Synchronization-based GPU Access Control

Wednesday March 7, 2018
Location: HH D-level Conference Room
Time: 4:30PM

Abstract

GPUs and other shared hardware accelerators are being increasingly used in computationally-intensive real-time and multimedia systems. Access control to the shared resource is a key aspect of system efficiency. Multiprocessor synchronization protocols such as the Multiprocessor Priority Ceiling Protocol (MPCP) and the FIFO-based Multiprocessor Locking Protocol (FMLP) have been found to be reasonably feasible in practice. However, the existing synchronization protocols assume that tasks may not suspend during execution on shared external resources such as GPU. In such cases, busy-waiting for the entire critical-section duration significantly compromises CPU utilization. To avoid this problem, it is becoming useful in practice to extend synchronization protocols to allow suspensions within critical sections. In this paper, we extend MPCP to this particular context. We examine the feasibility and benefits of suspension-based MPCP over the original busy-waiting approach by presenting a case study on an NVIDIA TX2 embedded platform using real-world vision applications. Experimental results indicate that suspension-based MPCP can offer significantly lower response-times than the busy-waiting approach in practice, especially for lower-priority tasks, as it allows other tasks to use the CPU during critical section execution.

Bio

Iljoo Baek, the 2nd year PhD student, ECE Department Advisor : Prof. Rajkumar