An OpenMP Implementation of the TVD-Hopmoc Method Based on a Synchronization Mechanism Using Locks Between Adjacent Threads on Xeon Phi (TM) Accelerators

摘要

This work focuses on the study of the 1-D TVD-Hopmoc method executed in shared memory manycore environments. In particular, this paper studies barrier costs on Intel® Xeon Phi™ (KNC and KNL) accelerators when using the OpenMP standard. This paper employs an explicit synchronization mechanism to reduce spin and thread scheduling times in an OpenMP implementation of the 1-D TVD-Hopmoc method. Basically, we define an array that represents threads and the new scheme consists of synchronizing only adjacent threads. Moreover, the new approach reduces the OpenMP scheduling time by employing an explicit work-sharing strategy. In the beginning of the process, the array that represents the computational mesh of the numerical method is partitioned among threads, instead of permitting the OpenMP API to perform this task. Thereby, the new scheme diminishes the OpenMP spin time by avoiding OpenMP barriers using an explicit synchronization mechanism where a thread only waits for its two adjacent threads. The results of the new approach is compared with a basic parallel implementation of the 1-D TVD-Hopmoc method. Specifically, numerical simulations shows that the new approach achieves promising performance gains in shared memory manycore environments for an OpenMP implementation of the 1-D TVD-Hopmoc method.