Preventing Network Bottlenecks: Accelerating Datacenter Services with Hotspot-Aware...

on YouTube ↗18m

" summary

Analysis of Google datacenter ToR hotspots shows links that persist hot for hours and often exceed 75–90% utilization, causing over 2× end-to-end latency for some services. Root causes are bandwidth demand–supply imbalance and placement that ignores network capacity. Practical fixes add network-awareness to Borg scheduling and Colossus placement: avoid placing or migrate network‑heavy tasks away from hot ToRs and place data under ToRs with available bandwidth, yielding 90% fewer hot ToRs and p95 latency down 50% in deployed clusters.

AI-written summary · about

" chapters (9)

1. 0:00Network hotspots in datacenters
2. 1:41Why hotspots occur
3. 5:35Hotspots on storage systems
4. 7:19Quantifying hotspot effects
5. 9:48Hotspot sensitivity in queries
6. 11:52Network-aware scheduling
7. 13:44Hotspot-aware placement in Borg
8. 15:11Summary and future work
9. 17:16Q&A

" related talks

• 12mBreaking the Transience-Equilibrium Nexus: A New Approach to Datacenter Packet TransportMohammad Alizadeh, Anirudh Sivaraman, Shiyu Liu, Ahmad Ghalayini, Balaji Prabhakar, Mendel Rosenblum · NSDI ’21
• 18mMeasuring Congestion in High Performance Datacenter InterconnectsArchit Patke, Saurabh Jha, Jim Brandt, Ann Gentile, Benjamin Lim, Mike Showerman, Greg Bauer, Larry Kaplan, William Kramer, Ravi Iyer, Zbigniew T. Kalbarczyk · NSDI 2020
• 17mNetHint: White-Box Networking for Multi-Tenant Data CentersWei Zhang, Ion Stoica, Jingrong Chen, Danyang Zhuo, Hong Zhang, Liang Luo, Jeffrey Chase · 19th USENIX Symposium on Networked Systems Design and Implementation (NSDI '22)
• 31mTuning Linux TCP for data-center networksYuchung Cheng · Linux Plumbers Conference 2022
• 16mOptimizing Network Provisioning through CooperationRanjita Bhagwan, Yogesh Bansal, Harsha Sharma, Parth Thakkar, Sagar Bharadwaj, Venkata N. Padmanabhan, Vijay Kumar, Kathleen Voelbel · 19th USENIX Symposium on Networked Systems Design and Implementation (NSDI '22)
• 29m"Datacenter Infrastructure Innovation"James Hamilton · Velocity 2010
• 18mArchitecting Multi-tenant Data-center Networks for GPU CustomersChang Kim, Weilong Cui · @Scale: Networking 2025
• 14mRDC: Energy-Efficient Data Center Network Congestion Relief with TopologicalWeitao Wang, T. S. Eugene Ng, Ang Chen, Dingming Wu, Sushovan Das, Afsaneh Rahbar · 19th USENIX Symposium on Networked Systems Design and Implementation (NSDI '22)