Don't Work on Individual Data Plane Algorithms. Put Them Together!

Abstract

Algorithms and data structures for data plane network functions have been extensively studied in the literature. Recently various compact data structures and algorithms have been used in data plane to achieve less memory cost and higher throughput. However, most of these studies only focus on individual network functions, such as packet forwarding information base (FIB), traffic measurement, and load balancing. To our knowledge, no study has been conducted to design compact data structures and algorithms for multiple and co-located network functions. We argue that there is a huge space of optimization if we design algorithms and data structures considering multiple co-located network functions, compared to designing them individually. It is because many of them share similar design goals and building blocks. We use two recently published methods as examples and present a new memory-compact design that serves both FIB and traffic measurement functions by a novel integration of the two methods. The preliminary results show that the new design can achieve almost 2x throughput compared to running them individually while achieving a higher accuracy of measurement using the same memory. In addition, we will discuss potential research directions and challenges. CCS CONCEPTS • Theory of computation → Sketching and sampling; Bloom filters and hashing; • Networks → Data path algorithms.