Network-accelerated cluster scheduler

Abstract

Efficient use of computing clusters is crucial in large-scale data centers: even small gains in utilization can save millions of dollars. However, as the number of microsecond-scale tasks increases, using a CPU to schedule tasks becomes inefficient. Cluster scheduling running within the network can solve this problem, and brings additional benefits in scalability, performance and power efficiency. However, the resource constraints of programmable network devices make network-accelerated cluster scheduling hard. In this paper we propose P4-K8s-Scheduler, a network-accelerated cluster scheduler for Kubernetes implemented on a programmable network device. Preliminary results show that by scheduling Pods in the network at line-rate, P4-K8s-Scheduler can reduce the scheduling overheads by an order of magnitude compared to state-of-the-art Kubernetes schedulers.