Enhancing Network Failure Mitigation with Performance-Aware Ranking
summary
Swarm ranks mitigation actions by directly estimating end-to-end flow metrics using fast approximations: epoched traffic samples, splitting short/long flows, computing loss-limited vs capacity-limited rates, then a resource-allocation step. It reports orders of magnitude improvements in flow completion time and throughput on real cloud incidents, reduces baseline performance penalties from >80% to ≤0.1% on top-priority objectives and <5% elsewhere, and finds mitigations for a 16,000-server datacenter in <4 minutes. The system targets 1% throughput and other percentile objectives.
AI-written summary · about
chapters (9)
- 0:00Swarm and network failure mitigation
- 1:08Limitations of mitigation strategies
- 4:24Swarm’s performance-aware ranking
- 6:56Estimating performance with approximations
- 9:27Long flow throughput transport model
- 11:35Handling traffic uncertainty
- 12:37Generality and flexibility of Swarm
- 13:40Evaluation and performance results
- 15:26Conclusion and future directions
related talks
12mCost-effective Cloud Edge Traffic Engineering with Cascara
21mHigh Network Reliability & Availability in FE & BE for Scalable Training Solutions
25mResQ: Enabling SLOs in Network Function Virtualization
16mOptimizing Network Provisioning through Cooperation
18mMeasuring Congestion in High Performance Datacenter Interconnects
17mNetHint: White-Box Networking for Multi-Tenant Data Centers
37mAutomatically Detect the Top Performance & Scalability Issues in Distributed
24mHarp: Improving VPC Network Availability via Efficient Failure Detection and Rerouting...