Work in Progress paper: Experiment Planning for Heterogeneous Programmable Networks

Abstract

Private and publicly-funded cloud infrastructure and testbeds increasingly feature programmable network hardware. Programmable network cards and switches support the execution of increasingly-complex in-network programs that can operate independently of end-hosts to improve the network’s performance, resilience and utilisation. Reasoning about in-network programs, their placement, and workloads is needed to plan jobs on programmable networks. On programmable testbed networks, this reasoning feeds into resource allocation, fairness and reproducible research. But this reasoning is made challenging by the performance and resource diversity of hardware and by the failure modes that can arise in a distributed system.Flightplanner is currently the most comprehensive reasoning system for distributed and heterogeneous in-network programs but it uses a custom formalism and tool implementation, making it difficult to understand, extend, and scale.This paper describes Lightplanner, a generalisation of Flight-planner’s reasoning system that has been implemented on Prolog. It provides an executable formalisation in a well-understood logic. By relying on Prolog’s proof search, Lightplanner is 10 smaller than Flightplanner’s implementation in C++, making×it better suited for others to understand, extend, and scale. A benchmark of publicly-available in-network programs is used to evaluate Lightplanner against Flightplanner. Though the time overhead is slightly larger, Lightplanner can find better allocations than the original, more complex C++ implementation.Lightplanner is being incubated to plan experiments in a local programmable network testbed at Illinois Tech, and as a future step it will be extended to work across federated networks such as FABRIC.