Aquila - A Unified, Low-Latency Fabric for Datacenter Networks

Explore the cutting-edge Aquila datacenter network fabric in this 15-minute conference talk from NSDI '22. Dive into the design and implementation of a unified, low-latency network architecture that addresses the evolving needs of modern datacenter workloads. Learn how Aquila achieves ultra-low latency communication through its innovative Layer 2 cell-based protocol, GNet, integrated switch, and custom ASIC with Remote Memory Access capabilities. Discover how this experimental fabric significantly reduces tail latency for key-value stores and supports both traditional and tightly coupled workloads. Gain insights into Aquila's cost-effective topology, design goals, and performance metrics, including sub-40 microsecond tail fabric Round Trip Time for IP traffic and sub-10 microsecond RMA execution time across hundreds of host machines.

Intro
A call for new Datacenter Architecture
Cliques as a Datacenter Building Block
Introducing Aquila
Design Goals
Cost-effective Topology: Dragonfly[1]
Building TIN
Life of a Packet
Experimentation Setup
1RMA and IP Uniform Random Latency
Latency Isolation in Aquila
What does that mean for E2E application metrics?
Retrospective