Batchy: Batch scheduling Data Flow Graphs with Service level Objectives
summary
Batchy is a scheduler for run-to-completion packet processing that profiles per-node batch gains, sets atomic fractional buffer sizes, and runs a one-step receding horizon controller (control period ~100 ms) to tune backlogs and trigger pipeline decomposition or core offload when buffering alone cannot meet SLOs. An analytical model picks per-node backlog targets; experiments on five 5G benchmark pipelines show 2–3× throughput improvement while respecting tight delay SLOs (including millisecond-class targets).
AI-written summary · about
chapters (12)
- 0:00Intro
- 1:21Batch Scheduling 101
- 1:57Batch Defragmentation
- 5:03Profile node processing times
- 7:14System Model
- 8:50Architecture
- 9:30Controllers
- 10:16Basic IP Router (L2L3)
- 11:22Static L2L3 results
- 12:17Mobile Gateway
- 13:52Systern Dynamics
- 15:16Conclusion
related talks
29mPredicting order batchability
20mBatchGen: An Architecture for Scalable and Efficient Batch Inference
22mHigh throughput data processing with GenStage
18mDVABatch: Diversity-aware Multi-Entry Multi-Exit Batching for Efficient...
34mWG-Batch Updates: What’s New and What Is Next?
37mIntroduction to the Kubernetes WG Batch
24mVolcano – Cloud Native Batch System for AI, BigData and HPC
32mCase Study of Batch Processing With Spring Cloud Data Flow Server in Cloud Foundry