Tree Ensemble Classifiers on heterogeneous platforms: Performance & Scalability Challenges

Abstract

Tree ensemble methods (Random Forest, Gradient Boosting) are widely used in ML but can be inefficient in cloud-based, multi-threaded environments due to uneven workload distribution across heterogeneous CPU cores. This talk analyzes performance trade-offs in existing ONNX-based implementations, introduces a custom C++ wrapper for optimized task scheduling, and demonstrates a 4x speedup in cloud-based inference workloads.