Exploitation of parallelism for applications with an input data stream: optimal resource-throughput tradeoffs

Pipeline applications simultaneously execute different instances of an input data stream. The iterative behavior of these applications makes the mapping process more difficult than that applied to classical task-parallel or data-parallel applications. In this paper we propose a new mapping algorithm, called ROUTE (resource optimization under throughput requirements), for pipelined applications, which provides assignments for given throughput constraints and minimizes the number of processors. The effectiveness of ROUTE is compared with a previous mapping algorithm proposed in the literature, for a representative set of synthetic applications. In all cases, ROUTE provides mappings that effectively meet the input throughput constraints, and is able to improve resource optimization by decreasing the number of required processors.