Modern graphics processing units (GPUs) with many-core architectures have emerged as general-purpose parallel computing platforms that can accelerate simulation science applications tremendously. While multi-GPU workstations with several TeraFLOPS of peak computing power are available to accelerate computational problems, larger problems require even more resources. Conventional clusters of central processing units (CPU) are now being augmented with multiple GPUs in each compute-node to tackle large problems. The heterogeneous architecture of a multi-GPU cluster with a deep memory hierarchy creates unique challenges in developing scalable and efficient simulation codes. In this study, we pursue mixed MPI-CUDA implementations and investigate three strategies to probe the efficiency and scalability of incompressible flow computations on the Lincoln Tesla cluster at the National Center for Supercomputing Applications (NCSA). We exploit some of the advanced features of MPI and CUDA programming to overlap both GPU data transfer and MPI communications with computations on the GPU. We sustain approximately 2.4 TeraFLOPS on the 64 nodes of the NCSA Lincoln Tesla cluster using 128 GPUs with a total of 30,720 processing elements. Our results demonstrate that multi-GPU clusters can substantially accelerate computational fluid dynamics (CFD) simulations.
We investigate multi-level parallelism on GPU clusters with MPI-CUDA and hybrid MPI-OpenMP-CUDA parallel implementations, in which all computations are done on the GPU using CUDA. We explore efficiency and scalability of incompressible flow computations using up to 256 GPUs on a problem with approximately 17.2 billion cells. Our work addresses some of the unique issues faced when merging fine-grain parallelism on the GPU using CUDA with coarse-grain parallelism that use either MPI or MPI-OpenMP for communications. We present three different strategies to overlap computations with communications, and systematically assess their impact on parallel performance on two different GPU clusters. Our results for strong and weak scaling analysis of incompressible flow computations demonstrate that GPU clusters offer significant benefits for large data sets, and a dual-level MPI-CUDA implementation with maximum overlapping of computation and communication provides substantial benefits in performance. We also find that our tri-level MPI-OpenMP-CUDA parallel implementation does not offer a significant advantage in performance over the dual-level implementation on GPU clusters with two GPUs per node, but on clusters with higher GPU counts per node or with different domain decomposition strategies a tri-level implementation may exhibit higher efficiency than a dual-level implementation and needs to be investigated further.
Numerical computations of incompressible flow equations with pressure-based algorithms necessitate the solution of an elliptic Poisson equation, for which multigrid methods are known to be very efficient. In our previous work we presented a dual-level (MPI-CUDA) parallel implementation of the Navier-Stokes equations to simulate buoyancy-driven incompressible fluid flows on GPU clusters with simple iterative methods while focusing on the scalability of the overall solver. In the present study we describe the implementation and performance of a multigrid method to solve the pressure Poisson equation within our MPI-CUDA parallel incompressible flow solver. Various design decisions and algorithmic choices for multigrid methods are explored in light of NVIDIA's recent Fermi architecture. We discuss how unique aspects of an MPI-CUDA implementation for GPU clusters is related to the software choices made to implement the multigrid method. We propose a new coarse grid solution method of embedded multigrid with amalgamation and show that the parallel implementation retains the numerical efficiency of the multigrid method. Performance measurements on the NCSA Lincoln and TACC Longhorn clusters are presented for up to 64 GPUs.
High performance computing using graphics processing units (GPUs) is gaining popularity in the scientific computing field, with many large compute clusters being augmented with multiple GPUs in each node. We investigate hybrid tri-level (MPI-OpenMP-CUDA) parallel implementations to explore the efficiency and scalability of incompressible flow computations on GPU clusters up to 128 GPUS. This work details some of the unique issues faced when merging fine-grain parallelism on the GPU using CUDA with coarse-grain parallelism using OpenMP for intra-node and MPI for inter-node communication. Comparisons between the tri-level MPI-OpenMP-CUDA and dual-level MPI-CUDA implementations are shown using computationally large computational fluid dynamics (CFD) simulations. Our results demonstrate that a tri-level parallel implementation does not provide a significant advantage in performance over the dual-level implementation, however further research is needed to justify our conclusion for a cluster with a high GPU per node density or when using software that can utilize OpenMP's fine-grain parallelism more effectively.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.
hi@scite.ai
10624 S. Eastern Ave., Ste. A-614
Henderson, NV 89052, USA
Copyright © 2024 scite LLC. All rights reserved.
Made with 💙 for researchers
Part of the Research Solutions Family.