Aquila: A unified, low-latency fabric for datacenter networks
Talkat 19th USENIX Symposium on Networked Systems Design and Implementation (NSDI '22) · 2022-04-04..04-06·
summary
Aquila builds a unified datacenter fabric using a GNet cell-based L2 protocol, a custom Tin/TORonic ASIC and integrated switch, and Dragonfly topology with 40 pods (~1,000 hosts) to cut cost and blast radius. It uses shallow-buffer cell switching, credit-based flow control, virtual channels, adaptive routing, and PCIe-integrated RMA protocol engines. Measured results show sub-10 microsecond RMA execution and 40 microsecond tail RTT for IP at scale, and a >5x tail-latency reduction for a co-designed key-value store.
AI-written summary · about
chapters (13)
- 0:00Intro
- 0:48A call for new Datacenter Architecture
- 1:34Cliques as a Datacenter Building Block
- 2:26Introducing Aquila
- 3:10Design Goals
- 4:01Cost-effective Topology: Dragonfly[1]
- 6:55Building TIN
- 8:38Life of a Packet
- 10:11Experimentation Setup
- 11:061RMA and IP Uniform Random Latency
- 11:36Latency Isolation in Aquila
- 12:42What does that mean for E2E application metrics?
- 13:03Retrospective
related talks
15mAlibaba HPN: A Data Center Network for Large Language Model Training
20mVirgo: Scale-out Data Center Network for AI
18mArchitecting Multi-tenant Data-center Networks for GPU Customers
17mMeta’s Network Journey to Enable AI
18mThe Evolution of High Performance Networks: A Storage Perspective and Practices
19mDisaggregating Stateful Network Functions
25mShoal: A Network Architecture for Disaggregated Racks
25mSilicon and Networking
