Contracting Wide-area Network Topologies to Solve Flow Problems Quickly
summary
NCFlow contracts the WAN into clusters and replaces the full multi-commodity LP with a tiny aggregated inter-cluster flow plus parallel per-cluster subproblems, contracts the WAN into clusters, bundles inter-cluster demands into single aggregated edges, selects one inter-cluster link per cluster pair and iteratively shuffles assignments to avoid conflicts. On Azure and public topologies it reaches a median 98% of optimal total flow while running 8x faster and using 6x fewer FIB entries (maximum speedups up to 1300x) and can recompute after failures in ~10 ms.
AI-written summary · about
chapters (8)
- 0:00Introduction
- 0:29Traffic Engineering
- 1:28Slow Traffic Engineering
- 3:50Clustering
- 6:14Solving per cluster
- 8:11Solution
- 9:55Recap
- 10:27Results
related talks
14mOneWAN Traffic Engineering: Unifying Inter-Datacenter and Internet Backbones
21mTransparent and Flexible Network Management for Big Data Processing in the Cloud
15mMitigating Scalability Walls of RDMA-based Container Networks
17mNetHint: White-Box Networking for Multi-Tenant Data Centers
16mOneWAN is better than two: Unifying a split WAN architecture
16mOptimizing Network Provisioning through Cooperation
17mTopoOpt: Co-optimizing Network Topology and Parallelization Strategy for Distributed...
21mTowards An Application Objective-Aware Network Interface
