Performance vs. accuracy trade-offs for large-scale image analysis applications

In many data analysis applications, application-level parameters influence the execution time of the data analysis method or program. Some of these parameters also affect the accuracy of output of the analysis. In this work, we investigate execution strategies for adaptive data analysis applications where the user is willing to trade-off accuracy of output for performance gain and vice-versa. In order to meet the user defined quality of service requirements, the system must dynamically select values for the parameters during execution. We propose algorithms for adaptive processing of image tiles at different resolutions so that user defined requirements in terms of accuracy of the result and execution time constraints can be satisfied. We develop heuristics for estimation of accuracy vs performance characteristics of image tiles and for scheduling of the tiles for processing. We implement a demand-driven strategy for parallel execution of these heuristics on a parallel machine. We evaluate our approach for analysis of large images from digitized microscopy scanners.