Análise do Impacto de Runtimes Task-Based no Desempenho de Algoritmos de Álgebra Linear Densa
Resumo
Runtimes de escalonamento de tarefas dependentes como StarPU e PaRSEC abstraem a complexidade de plataformas heterogêneas, mas adotam estratégias diferentes de escalonamento e gerenciamento de dados. Este trabalho compara o desempenho dos dois runtimes nas fatorações de Cholesky e LU sem pivotamento da biblioteca Chameleon, mantendo fixa a biblioteca de álgebra linear para isolar o efeito do runtime. Este estudo exigiu estender a interface Dynamic Task Discovery (DTD) do PaRSEC e os codelets do Chameleon com suporte a GPU. Os experimentos mostram que o StarPU é superior ou equivalente na maior parte dos casos. O parsec:gd só obtém vantagem na máquina com RTX 4070 (455 GFLOPS em FP64, aproximadamente metade da capacidade da CPU). Na máquina com RTX 4090 (1.290 GFLOPS, quase o dobro da CPU), o StarPU vence em toda a faixa avaliada, com cerca de 30% de vantagem mesmo no tamanho de bloco ótimo de cada runtime.Referências
Agullo, E., Augonnet, C., Dongarra, J., Faverge, M., Langou, J., Ltaief, H., and Tomov, S. (2011). Lu factorization for accelerator-based systems. In 2011 9th IEEE/ACS International Conference on Computer Systems and Applications (AICCSA), pages 217–224.
Agullo, E., Augonnet, C., Dongarra, J., Ltaief, H., Namyst, R., Thibault, S., and Tomov, S. (2012). Chapter 34 - a hybridization methodology for high-performance linear algebra software for gpus. In mei W. Hwu, W., editor, GPU Computing Gems Jade Edition, Applications of GPU Computing Series, pages 473–484. Morgan Kaufmann, Boston.
Agullo, E., Beaumont, O., Eyraud-Dubois, L., Herrmann, J., Kumar, S., Marchal, L., and Thibault, S. (2015). Bridging the gap between performance and bounds of cholesky factorization on heterogeneous platforms. In 2015 IEEE International Parallel and Distributed Processing Symposium Workshop, pages 34–45.
Augonnet, C., Thibault, S., Namyst, R., and Wacrenier, P.-A. (2011). Starpu: a unified platform for task scheduling on heterogeneous multicore architectures. Concurrency and Computation: Practice and Experience, 23(2):187–198.
Bosilca, G., Bouteiller, A., Danalis, A., Herault, T., Lemariner, P., and Dongarra, J. (2011). DAGuE: a generic distributed DAG engine for high performance computing. In Proceedings of the Workshops of the 25th IEEE International Symposium on Parallel and Distributed Processing (IPDPS 2011 Workshops), pages 1151–1158, Anchorage, Alaska, USA. IEEE.
Buttari, A., Langou, J., Kurzak, J., and Dongarra, J. J. (2007). A class of parallel tiled linear algebra algorithms for multicore architectures. CoRR, abs/0709.1272.
Garcia Pinto, V., Mello Schnorr, L., Stanisic, L., Legrand, A., Thibault, S., and Danjean, V. (2018). A visual performance analysis framework for task-based parallel applications running on hybrid clusters. Concurrency and Computation: Practice and Experience, 30(18):e4472.
Hoque, R., Herault, T., Bosilca, G., and Dongarra, J. (2017). Dynamic task discovery in parsec: a data-flow task-based runtime. In Proceedings of the 8th Workshop on Latest Advances in Scalable Algorithms for Large-Scale Systems, ScalA ’17, New York, NY, USA. Association for Computing Machinery.
Lacoste, X., Faverge, M., Bosilca, G., Ramet, P., and Thibault, S. (2014). Taking advantage of hybrid systems for sparse direct solvers via task-based runtimes. In 2014 IEEE International Parallel & Distributed Processing Symposium Workshops (IPDPSW), pages 29–38.
Lisito, A., Faverge, M., Kuhn, M., Pruvost, F., and Ramet, P. (2025). Scalable and portable lu factorization with partial pivoting on top of runtime systems. In 2025 IEEE International Parallel and Distributed Processing Symposium (IPDPS), pages 1011–1022.
Pei, Y., Bosilca, G., and Dongarra, J. (2022). Sequential task flow runtime model improvements and limitations. In 2022 IEEE/ACM International Workshop on Runtime and Operating Systems for Supercomputers (ROSS), pages 1–8.
Pinto, V. G., Nesi, L. L., Miletto, M. C., and Schnorr, L. M. (2021). Providing indepth performance analysis for heterogeneous task-based applications with starvz. 2021 IEEE International Parallel and Distributed Processing Symposium Workshops (IPDPSW), pages 16–25.
Thibault, S. (2018). On Runtime Systems for Task-based Programming on Heterogeneous Platforms. Habilitation à diriger des recherches (hdr), Université de Bordeaux. HAL Id: tel-01959127.
Agullo, E., Augonnet, C., Dongarra, J., Ltaief, H., Namyst, R., Thibault, S., and Tomov, S. (2012). Chapter 34 - a hybridization methodology for high-performance linear algebra software for gpus. In mei W. Hwu, W., editor, GPU Computing Gems Jade Edition, Applications of GPU Computing Series, pages 473–484. Morgan Kaufmann, Boston.
Agullo, E., Beaumont, O., Eyraud-Dubois, L., Herrmann, J., Kumar, S., Marchal, L., and Thibault, S. (2015). Bridging the gap between performance and bounds of cholesky factorization on heterogeneous platforms. In 2015 IEEE International Parallel and Distributed Processing Symposium Workshop, pages 34–45.
Augonnet, C., Thibault, S., Namyst, R., and Wacrenier, P.-A. (2011). Starpu: a unified platform for task scheduling on heterogeneous multicore architectures. Concurrency and Computation: Practice and Experience, 23(2):187–198.
Bosilca, G., Bouteiller, A., Danalis, A., Herault, T., Lemariner, P., and Dongarra, J. (2011). DAGuE: a generic distributed DAG engine for high performance computing. In Proceedings of the Workshops of the 25th IEEE International Symposium on Parallel and Distributed Processing (IPDPS 2011 Workshops), pages 1151–1158, Anchorage, Alaska, USA. IEEE.
Buttari, A., Langou, J., Kurzak, J., and Dongarra, J. J. (2007). A class of parallel tiled linear algebra algorithms for multicore architectures. CoRR, abs/0709.1272.
Garcia Pinto, V., Mello Schnorr, L., Stanisic, L., Legrand, A., Thibault, S., and Danjean, V. (2018). A visual performance analysis framework for task-based parallel applications running on hybrid clusters. Concurrency and Computation: Practice and Experience, 30(18):e4472.
Hoque, R., Herault, T., Bosilca, G., and Dongarra, J. (2017). Dynamic task discovery in parsec: a data-flow task-based runtime. In Proceedings of the 8th Workshop on Latest Advances in Scalable Algorithms for Large-Scale Systems, ScalA ’17, New York, NY, USA. Association for Computing Machinery.
Lacoste, X., Faverge, M., Bosilca, G., Ramet, P., and Thibault, S. (2014). Taking advantage of hybrid systems for sparse direct solvers via task-based runtimes. In 2014 IEEE International Parallel & Distributed Processing Symposium Workshops (IPDPSW), pages 29–38.
Lisito, A., Faverge, M., Kuhn, M., Pruvost, F., and Ramet, P. (2025). Scalable and portable lu factorization with partial pivoting on top of runtime systems. In 2025 IEEE International Parallel and Distributed Processing Symposium (IPDPS), pages 1011–1022.
Pei, Y., Bosilca, G., and Dongarra, J. (2022). Sequential task flow runtime model improvements and limitations. In 2022 IEEE/ACM International Workshop on Runtime and Operating Systems for Supercomputers (ROSS), pages 1–8.
Pinto, V. G., Nesi, L. L., Miletto, M. C., and Schnorr, L. M. (2021). Providing indepth performance analysis for heterogeneous task-based applications with starvz. 2021 IEEE International Parallel and Distributed Processing Symposium Workshops (IPDPSW), pages 16–25.
Thibault, S. (2018). On Runtime Systems for Task-based Programming on Heterogeneous Platforms. Habilitation à diriger des recherches (hdr), Université de Bordeaux. HAL Id: tel-01959127.
Publicado
03/11/2026
Como Citar
TREGNAGO, Matheus Augusto; NAVAUX, Philippe O. A.; SCHNORR, Lucas Mello.
Análise do Impacto de Runtimes Task-Based no Desempenho de Algoritmos de Álgebra Linear Densa. In: SIMPÓSIO EM SISTEMAS COMPUTACIONAIS DE ALTO DESEMPENHO (SSCAD), 27. , 2026, Natal/RN.
Anais [...].
Porto Alegre: Sociedade Brasileira de Computação,
2026
.
p. 99-111.
DOI: https://doi.org/10.5753/sscad.2026.34369.
