Towards adaptive GPU resource management for embedded real-time systems

Abstract

In this paper, we present two conceptual frameworks for GPU applications to adjust their task execution times based on total workload. These frameworks enable smart GPU resource management when many applications share GPU resources while the workloads of those applications vary. Application developers can explicitly adjust the number of GPU cores depending on their needs. An implicit adjustment will be supported by a run-time framework, which dynamically allocates the number of cores to tasks based on the total workload. The runtime support of the proposed system can be realized using functions which measure the execution times of the tasks on GPU and change the number of GPU cores. We motivate the necessity of this framework in the context of self-driving technologies, and we believe that our frameworks for GPU programming are useful contributions given the increasing emphasis on parallel heterogeneous computing.