“Network coding data planes with programmable switches”
Master’s thesis, Mestrado em Informática, Departamento de Informática, Faculdade de Ciências da Universidade de Lisboa, Oct. 2017
Abstract: Network Coding (NC) is a technique that can be used to improve a network’s throughput. In addition, it has significant potential to improve the security, manageability, resilience (to packet losses, link failures and node departures) and the support of quality of service, in both wired and wireless network environments. The idea is to allow intermediate nodes of the network (i.e. switches and/or routers) to mix the contents of incoming data packets before forwarding them. Something that, traditionally carried out at source nodes, is therefore extended to the network, creating an array of new options. The difficulty of deploying NC on traditional switches lies in the impossibility to change or extend their operation with the requirements of this new paradigm. The devices are closed, the software and underlying hardware are vendor specific, and follow a fixed set of protocols and processing pipeline. This rigidity precludes NC in today’s switches and routers. Fortunately, programmable switches are beginning to emerge, with some already achieving production-levels and reaching the market (e.g., Barefoot Tofino). A new high-level language to program these switches has recently been proposed: P4. The P4 language allows the precise definition of how packets are processed in these programmable switches. Namely, it enables the definition of headers, parsers, match-action tables, and the processing pipeline itself. Therefore, by taking advantage of these constructs, P4 enables the deployment of NC, on the switch’s data plane, for the first time. In this dissertation, we design and implement two NC switches using the P4 language. Both switches employ Linear Network Coding (LNC). The main difference is that the first (P4-XOR Switch), simply performs the XOR of packets (i.e., a linear code with field size 2). The second (P4-RLNC Switch) is more generic, allowing larger field sizes. For this purpose it performs Random Linear Network Coding (RLNC), which is a random variant of LNC. The evaluation was performed on Mininet (a network emulator) and focused on the functionality of both switches. Additionally, the performance of the P4-XOR Switch was tested as well. The main conclusion is that our implementations correctly perform the required operations allowing, for the first time, NC to be performed in real data planes.
Research line(s): Fault and Intrusion Tolerance in Open Distributed Systems (FIT)