The hydrodynamic approach to modeling astrophysics problems has several disadvantages in terms of the implementation of a parallel computing code. One of the main drawbacks is the low arithmetic intensity of the methods that implement the computational problem. This peculiarity produces the performance limitation associated with the performance limitations of the DRAM memory of high-performance computing systems. One of the solutions to this problem is data structuring based on the characteristics of processors and memory of a computer system on which supercomputer simulation is to be carried out. In this work, the authors use the specialized Intel SDLT library, which allows you to organize data in a special way that can help the compiler to vectorize a computational code for Intel server processors. The use of this library made it possible to speed up the computational code by fifty times, and for the first time bring the performance of some code functions to the performance limits of server processors on vector FMA instructions.
Язык оригиналаАнглийский
Название основной публикацииLecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Подзаголовок основной публикацииbook
РедакторыV. Voevodin, S. Sobolev, M. Yakobovskiy, R. Shagaliev
ГлаваChapter 21
Число страниц14
Том14388 LNCS
ISBN (печатное издание)978-303149431-4
СостояниеОпубликовано - 5 янв. 2024

Серия публикаций

ISSN (печатное издание)0302-9743
ISSN (электронное издание)1611-3349

    Предметные области ASJC Scopus

  • Компьютерные науки в целом
  • Theoretical Computer Science

ID: 51655911