An Experimentation and Analytics Framework for Large-Scale AI Operations Platforms
summary
A trace-driven experimentation and analytics framework models IBM production AI ops traces, builds a SimPy-based discrete-event simulator and generates synthetic workloads so researchers can test pipeline schedulers and resource policies. Task durations map to data size or ML framework, data assets sample from a Gaussian mixture for columns/rows/bytes, and arrivals use 168 hourly inter-arrival clusters. A one-day run simulates ~2,000 pipeline executions in ~5 seconds, exposing queueing, saturation and scheduler effects; runtime behavior modeling remains work in progress.
AI-written summary · about
chapters (17)
- 0:00Introduction
- 0:43Operationalizing AI
- 1:31User Platform Platform
- 3:47Pipeline Scheduling Challenges
- 4:06Simulation Framework
- 4:31Components
- 5:16System Model
- 6:53Runtime Model
- 8:47Dataset
- 9:49Process Model
- 11:20Generating Workload
- 11:41Modeling Workload
- 12:33Open Questions
- 13:04Conditional Modeling
- 13:40Simulation
- 17:05Pipeline Traces
- 18:30Conclusion
related talks
26mEngineering Excellence: Platforms and Architectures for Real-Time AI Analytics
16mOperationalizing AI in Finance: Real-Time Analytics
55mLarge Scale Distributed AI on the JVM
50mThe tech behind modern AI/ML use cases in Gaming
17mProductionizing Data Science Workloads
13mBuilding Scalable AI/ML Platforms for Industrial IoT
30mData Science for Infrastructure: Observe, Understand, Automate
30mSMART DATA PIPELINES FOR AI