Tree Ensemble Classifiers on heterogeneous platforms
summary
Evaluation compares Random Forest and Gradient Boosting running on an octacore heterogeneous octacore (1 big, 3 mid, 4 small) using ONNX Runtime and a weather dataset with up to 100 trees. Stock multithreading shows a clear concurrency penalty as faster cores idle waiting for slow cores, while a worker pool pattern with 400–800 sample batches balances load and delivers about a 4–5× speedup without GPU involvement.
AI-written summary · about
chapters (12)
- 0:00Introduction to Performance and Scalability Challenges
- 0:12Understanding Heterogeneous CPUs
- 1:09Introduction to Machine Learning Models
- 1:24Decision Trees and Their Applications
- 2:27Random Forest and Gradient Boosting
- 3:21Experimental Settings and Data
- 3:50Performance Analysis of Random Forest
- 5:00Performance Analysis of Gradient Boosting
- 5:33Optimizing Multithreading for Heterogeneous CPUs
- 8:03Worker Pool Pattern for Improved Performance
- 10:32Conclusion and Key Takeaways
- 11:07Q&A and Closing Remarks
related talks
31mDo You Want To Build A Forest?
20mMulti-algorithm Ensemble Learning at Scale: Software...
41mBetter Than Deep Learning: Gradient Boosting Machines (GBM)
44mEnsemble Techniques for High Performance Machine Learning
28mDistributed XGBoost on Ray
18mCompetitive Analysis of the Top Gradient...
20mHow the cloud changed the way I do analytics and research
24mTianqi Chen, Computer Science PhD Student, University of Washington at MLconf Seattle 2017