sets generation.

Packet classification, core networks, prefix matching, FPGA, OpenFlow, ClassBench-ng.

MATOUŠEK, J.

12. 4. 2019

Faculty of Information Technology BUT

Brno

101

https://www.fit.vut.cz/study/phd-thesis/642/