CROSS-Sched: A Crossover-Aware ML Scheduler and its Benchmarking on Heterogeneous Systems

Modern computing systems increasingly feature heterogeneous architectures combining multicore CPUs with GPU accelerators, yet effectively distributing workloads across these diverse processing units remains challenging. This paper presents a comprehensive benchmarking study of machine learning-based adaptive scheduling for image convolution operations across CPU and GPU resources. We evaluate 96 experimental configurations spanning four diverse datasets with image sizes ranging from 28×28 to 224×224 pixels on a production system featuring dual Intel Xeon processors and an NVIDIA Tesla V100 GPU. Our RandomForest-based scheduler learns optimal device and thread count selection from historical performance data. Results demonstrate a critical performance crossover point: CPU-based OpenMP parallelization achieves superior performance for small images (up to 1.76× faster than GPU), while GPU acceleration dominates for larger images (11.80× speedup over sequential, 2.15× faster than best CPU). The adaptive scheduler achieves competitive performance with an 8.8% average gap to optimal, successfully navigating the CPU-GPU performance tradeoff without manual tuning. This work provides reference benchmarks and a complete reproducible artifact for future heterogeneous scheduling research.