On Packet Classification with Multidimensional Range Trees

Packet classification plays a crucial role in numerous network services, including Quality of Service (QoS), Firewalls, and data streaming. In our research, we introduce a cutting-edge method for packet classification utilizing a modified multi-dimensional range tree, which enables effective rule storage and swift querying. This method provides a comprehensive solution for packet classification across various rulesets. Our use of the multi-dimensional tree structure significantly enhances the efficiency of packet classification, leading to better overall performance. The worst-case time complexity for our method is
$O(\log^{d}n)$
, with
$n$
representing the number of rules and
$d$
the number of dimensions/fields. Our method surpasses PartitionSort, which has a worst-case time complexity of
$O(n^{1/d})$
, underscoring the improved efficiency of our technique. Our approach includes two types of classifiers: one that identifies the first applicable rule (Classify First), and another that finds the most specific rule (Classify Specific) for a given packet. To assess the effectiveness of our method, we performed empirical tests using the Classbench dataset to simulate real-world conditions. We created standard 5-tuple rulesets of various sizes and seeds and compared our method with other techniques such as PartitionSort, Priority Tuple Space Search (TSS), HyperSplit and HyperCuts. Our findings show that our first rule classifier surpasses other techniques in all ruleset sizes, being 600% faster than PartitionSort for 64k rulesets. The specific rule classifier also shows increased efficiency, nearly matching PartitionSort for smaller rule sets and exceeding it by 30% for 64k size rulesets. Furthermore, our Classify First approach is shown to be empirically faster than other well-studied techniques such as SmartSplit, HyperSplit, HyperCuts, TSS, and others.