Dynamic Overlay Networks for Image Processing Grids

Abstract

During the development and parametrization of 2D image-processing algorithms for surface inspection uses, you need to test a huge amount of image-data for each modification of the algorithms or parameters. For algorithm runtimes up to several seconds, this will take a long time. To speed up this process it is recommended to distribute the computation in a parallel computation environment. Compute Grids, which use the unused resources of existing hardware are the most cost efficient way to solve this problem. The most existing Grid-Concepts are based on flat connection structures with a scheduler on the top; for high job-rates the scheduler becomes the bottleneck of the whole system. Concepts to solve this problem organize the nodes in tree-structures to discharge the central scheduler. In heterogeneous Desktop-Grids where the different nodes are widely distributed the usually used random arrangement of the nodes in the tree-structure can be counterproductive, because the bandwidthes and latencies in a Grid can be varying. In this paper we will show a solution to arrange the nodes of the grid optimized by bandwidth and latency, using modified spanning-tree algorithms, so that the average response time is reduced and in result of this the job-throughput of the Compute-Grid is increased.