Slowpoke: End-to-end Throughput Optimization Modeling for Microservice Applications
summary
Slowpoke predicts how a hypothetical optimization changes end-to-end throughput for microservice apps without tracing or a priori call graphs, by slowing down all non-target services and solving a linear program that enforces bottleneck equivalence. It instruments L7 libraries, uses a per-pod poker control process to coordinate pauses via signals, and runs batched slowdown experiments. Evaluated on four real applications plus 108 synthetic topologies across 45 nodes, it reports ~2.07% RMSE in throughput prediction.
AI-written summary · about
chapters (9)
- 0:00Slowpoke introduction
- 1:21Optimization prediction challenges
- 2:26Predicting throughput improvements
- 3:51Causal profiling limits
- 5:14Bottleneck model with linear programming
- 7:43Distributed slowdown coordination
- 9:07Data plane components
- 10:11Benchmark prediction accuracy
- 12:37Summary and open-source
related talks
17mAutothrottle: A Practical Bi-Level Approach to Resource Management for SLO-Targeted...
14mPerformance Optimization in High-Volume Transaction Systems
42mPredictive Load-Balancing: Unfair but Faster & more Robust
30mOn scaling Machine Learning microservices
01h 01mContinuous Optimization of MIcroservices using Machine Learning
45mUnderstanding Microservice Performance
11mLightning Talk: Grasp the Real Worth of Your Microservices by Brewing a Truth Poti...
48mPerformance testing of microservices in action