Analytical Communication Performance Models as a metric in the partitioning of data-parallel kernels on heterogeneous platforms

Data partitioning on heterogeneous HPC platforms is formulated as an optimization problem. The algorithm departs from the communication performance models of the processes representing their speeds and outputs a data tiling that minimizes the communication cost. Traditionally, communication volume i...

ver descrição completa

Detalhes bibliográficos
Autores: Rico Gallego, Juan Antonio, Díaz Martín, Juan Carlos, Calvo Jurado, Carmen, Moreno Álvarez, Sergio, García Zapata, Juan Luis
Formato: artículo
Fecha de publicación:2019
País:España
Recursos:Universidad Nacional de Educación a Distancia
Repositorio:e-spacio. Repositorio Institucional de la UNED
Idioma:inglés
OAI Identifier:oai:e-spacio.uned.es:20.500.14468/24376
Acesso em linha:https://hdl.handle.net/20.500.14468/24376
Access Level:acceso abierto
Palavra-chave:12 Matemáticas::1203 Ciencia de los ordenadores ::1203.17 Informática
partitioning algorithms
communication performance models
communication optimization
hybrid data-parallel kernels
Descrição
Resumo:Data partitioning on heterogeneous HPC platforms is formulated as an optimization problem. The algorithm departs from the communication performance models of the processes representing their speeds and outputs a data tiling that minimizes the communication cost. Traditionally, communication volume is the metric used to guide the partitioning, but such metric is unable to capture the complexities introduced by uneven communication channels and the variety of patterns in the kernel communications. We discuss Analytical Communication Performance Models as a new metric in partitioning algorithms. They have not been considered in the past because of two reasons: prediction inaccuracy and lack of tools to automatically build and solve kernel communication formal expressions. We show how communication performance models fit the specific kernel and platform, and we present results that equal or even improve previous volume-based strategies.