ScholarWorks@숙명여자대학교

초록

GPU sharing technologies are being implemented to address the low resource utilization often seen in deep learning applications and to enhance performance in heterogeneous GPU clusters. However, there has been considerably less research on the effective use of time-slicing, Multi-Process Service (MPS), and Multi-Instance GPU (MIG) sharing technologies in relation to the specific characteristics of deep learning applications across different GPU architectures. This paper proposes optimal task placement criteria for these GPU sharing technologies by analyzing the execution scale, resource sensitivity, and resource utilization of tasks based on various deep learning models and parameters. In a heterogeneous GPU cluster environment that supports multiple sharing technologies, this study tested these optimal task placement criteria, resulting in reduced execution time and improved throughput.

키워드