Measuring Congestion in High Performance Datacenter Interconnects
summary
Describes a deployable tool Monet that detects credit based congestion on a 3D torus (Blue Waters) using a percent time stalled counter aggregated into spatial congestion regions. Empirical dataset shows regions persist median ~10 hours and up to 24 hours, yields ~3,500 congesting events while default mitigation fires only ~260 times over five months (about an 8% trigger rate) and often fails; topology aware fixes raised throughput 56% in production.
AI-written summary · about
chapters (15)
- 0:00Introduction
- 0:25High Performance Computing
- 1:07High Speed Networks
- 2:08Questions
- 2:29Methodology
- 3:27Key Findings
- 4:57Definition of Congestion
- 5:58Percent Time Stalled Metric
- 6:50Congestion on the Network
- 7:18Congestion Visualization
- 7:39Congestion Regions
- 9:05System Size
- 10:15Longlived Congestion
- 14:46Conclusion
- 15:00Questions Answers
related talks
17mNetHint: White-Box Networking for Multi-Tenant Data Centers
12mBreaking the Transience-Equilibrium Nexus: A New Approach to Datacenter Packet Transport
19mTackling DC Congestion and Bursts
16mPowerTCP: Pushing the Performance Limits of Datacenter Networks
18mCCC: Re-architecting Delay-based Congestion Control in Datacenter Networks
14mRDC: Energy-Efficient Data Center Network Congestion Relief with Topological
18mPreventing Network Bottlenecks: Accelerating Datacenter Services with Hotspot-Aware...
17mTowards provably performant congestion control