Open Access

ARTICLE

Comparison and Performance Analysis of Multiple CPU/GPU Computing Systems – Resin Infusion Flow Modeling Application

R.H. Haney¹, R.V. Mohan²

1 North Carolina A&T State University, Greensboro, NC, U.S.A.
2 North Carolina A&T State University, Greensboro, NC, U.S.A, For Correspondence.

Computer Modeling in Engineering & Sciences 2013, 95(5), 431-452. https://doi.org/10.3970/cmes.2013.095.431

Abstract

The use of Graphics Processing Units (GPUs) as co-processors for single CPU/GPU computing systems has become pronounced in high performance computing research, however the solution of truly large scale computationally intensive problems require the utilization of multiple computing nodes. Multiple CPU/GPU computing systems bring new complexities to the observed performance of computationally intensive applications, the more salient of which is the cost of local CPU-GPU host and intra-nodal communication. This paper compares and analyzes the performance of a computationally intensive application represented by resin infusion flow during liquid composite molding process for the manufacture of structural composites application via two distinct multiple CPU/GPU computing system architectures. Resin flow infusion modeling during liquid composite molding process is the engineering application of interest in the present study. The global domain is partitioned into a series of sub-domains each of which is solved at the local host and reassembled for the final solution as per the domain decomposition methodology. The candidate application, as with many scientific and engineering applications, uses the Finite Element Method (FEM) to computationally model the governing physics based mass and momentum conversation equations. FEM discretization results in large sparse linear equation systems that are solved iteratively for this class of free surface, moving boundary value problem. Computational analysis software for the GPU environment has been developed using CUDA API for the iterative linear equation system solver based on the preconditioned conjugate gradient method for the solution of linear system of equations. These linear equation systems are solved multiple times with intra-nodal communication utilized, resulting in the converged global system. The interplay of local host CPU/GPU and intra-nodal communication creates mixed performance results for the presented candidate application. The software/hardware factors that affect performance for each architecture are examined and discussed in this paper-understanding how the presented candidate application’s observed performance is effected by both the individual multiple CPU/GPU computing system architecture and algorithmic/software design is critical to optimize many modern high performance applications which employee the GPU as a hardware accelerator.

---

Keywords

---