Machine Learning-based Interference Detection in GPGPU Concurrent Kernel Execution
Negar Sadat Alizadeh, Mahmoud Momtazpour · 2020
Recent advancements in GPU architectures have made it possible to run multiple kernels concurrently on a single GPU, to avoid under-utilization of its resources. Fine-grain sharing of streaming multiprocessors (SMs) allows thread blocks of multiple kernels to be assigned to GPU resources altogether. However, this may cause resource contention and performance degradation if both kernels try to access a shared resource at the same time. Detecting these interferences is essential especially in high-performance computing (HPC) systems, in which multiple applications may issue different kernels to available shared GPUs. This paper proposes a machine learning-based approach to characterize kernels and predict interference before their concurrent execution. Random forest classifier is used to classify interfering and noninterfering kernels. Experimental results show that the proposed method can detect interfering kernels with up to 91.7% accuracy.