How Fast Can Graph Computations Go on Fine-grained Parallel Architectures

Fuente: arXiv
Guardado en:
Detalles Bibliográficos
Autores principales: Wang, Yuqing, Colley, Charles, Wheatman, Brian, Su, Jiya, Gleich, David F., Chien, Andrew A.
Formato: Preprint
Publicado: 2025
Materias:
Acceso en línea:
Etiquetas: Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
_version_ 1866909670454591488
author Wang, Yuqing
Colley, Charles
Wheatman, Brian
Su, Jiya
Gleich, David F.
Chien, Andrew A.
author_facet Wang, Yuqing
Colley, Charles
Wheatman, Brian
Su, Jiya
Gleich, David F.
Chien, Andrew A.
contents Large-scale graph problems are of critical and growing importance and historically parallel architectures have provided little support. In the spirit of co-design, we explore the question, How fast can graph computing go on a fine-grained architecture? We explore the possibilities of an architecture optimized for fine-grained parallelism, natural programming, and the irregularity and skew found in real-world graphs. Using two graph benchmarks, PageRank (PR) and Breadth-First Search (BFS), we evaluate a Fine-Grained Graph architecture, UpDown, to explore what performance codesign can achieve. To demonstrate programmability, we wrote five variants of these algorithms. Simulations of up to 256 nodes (524,288 lanes) and projections to 16,384 nodes (33M lanes) show the UpDown system can achieve 637K GTEPS PR and 989K GTEPS BFS on RMAT, exceeding the best prior results by 5x and 100x respectively.
format Preprint
id arxiv_https___arxiv_org_abs_2507_00949
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle How Fast Can Graph Computations Go on Fine-grained Parallel Architectures
Wang, Yuqing
Colley, Charles
Wheatman, Brian
Su, Jiya
Gleich, David F.
Chien, Andrew A.
Distributed, Parallel, and Cluster Computing
Hardware Architecture
Large-scale graph problems are of critical and growing importance and historically parallel architectures have provided little support. In the spirit of co-design, we explore the question, How fast can graph computing go on a fine-grained architecture? We explore the possibilities of an architecture optimized for fine-grained parallelism, natural programming, and the irregularity and skew found in real-world graphs. Using two graph benchmarks, PageRank (PR) and Breadth-First Search (BFS), we evaluate a Fine-Grained Graph architecture, UpDown, to explore what performance codesign can achieve. To demonstrate programmability, we wrote five variants of these algorithms. Simulations of up to 256 nodes (524,288 lanes) and projections to 16,384 nodes (33M lanes) show the UpDown system can achieve 637K GTEPS PR and 989K GTEPS BFS on RMAT, exceeding the best prior results by 5x and 100x respectively.
title How Fast Can Graph Computations Go on Fine-grained Parallel Architectures
topic Distributed, Parallel, and Cluster Computing
Hardware Architecture
url https://arxiv.org/abs/2507.00949