Scheduling of Signal Processing Tasks in a Computer Cluster

Abstract

Newer generations of radar signal processing systems have increasingly higher computational demands. This thesis aimed to investigate the impact of scheduling techniques, load balancing approaches, and parallel programming models in real-time signal processing applications utilizing a cluster environment. To do so, a representative scenario was created, which was intended to resemble a real application scenario. Several scheduling algorithms were implemented in an iterative manner and evaluated in the representative scenario using a homogenous cluster system consisting of four nodes. Some key findings involved the potential of dynamically varying the number of workers and their resources to better adapt to the dynamic environment of radar signal processing. These techniques could also reduce contention for memory resources and the negative impacts of simultaneous multithreading for execution times. By allocating sparingly used backup workers that ran tasks at an easier difficulty, additional increases in overall performance and robustness could be established. The results indicate that a scheduler implemented in cluster-oriented programming models can utilize the system resources to meet the increased performance demands of signal processing systems. However, challenges such as development overhead, process allocation, and adaptation to cluster architecture must be considered for optimal performance in an arbitrary cluster environment.