Adaptive relaxed synchronization through the use of supervised learning methods

Abstract Several authors have proposed the use of relaxed synchronization to speed up the execution of parallel applications that admit tradeoffs between quality and execution time. However, most of these works propose the complete removal of synchronization primitives and do not anticipate the quality of the results to be obtained with different input data. In this paper, we propose a novel strategy for relaxing synchronization, evaluating the feasibility of using supervised learning methods to ensure that the relaxed synchronization technique provides results within acceptable limits of error. We use a varied set of program inputs to create a control base, providing data for the training of supervised learning methods. When the user wishes to execute his/her application with new input data (in the same execution environment), the trained classification algorithm will suggest the relax factor that is best suited for the triple application/input/execution environment. Using this methodology, we obtained a gain of 3.5x for the K-means algorithm applied to videos while maintaining the desired quality rate.