DcNetCache: Datacenter In-Network Caching for Large-Scale Key-Value Stores

Leveraging programmability and high-perfonnance packet-processing capability of network switches, we design DcNetCache, a new in-network caching architecture for large-scale key-value stores housed in a fat-tree datacenter network. At the heart of DcNetCache lie three main components, including topology-cognizant cache allocation that utilizes independent hash functions to divide labor of caching among switches, load-aware query routing with the power-of-two-choices, and a twophase consistency protocol with a chain cache update mechanism for cache coherence. We implement a prototype of DcNetCache atop NS3, a discrete-event network simulator. We evaluate performance of DcNetCache through detailed simulations, and show that DcNetCache can dramatically improve throughput and latency for skewed workloads. We also demonstrate that DcNetCache reacts promptly to changes of key popularity and handles cache coherence gracefully.