Decentralized cloud wide-area network traffic engineering with BLASTSHIELD
Talkat 19th USENIX Symposium on Networked Systems Design and Implementation (NSDI '22) · 2022-04-04..04-06·
summary
BLASTSHIELD decentralizes Microsoft’s WAN by slicing the network and running independent slice controllers that program only local path segments (via MPLS) and require no inter-controller coordination. An online TE scheduler uses max‑flow/path‑diversity search to pick constrained paths, respecting a label stack depth of four and fallbacks so missing segment routes take conventional routes. Deployed in production carrying a majority of backbone traffic, it tolerates controller failures, reduces traffic loss by 60%, shows ~3% controller weight mismatch and reserves ~15% scratch capacity.
AI-written summary · about
chapters (14)
- 0:00Intro
- 0:28Software-driven WAN
- 2:22SWAN traffic
- 3:01SWAN outage of global scope
- 5:31BlastShield slices
- 6:24Design assumptions
- 7:52BlastShield controller
- 9:21Inter-slice routing
- 10:48Blast ripple and routing loops
- 11:52Source routing
- 13:41Traffic engineering scheduler
- 15:45Symphony or cacophony
- 16:51Blast radius reduction
- 18:18Summary
related talks
14mOneWAN Traffic Engineering: Unifying Inter-Datacenter and Internet Backbones
17mBLADE: Adaptive Wi-Fi Contention Control for Next-Generation Real-Time Communication
33mNetwork and application protection on AWS
16mExpress Lane to Efficiency and Reliability: Multi-Dimensional Control in Meta’s Express..
18mTGW: Operating an Efficient and Resilient Cloud Gateway at Scale
35mIntelligent Load Balancing in Kubernetes
42mProtect your workloads from DDoS attacks with Shield Advanced
32mIntelligent cloud networking: Create simplicity, remove complexity (PRT214)