A Processor Workload Distribution Algorithm for Massively Parallel Applications

Abstract

Directed Acyclic Graph (DAG) is a standard model used to describe tasks that execute according to precedence constraints and that allows intra-task parallelism. This model is well suited to camera-based applications where multiple treatments must be executed in parallel according to the camera input, such applications found for example in self-driving cars or image recognition via convolutional neural network (CNN). Such applications are used on embedded systems and therefore require low energy cost and a limited hardware space. The main contribution of this paper is to present a new partitioning algorithm based on a DAG stretching technique. This stretching algorithm frees processor cores and thus implies energy savings and leads to new hardware design using a reduced number of processors. We present an experimental evaluation of this algorithm to show its efficiency.