Concurrent kernel execution and interference analysis on GPUs using deep learning approaches

Abstract

Graphic processing units (GPUs) are becoming gradually popular in large-scale data centers and cloud computing. Sharing a GPU across different applications is an important feature in these platforms to improve resource utilization and system productivity. However, in situations where GPUs are competitively shared, some challenges arise. In this paper, concurrent kernel execution and interference detection are investigated using Extreme Gradient Boosting, Convolutional Neural Network, Gated Recurrent Unit, and Conditional Generative Adversarial Network learning techniques on Tesla P100 and RTX-2080 GPU architectures. The experimental results obtained by applying the deep learning models on the dataset, which consists of four GPU resources: blocks per grid, threads per block, number of registers, and shared memory, show that the Extreme Gradient Boosting and Multi-Channel Convolutional Neural Network models have the promising capability of concurrent kernel execution classification and interference detection. In comparison to the existing work, the proposed models investigated using real and synthesized datasets show clear outperformance and generalization capability in terms of recall and precision.