Stateless Datacenter Load-balancing with Beamer
summary
Beamer replaces per-connection state with stateless muxes that attach the previous server and reassignment timestamp to each packet so a newly chosen backend can daisy chain mid-connection packets back to the server that has state. A central controller maps millions of buckets (test used 6.4 million buckets) to servers and updates muxes; the software implementation reports ~40 Gbps on a few cores and is about twice as fast as Google’s Maglev at the cost of ~16 extra bytes per encapsulated packet.
AI-written summary · about
chapters (19)
- 0:00Intro
- 0:50Datacenter load balancing today
- 1:53Strawman approach
- 2:22Load balancers use state to ensure connection affinity
- 4:00Beamer: stateless load balancing
- 4:30Beamer daisy chaining
- 5:09Balancing packets in Beamer Which hashing algorithm is best?
- 6:27Beamer hashing
- 7:02Beamer at work
- 9:16Beamer mux performance
- 9:42Single mux performance
- 10:26Realistic traffic
- 11:03Testbed evaluation
- 11:25Adding and removing muxes
- 12:55Adding servers
- 13:24Connection affinity under SYN flood attacks
- 15:43Control plane experiments
- 17:18Conclusions
- 18:02THE ADVANCED COMPUTING SYSTEMS ASSOCIATION
related talks
19mA High-Speed Load-Balancer Design with Guaranteed Per-Connection-Consistency
13mEdgeBalance: Model-Based Load Balancing for Network Edge Data Planes
41mLoad Balancing Roundup
46mKeeping the Balance: Load Balancing Demystified
19mCome Hell or Still Water: Alleviating Tail Latency in Cloud Block Store
42mPredictive Load-Balancing: Unfair but Faster & more Robust
26mIntegrating Stateful Load Balancing At Scale In Envoy
16mMedley: Optimizing Midgress Bandwidth for Commercial Live Streaming CDNs