Response time analysis of parallel tasks on accelerator-based heterogeneous platforms

Abstract

Due to the inherent high parallelism and heterogeneity, accelerator-based heterogeneous platforms have been regarded as promising solutions for computation-intensive applications. With respect to latency and power consumption, accelerator often performs better than general-purpose processors. The directed acyclic graph (DAG) model is widely used to represent parallel applications, where each vertex represents an independent workload and each edge represents the precedence constraint between two vertices. The response time analysis of DAG tasks is of paramount importance to the judgment of task schedulability. However, the existing response time analysis method of DAG tasks on accelerator-based heterogeneous platforms just offloading one vertex to the accelerator cannot comply with the actual application scenarios. The well-known worst-case response time (WCRT) bounds of DAG tasks on heterogeneous platforms are pessimistic for judging the task schedulability. This paper studies the response time analysis method of DAG tasks that offload multiple vertices to the accelerator. We transform the graph structure of the DAG task to reduce the interference of vertices running on the accelerator to vertices on the general-purpose processor. Finally, we propose a new WCRT bound for the transformed DAG. The experimental results show that the new WCRT bound is more precise than the existing bounds of DAG tasks.