A Server-based Approach for Predictable GPU Access with Improved Analysis

摘要

We propose a server-based approach to manage a general-purpose graphicsprocessing unit (GPU) in a predictable and efficient manner. Our proposedapproach introduces a GPU server that is a dedicated task to handle GPUrequests from other tasks on their behalf. The GPU server ensures bounded timeto access the GPU, and allows other tasks to suspend during their GPUcomputation to save CPU cycles. By doing so, we address the two majorlimitations of the existing real-time synchronization-based GPU managementapproach: busy waiting and long priority inversion. We have implemented aprototype of the server-based approach on a real embedded platform. This casestudy demonstrates the practicality and effectiveness of the server-basedapproach. Experimental results indicate that the server-based approach yieldssignificant improvements in task schedulability over the existingsynchronization-based approach in most practical settings. Although we focus ona GPU in this paper, the server-based approach can also be used for other typesof computational accelerators.