Fine-Tuning an OpenMP-Based TVD–Hopmoc Method Using Intel® Parallel Studio XE Tools on Intel® Xeon® Architectures

Otimização do Método HOPMOC 1D com auxílio das ferramentas Intel Parallel Studio

“…Este trabalho difere de [Cabral et al 2018a]à medida que realiza um estudo comparativo não apenas entre valores de speedup mas também de Spin time e CPU time, além de incluir nesse estudo uma implementação MPI do método. O trabalho [Cabral et al 2018b] estudou o desempenho da metodologia apresentada no trabalho anterior para uma versão TVD-HOPMOC e para diversas arquiteturas de memória compartilhada Intel Xeon. O trabalho de agora compara o desempenho do método HOPMOC em um ambiente de memória compartilhada Xeon para uma versão em MPI em relação a versão OpenMP apresentada no trabalho [Cabral et al 2018a] chamada de HOPMOC EWS.…”

An evaluation of MPI and OpenMP paradigms in finite‐difference explicit methods for PDEs on shared‐memory multi‐ and manycore systems

“…OpenMP is also efficient for sharing data between threads because this standard can access shared memory directly without the need for communicating explicitly and copying data on the same node. However, publications report that achieving high efficiency with OpenMP when using a large number of cores and threads is not trivial . Additionally, it is not clear if all applications and hardware platforms can effectively benefit from an MPI/OpenMP approach (see the work of Bassi et al and references therein).…”

“…Previous work has found that a naïve OpenMP‐based implementation of an explicit method for solving partial differential equations (PDEs) is inefficient . Thus, a previous publication proposed an OpenMP‐based one‐dimensional (1D) Hopmoc method in tandem with total variation diminishing (TVD) for numerical time integration of evolutionary differential equations.…”

“…Additionally, we investigated a coarse‐grained implementation that employs an explicit synchronization mechanism capable of providing high speedups in an OpenMP‐based TVD‐Hopmoc method, even in simulations involving a small number of time steps. As a result, Cabral et al proposed a scheme consisting of synchronizing only adjacent threads. The approach employs an explicit work‐sharing (EWS) strategy to reduce the OpenMP scheduling time.…”

“…Thus, we investigated the selection of the best paradigm for each specific application running on multicore or manycore platforms. Specifically, this paper evaluates the use of pure MPI, the OMP‐EWS approach, and hybrid MPI/OMP‐EWS approaches with scatter and compact thread binding policies aiming at improving the parallel efficiency of three explicit finite‐difference (FD) methods for solving PDEs running on Intel ® Many Integrated Core and Xeon ® Scalable Processor shared‐memory architectures. We apply the approaches to Laplace's, heat conduction, and advection‐diffusion equations.…”