@MastersThesis{Souza:2018:AvTéAv,
author = "Souza, Carlos Renato de",
title = "Avalia{\c{c}}{\~a}o de t{\'e}cnicas avan{\c{c}}adas de
comunica{\c{c}}{\~a}o MPI na execu{\c{c}}{\~a}o do modelo
BRAMS",
school = "Instituto Nacional de Pesquisas Espaciais (INPE)",
year = "2018",
address = "S{\~a}o Jos{\'e} dos Campos",
month = "2018-05-28",
keywords = "BRMAS, shared memory, modelo num{\'e}rico de previs{\~a}o, MPI,
numerical model forecast.",
abstract = "O Centro de Previs{\~a}o de Tempo e Estudos Clim{\'a}ticos
(CPTEC) do Instituto Nacional de Pesquisas Espaciais (INPE)
executa operacionalmente modelos num{\'e}ricos de previs{\~a}o
de tempo e clima. Em particular, utiliza o modelo regional BRAMS
(Brazilian developments on the Regional Atmospheric Modeling
System). Esse modelo foi paralelizado com a biblioteca de
comunica{\c{c}}{\~a}o por troca de mensagens Message Passing
Interface (MPI) e {\'e} executado no supercomputador Tup{\~a} do
CPTEC, o qual {\'e} composto de mais de um milhar de n{\'o}s de
processamento. Cada n{\'o} possui dois processadores
multin{\'u}cleos numa arquitetura de mem{\'o}ria compartilhada.
Na execu{\c{c}}{\~a}o paralela do modelo BRAMS, seu
dom{\'{\i}}nio de c{\'a}lculo {\'e} dividido em
subdom{\'{\i}}nios entre processos executados nos n{\'u}cleos
dos muitos n{\'o}s. Eventuais depend{\^e}ncias de dados entre
subdom{\'{\i}}nios diferentes implicam na troca de mensagens MPI
entre processos sejam eles do mesmo n{\'o} ou n{\~a}o. O BRAMS
utiliza troca de mensagens MPI bilateral no modo
ass{\'{\i}}ncrono e sem bloqueio, dispon{\'{\i}}vel desde a
primeira vers{\~a}o do padr{\~a}o MPI. O padr{\~a}o MPI tem
evolu{\'{\i}}do, oferecendo novas t{\'e}cnicas para otimizar a
comunica{\c{c}}{\~a}o entre processos. Assim, visando otimizar o
desempenho da comunica{\c{c}}{\~a}o, o padr{\~a}o MPI-2
introduziu a chamada comunica{\c{c}}{\~a}o unilateral por acesso
remoto {\`a} mem{\'o}ria, que permite a um processo pode fazer
leituras ou escritas por meio de fun{\c{c}}{\~o}es MPI na
mem{\'o}ria de outro, seja do mesmo n{\'o} ou n{\~a}o,
permitindo a troca de dados entre processos, sem que o processo
alvo participe da comunica{\c{c}}{\~a}o explicitamente. A
comunica{\c{c}}{\~a}o unilateral foi aperfei{\c{c}}oada no
MPI-3, mas uma nova funcionalidade foi introduzida, a
comunica{\c{c}}{\~a}o unilateral por mem{\'o}ria compartilhada,
que permite a processos MPI executados num mesmo n{\'o} definirem
uma janela comum de mem{\'o}ria local e efetuar leituras e
escritas diretas na {\'a}rea da janela de outros processos
locais. Este trabalho visa avaliar o desempenho da
execu{\c{c}}{\~a}o paralela do modelo regional BRAMS ao utilizar
a comunica{\c{c}}{\~a}o unilateral de mem{\'o}ria compartilhada
na comunica{\c{c}}{\~a}o intra-n{\'o} e mantendo a
comunica{\c{c}}{\~a}o bilateral ass{\'{\i}}ncrona e sem
bloqueio na comunica{\c{c}}{\~a}o inter-n{\'o} e preservando a
mesma divis{\~a}o de dom{\'{\i}}nio de sua vers{\~a}o paralela
original. ABSTRACT: The Center for Weather Forecasts and Climate
Studies (CPTEC) of the Brazilian National Institute for Space
Research (INPE) executes several climate and weather numerical
forecast models on an operational basis, specifically using the
regional model nominated BRAMS (Brazilian developments on the
Regional Atmospheric Modeling System). This model was parallelized
using the Message Passing Interface (MPI) communication library,
being executed by the CPTECs {"}Tup{\~a}{"}, a supercomputer
composed of hundreds of processing nodes. Each node has two
multi-core processors in a shared memory architecture. In the
paralell execution of BRAMS, its calculation domain is divided
among processes executed in the cores of the many nodes. Data
dependencies between different subdomains require the exchange of
MPI messages between the processes, either intra-node or
inter-node. BRAMS employs asynchronous non-blocking point-to-point
communication, available since the first version of the MPI
standard. The MPI standard has evolved through the years bringing
new techniques to optimize the communication between processes.
Thus MPI-2 introduced the one-sided communication by remote memory
access. It allows a process to execute reads or writes using MPI
functions to the memory of other process, either in the same node
or not, exchanging data betweem processes without the explicit
cooperation of the target process. One-sided communication was
improved in the MPI-3 standard, but a new technique was added, the
shared memory one-sided communication. MPI processes executed in
the same computational node may define a common shared memory
window and execute direct reads and writes in the window part of
another process. The purpose of this work is to evaluate the
parallel performance of the BRAMS model using the shared memory
one-sided communication for the intra-node communication while
keeping the asynchronous non-blocking point-topoint inter-node
communication, and preserving the domain decomposition of its
original parallel version.",
committee = "Mendes, Celso Luiz (presidente) and Stephany, Stephan (orientador)
and Panetta, Jairo (orientador) and Campos Velho, Haroldo Fraga de
and Fazenda, {\'A}lvaro Luiz",
englishtitle = "Evaluation of advanced techniques for MPI communication in the
execution of the BRAMS model",
language = "pt",
pages = "108",
ibi = "8JMKD3MGP3W34R/3R6QKR8",
url = "http://urlib.net/ibi/8JMKD3MGP3W34R/3R6QKR8",
targetfile = "publicacao.pdf",
urlaccessdate = "15 jun. 2024"
}