Using machine learning techniques to analyze the performance of concurrent kernel execution on GPUs

Abstract

Heterogeneous systems employing CPUs and GPUs are becoming increasingly popular in large-scale data centers and cloud environments. In these platforms, sharing a GPU across different applications is an important feature to improve hardware utilization and system throughput. However, under scenarios where GPUs are competitively shared, some challenges arise. The decision on the simultaneous execution of different kernels is made by the hardware and depends on the kernels resource requirements. Besides that, it is very difficult to understand all the hardware variables involved in the simultaneous execution decisions in order to describe a formal allocation method. In this work, we use machine learning techniques to understand how the resource requirements of the kernels from the most important GPU benchmarks impact their concurrent execution. We focus on making the machine learning algorithms capture the hidden patterns that make a kernel interfere in the execution of another one when they are submitted to run at the same time. The techniques analyzed were k-NN, Logistic Regression, Multilayer Perceptron and XGBoost (which obtained the best results) over the GPU benchmark suites, Rodinia, Parboil and SHOC. Our results showed that, from the features selected in the analysis, the number of blocks per grid, number of threads per block, and number of registers are the resource consumption features that most affect the performance of the concurrent execution.