USO DE MODELOS PGAS EN LA SOLUCION DIRECTA DE SISTEMAS LINEALES DENSOS

SARRAF, SOFÍA; LOPEZ, EZEQUIEL; RIOS RODRIGUEZ, GUSTAVO ADOLFO; BATTAGLIA, LAURA; D'ELIA, JORGE

Bahía Blanca

Congreso; XXXVIII Congreso Argentino de Mecanica Computacional; 2022

Universidad Nacional del Sur (UNS), Avenida Colón 80, Bahía Blanca

Resumen. El espacio de direcciones global particionado (PGAS, por Partitioned Global Address Space) es un modelo de programación paralela desarrollado para computadoras con memoria compartida o distribuida, y es la base para los lenguajes de programación con comunicaciones unilaterales de alto nivelcomo UPC (Unified Parallel C) y Coarray Fortran (CAF), entre otros, así como también de bibliotecas especializadas. Todos estos recursos son extensiones para proporcionar una comunicación unilateral de alto nivel. En este trabajo se muestra el empleo de CAF utilizando el compilador gfortran a través deOpen Coarrays, y se lo aplica a la clásica factorización LU (Lower-Upper), sin pivoteo, de un sistema lineal denso regular, un tópico de interés en el método de elementos de borde (o BEM, por Boundary Element Method), utilizando memoria primaria compartida en computadores multi-núcleo, comparandorendimientos versus recursos clásicos como HPL (High-Performance Linpack Benchmark).Keywords: coarray fortran (CAF), High Performance Computing (HPC), one-sided communication, parallel computation, Partitioned Global Address Space (PGAS).Abstract. The Partitioned Global Address Space (PGAS) is a parallel programming model that has been developed for computers with shared or distributed memory, and is the basis for programming languages with unilateral high-level communications such as Unified Parallel C (UPC) and Coarray Fortran (CAF), among others, as well as other specialized libraries. All of these resources are extensions to provide one-sided high-level communication. In this work its use is shown through CAF utilising the gfortrancompiler with Open Coarrays, and it is applied to the classic LU (Lower-Upper) factorization, without pivoting, of a regular dense linear system, which is a topic of interest in the Boundary Element Method (BEM). The tests are performed using shared primary memory in multi-core computers and compared with the performance of classic resources such as the High-Performance Linpack Benchmark (HPL).