TY - GEN

T1 - Comparative Efficiency Analysis of MPI Blocking and Non-blocking Communications with Coarray Fortran

AU - Reshetova, Galina

AU - Cheverda, Vladimir

AU - Koinov, Vitaly

N1 - Funding Information: Acknowledgements. Galina Reshetova and Vladimir Cheverda have been supported by the Russian Science Foundation, project 20-11-20112. Publisher Copyright: © 2021, Springer Nature Switzerland AG.

PY - 2021

Y1 - 2021

N2 - The MPI is the most widespread data exchange interface standard used in parallel programming for clusters and supercomputers with many computer platforms. The primary means of the MPI communication between processes is passing messages based on basic point-to-point blocking and non-blocking routines. The choice of the optimal implementation of exchanges is essential to minimize the idle and transmission times to achieve parallel algorithm efficiency. We used three realizations of data exchange processes based on blocking, non-blocking point-to-point MPI routines and new features of the Coarray Fortran technique to determine the most efficient parallelization strategy. For the study, the two-dimensional wave equation was used as a test problem. During the experiments, the problem size and the approaches to the data exchange for transferring data between processes were changed. For each version, we measured the computation time and the acceleration factor. The research carried out shows that the larger the problem size, the greater the benefits of delayed non-blocking routines and Coarray Fortran. The efficiency of delayed non-blocking operations is due to overlapping the data transfer in the computations background. The Coarray Fortran acceleration is achieved by using Coarray variables with shared memory. The Coarray approach starts to win with the growth of problem size.

AB - The MPI is the most widespread data exchange interface standard used in parallel programming for clusters and supercomputers with many computer platforms. The primary means of the MPI communication between processes is passing messages based on basic point-to-point blocking and non-blocking routines. The choice of the optimal implementation of exchanges is essential to minimize the idle and transmission times to achieve parallel algorithm efficiency. We used three realizations of data exchange processes based on blocking, non-blocking point-to-point MPI routines and new features of the Coarray Fortran technique to determine the most efficient parallelization strategy. For the study, the two-dimensional wave equation was used as a test problem. During the experiments, the problem size and the approaches to the data exchange for transferring data between processes were changed. For each version, we measured the computation time and the acceleration factor. The research carried out shows that the larger the problem size, the greater the benefits of delayed non-blocking routines and Coarray Fortran. The efficiency of delayed non-blocking operations is due to overlapping the data transfer in the computations background. The Coarray Fortran acceleration is achieved by using Coarray variables with shared memory. The Coarray approach starts to win with the growth of problem size.

KW - Acoustic wave equation

KW - Coarray Fortran

KW - Domain decomposition method

KW - Finite difference schemes

KW - MPI

KW - Parallel programming

UR - http://www.scopus.com/inward/record.url?scp=85123315141&partnerID=8YFLogxK

UR - https://www.mendeley.com/catalogue/dec692d1-4367-3e34-afad-6d1a032efb85/

U2 - 10.1007/978-3-030-92864-3_25

DO - 10.1007/978-3-030-92864-3_25

M3 - Conference contribution

AN - SCOPUS:85123315141

SN - 978-3-030-92863-6

T3 - Communications in Computer and Information Science

SP - 322

EP - 336

BT - Supercomputing - 7th Russian Supercomputing Days, RuSCDays 2021, Revised Selected Papers

A2 - Voevodin, Vladimir

A2 - Sobolev, Sergey

PB - Springer Science and Business Media Deutschland GmbH

T2 - 7th Russian Supercomputing Days Conference, RuSCDays 2021

Y2 - 27 September 2021 through 28 September 2021

ER -