Holistic In-Network Acceleration for Heavy-Tailed Storage Workloads

Abstract

Storage workloads are typically heavy-tailed, and a small number of large requests incur a burdensome performance overhead. To this end, we present NetStore, an in-network storage accelerator that exploits the capability of emerging programmable switches. The key idea of NetStore is to directly process large requests in the network by leveraging switches as an in-network request processor. NetStore not only mitigates head-of-line blocking but also provides extra computational power for data storage. To overcome the strict resource constraints of the switch ASIC, we take a holistic approach that carefully co-designs the switch data plane and the switch control plane. Specifically, we design a custom control plane that acts as a dedicated request processor and a custom data plane that performs size-aware request scheduling and large object tracking. Our solution can be implemented on a commodity programmable switch at a line rate using only 6.82% of the switch memory. We implement a NetStore prototype on an Intel Tofino switch and conduct a series of testbed experiments. Our experimental results demonstrate that NetStore can improve throughput, the median latency, and the 99th percentile latency by up to $1.19\times $ , $21.29\times $ , and $2.91\times $ , respectively.

Keywords

• Programmable data plane
• P4
• in-network computing