Learning Multicast Patterns for Efficient BIER Forwarding With P4

Bit Index Explicit Replication (BIER) is an efficient domain-based transport mechanism for IP multicast (IPMC) that indicates receivers of a packet through a bitstring in the packet header. Recently, BIER forwarding has been implemented on 100 Gbit/s per port hardware using the P4 programming language. However, the implementation requires packet recirculation to iteratively serve one next-hop after another. The objective of this paper is to reduce this inefficiency. Static multicast groups can be configured on P4 switches so that traffic can be sent to all next-hops without recirculation. We leverage that feature to make BIER forwarding more efficient. However, only a limited number of static multicast groups can be configured on a switch, which is not sufficient to cover all potential port patterns. In a first step, we develop efficient BIER forwarding that utilizes static multicast groups derived from so-called configured port clusters. Then, we design port clustering algorithms that observe multicast patterns and compute configured port clusters which are more efficient than randomly selected port clusters. These methods are based on Spectral Clustering, an unsupervised machine learning technique. We perform simulations that underline the effectiveness of this approach to reduce inefficient packet recirculations. We further implement the new forwarding behaviour on programmable hardware and provide a controller that samples BIER packets on the switch, runs the port clustering algorithms, and updates the configured static multicast groups. We validate this open source implementation in a testbed and show that the experimental results are in line with the simulation results.