EdgeBalance: Model-Based Load Balancing for Network Edge Data Planes
summary
EdgeBalance is a model-based load balancer that predicts per-server CPU from packet-level features and uses a PID controller to adjust dispatch weights, aiming to equalize predicted CPU across cores. It enforces bi-directional flow affinity even when middleboxes rewrite headers by maintaining forward and reverse flow tables and a change flag. Implementation uses DPDK with per-core threads and lockless flow tables to run at millions of packets per second. Evaluation on NSF CloudLab shows faster convergence and higher throughput than static or polling-based approaches.
AI-written summary · about
chapters (12)
- 0:00Start
- 0:27Outline
- 1:43Edge Computing runs NFS
- 2:28Cloud Load Balancer
- 4:12Static LB is not enough
- 6:39Bidirectional Affinity
- 7:22Accurate CPU Prediction
- 8:35Dynamic Load Balancing
- 9:17High Performance
- 10:12Experiment Setup
- 10:59Evaluation: Outperform
- 11:29Conclusions
related talks
35mIntelligent Load Balancing in Kubernetes
46mKeeping the Balance: Load Balancing Demystified
33mCloud Native Load Balancing From Scratch
23mBalancing on the Edge: Transport Affinity without Network State
23mBuilding Unified Global Load Balancing for the Edge With k8gb
41mLoad Balancing Roundup
16mFLB: Fine-grained Load Balancing for Lossless Datacenter Networks
21mStateless Datacenter Load-balancing with Beamer