Map-Reduce based tipping point scheduler for parallel image processing

Abstract

Nowadays, Big Data image processing is very much in need due to its proven success in the field of business information system, medical science and social media. However, as the days are passing by, the computation of Big Data images is becoming more complex which ultimately results in complex resource management and higher task execution time. Researchers have been using a combination of CPU and GPU based computing to cut down the execution time, however, when it comes to scaling of compute nodes, then the combination of CPU and GPU based computing still remains a challenge due to the high communication cost factor. In order to tackle this issue, the Map-Reduce framework has come out to be a viable option as its workflow optimization could be enhanced by changing its underlying job scheduling mechanism. This paper presents a comparative study of job scheduling algorithms which could be deployed over various Big Data based image processing application and also proposes a tipping point scheduling algorithm to optimize the workflow for job execution on multiple nodes. The evaluation of the proposed scheduling algorithm is done by implementing parallel image segmentation algorithm to detect lung tumor for up to 3GB size of image dataset. In terms of performance comprising of task execution time and throughput, the proposed tipping point scheduler has come out to be the best scheduler followed by the Map-Reduce based Fair scheduler. The proposed tipping point scheduler is 1.14 times better than Map-Reduce based Fair scheduler and 1.33 times better than Map-Reduced based FIFO scheduler in terms of task execution time and throughput. In terms of speedup comparison between single node and multiple nodes, the proposed tipping point scheduler attained a speedup of 4.5 X for multi-node architecture.