zD

Abstract

Modern end-host network stacks have to handle tra c from tens of thousands of ⇥ows and hundreds of virtual machines per single host, to keep up with the scale of modern clouds. This can cause congestion for tra c egressing from the end host. The e⇤ects of this congestion have received little attention. Currently, an over⇥owing queue, like a kernel queuing discipline, will drop incoming packets. Packet drops lead to worse network and CPU performance by in⇥ating the time to transmit the packet as well as spending extra e⇤ort on retansmissions. In this paper, we show that current end-host mechanisms can lead to high CPU utilization, high tail latency, and low throughput in cases of congestion of egress tra c within the end host. We present zD, a framework for applying backpressure from a congested queue to tra c sources at end hosts that can scale to thousands of ⇥ows. We implement zD to apply backpressure in two settings: i) between TCP sources and kernel queuing discipline, and ii) between VMs as tra c sources and kernel queuing discipline in the hypervisor. zD improves throughput by up to 60%, and improves tail RTT by at least 10x at high loads, compared to standard kernel implementation. CCS CONCEPTS • Networks ⌅ Network architectures; • Software and its engineering ⌅ Communications management.