Evaluation of global synchronization for iterative algebra algorithms on many-core

Massively parallel computing is applied extensively in various scientific and engineering domains. With the growing interest in many-core architectures and due to the lack of explicit support for inter-block synchronization specifically in GPUs, synchronization becomes necessary to minimize inter-block communication time. In this paper, we have proposed two new inter-block synchronization techniques: 1) Relaxed Synchronization, and 2) Block-Query Synchronization. These schemes are used in implementing numerical iterative solvers where computation/communication overlapping is one used optimization to enhance application performance. We have evaluated and analyzed the performance of the proposed synchronization techniques using Jacobi Iterative Solver in comparison to the state of the art inter-block lock-free synchronization techniques. We have achieved about 1-8% performance improvement in terms of execution time over lock-free synchronization depending on the problem size and the number of thread blocks. We have also evaluated the proposed algorithm on GPU and MIC architectures and obtained about 8-26% performance improvement over the barrier synchronization available in OpenMP programming environment depending on the problem size and number of cores used.