| Title |
Fairness Analysis of Multi-tenant Applicationson Multi-Instance GPUs |
| Authors |
이제인(Jane Rhee) ; 윤명국(Myung Kuk Yoon) |
| DOI |
https://doi.org/10.5573/ieie.2023.60.4.11 |
| Keywords |
Multi-Instance GPU; Multi-tenant; Fairness; Slice |
| Abstract |
Recently Multi-Instance Graphics Processing Units (GPUs) have been widely used in multi-tenant cloud computing environments, where multiple concurrent applications are executed on a single GPU, sharing limited resources. Nevertheless, studies lack in the sphere of fairness between applications executing on Multi-Instance GPUs. This paper conducts a detailed analysis of the fairness of concurrently executing applications on Multi-Instance GPUs. First, we analyze the performance of an application according to the number of streaming multiprocessor slices and memory slices used in Multi-Instance GPUs. Then, based on the analyzed performance patterns, we measure the fairness of applications and reveal that the highest fairness is guaranteed when slices are divided evenly, or in an asymmetric form considering the performance saturation point. In summary, this study makes three major contributions. (i) We define the three types of applications classified for the performance patterns with the increase of the number of slices used. (ii) We present an algorithm of the case with the highest fairness when intra-type applications are executed concurrently. (iii) We also present an algorithm for the highest fairness in those situations where inter-type applications are executed concurrently. |