Elastic RSS: Co-Scheduling Packets and Cores Using Programmable NICs

Abstract

Meeting Service-Level Objectives (SLOs) for workloads in today's datacenter environments places stringent demands on end-host servers: work conservation, tolerance to varying request service time distributions, high throughput, and CPU efficiency. Beginning with Receive Side Scaling (RSS), various schedulers have been proposed to steer packets to cores while preserving locality. However, these techniques are either too inflexible (randomly steering traffic at the NIC) or slow (bottlenecked by a central CPU-based scheduler). In this paper, we present Elastic RSS (eRSS), a system that extends traditional RSS by scheduling packets and cores using emerging programmable NICs with new abstractions (e.g., map-reduce). Operating at the NIC with minimal intervention from the host CPU, eRSS responds to load shifts at line rate and on a per-packet basis. eRSS also supports distributed packet stealing and fast preemption per-core to improve tail latency under heavy-tailed service time distributions. Our preliminary evaluation shows that eRSS increases CPU efficiency while responding to rapid load changes and meeting real-world tail latency constraints of 100 μs.