Busca avançada
Ano de início
Entree


How to Efficiently Parallelize Irregular DOACROSS Loops Using Fine Granularity and OpenMP Tasks: The SPEC mcf Case

Texto completo
Autor(es):
Salamanca, Juan ; Baldassin, Alexandro
Número total de Autores: 2
Tipo de documento: Artigo Científico
Fonte: OPENMP: ADVANCED TASK-BASED, DEVICE AND COMPILER PROGRAMMING, IWOMP 2023; v. 14114, p. 16-pg., 2023-01-01.
Resumo

There are certain loops that are considered hard to parallelize. Examples of this type of loops are those that have loop-carried dependencies (DOACROSS loops) and that are also irregular, that is, the dependencies between iterations vary depending on the context. Many techniques have been studied before to be able to parallelize this type of loops, however in OpenMP standard there is no efficient way to parallelize them. From the literature, it is known that many of these loops can be efficiently parallelized using fine-grained techniques (identifying strongly connected components). On the other hand, the most efficient way to parallelize this type of loops using OpenMP tasks has not been explored. Thus, this paper discusses the various forms of parallelization of this type of loops using SPEC 429.mcf as a case study; particularly, how to parallelize mcf using fine granularity in tasks. For that, this paper proposes new constructs (ste_for and ste) and speculative dependency-types (spec_in, spec_out, and spec_inout). An initial evaluation using different implementations to parallelize the mcf hottest loop shows that it is possible to achieve speed-ups of up to 2.44x with respect to the task-depend version using Speculative Task Execution. (AU)

Processo FAPESP: 18/15519-5 - Otimizações de desempenho para arquiteturas multicore
Beneficiário:Alexandro José Baldassin
Modalidade de apoio: Auxílio à Pesquisa - Jovens Pesquisadores - Fase 2
Processo FAPESP: 18/07446-8 - Integrando Execução Especulativa à Paralelização de Laços e Tarefas
Beneficiário:Juan Jesús Salamanca Guillén
Modalidade de apoio: Bolsas no Brasil - Pós-Doutorado
Processo FAPESP: 20/01665-0 - Integrando execução especulativa à paralelização de tarefas
Beneficiário:Juan Jesús Salamanca Guillén
Modalidade de apoio: Bolsas no Exterior - Estágio de Pesquisa - Pós-Doutorado