Pecan: Cost-Efficient ML Data Preprocessing with Automatic Transformation...
summary
Pecan is a service that automatically places preprocessing workers on accelerator hosts and applies automatic transformation reordering to cut remote CPU usage. The auto-placement policy trades remote workers for local ones (example: five remote → two remote plus six local) to hit peak ingestion while preserving auxiliary CPU needs. The auto-order policy computes an inflation factor to move low-volume ops earlier, raising throughput with negligible accuracy loss (vision Δ < 0.4%). Results: ~87% average preprocessing cost reduction and up to 60% total training cost savings.
AI-written summary · about
chapters (12)
- 0:00Pecan system intro
- 1:48Disaggregating input pipeline
- 3:34Cost analysis setup
- 5:02Pecan paper contributions
- 6:26Balancing local remote resources
- 8:32Auto placement policy
- 11:45Transformation aware preprocessing
- 13:32Auto order policy
- 16:26Policies impact on accuracy cost
- 17:28Pecan policies combined impact
- 18:51Summary of results
- 19:11Conclusion
related talks
18mPreparation Meets Opportunity: Enhancing Data Preprocessing for ML Training With Seneca
21mCost Effectively Deploy Thousands of Fine Tuned Gen AI Models Like...
15mPET: Optimizing Tensor Programs with Partially Equivalent Transformations and Automated
40mReaching digital transformation excellence
12mPreview: ML Serving
08mMaximizing ROI Through Distributed Computing and Multi-Cloud
13mPreview: ML Training
35mAccelerate your ML innovation with AWS ML infrastructure services