SmartNIC-based scheduling

Details

Supervisors

Barbette, Tom

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] en sciences informatiques, à finalité spécialisée

Abstract

Following Moore's law slowing down this last decade, single CPU cores aren't able to match always increasing network capacities anymore. In this situation, sharding imposes itself as a very scalable solution to fill this performance gap but requires flows to be processed by the same cores each time. The literature focused on finding a way to provide such characteristics while providing a balanced distribution between cores to maximize the available performances and reduced resource consumption. But with the emergence of SmartNICs, new opportunities open as hardware offloading becomes accessible. This document proposes oRSS, a packet scheduler that leverages SmartNIC offloading to provide a balanced flow distribution among cores based on processor utilization while maximizing the CPU cycles available for application processing. We prove that such a scheduler is possible and can provide better performances than well-known techniques such as RSS with a throughput gain that can reach 20% on average.