Load Balancing Roundup
summary
Overview of load balancing choices with concrete tests and a tool: a Perl-based lb_emu emulator models servers (cores, ticks, keepalives) and compares algos. L4 vs L7 behavior, HAProxy/nginx for static L4, perlbal for persistent L7, and tradeoffs are quantified: under 400 concurrent (80 backend slots) the emulator shows about 346 req/s with ~1.2s median latency during overload; under 40 concurrent it shows ~288 req/s and ~127 ms per request, demonstrating when queuing beats overloading.
AI-written summary · about
chapters (21)
- 0:00Intro
- 0:51OSS Load Balancers (just a few)
- 3:21First LB Algorithm Tip
- 6:10Balancing Methods
- 6:40L4 Roundrobin
- 10:34Smarter L4 Issues
- 12:17L7 (HTTP level)
- 14:49Applied Synthetics
- 15:25The Machine
- 17:16The Balancer
- 17:43Lbemu Config Example
- 19:14Peribal Config Example
- 22:29The Tests
- 23:42The Scenarios
- 24:25400 Concurrent Requests
- 24:45400, Ideal
- 25:35400, Swappy Bastard
- 26:01400, Sloppy Sysadmins (uneven)
- 26:3140 Concurrent
- 26:4940, Ideal
- 29:39Questions?
related talks
46mKeeping the Balance: Load Balancing Demystified
25mScheduling and Load Balancing in RHEL8 and Beyond
19mLoad is not what you should balance: Introducing Prequal
44msmart balancing with nginx+lua
33mCloud Native Load Balancing From Scratch
33mCloud Native Load Balancing From Scratch
19mFlash Sale Engineering
42mPredictive Load-Balancing: Unfair but Faster & more Robust