Start Date

10-17-2017 10:00 AM

Description

We present the performance gains of an openMP implementation of a fully adaptive nonlinear full multigrid (FMG) algorithm to simulate three-dimensional multispecies desmoplastic tumor growth on computer systems of varying processing capabilities. The FMG algorithm is applied to solve a recently published thermodynamic mixture model that uses a diffuse interface approach with fourth-order reaction-advection-diffusion PDEs (Cahn-Hilliard-type equations) that are coupled, nonlinear, and numerically stiff. The model includes multiple cell species and extracellular matrix (ECM), with adhesive and elastic energy contributions in chemical potential terms, as well as including blood and lymphatic vessels represented as continuous vasculatures. Advection-reaction-diffusion PDEs are employed for the cell-ECM components, whereas reaction-diffusion/advection-reaction-diffusion PDEs are used for the cell substrate and vessel species. This desmoplastic tumor model exhibits an extracellular matrix rich tumor microenvironment and may be beneficial when applied to studying fibrotic tumors such as pancreatic adenocarcinoma.

After adding openMP to the FMG code, the program was run for a single time step on a i 7-4600U processor in single core and dual code configurations. Timing macros were used to determine how effective openMP improved the performance of the program from the single core to dual code execution. The model was then timed on a “FAT Node” of the Cardinal Research Cluster (CRC) for 1, 2, 4, 6, 8, 16, and 32 cores.

The resulting data indicate that, relative to a single core system, openMP applied to the FMG algorithm renders the initial time step 1. 7 times faster on a dual core system and approximately 3 times faster on a quad core system. However, overhead between processing cores overtakes the benefits of using openMP on CPUs with more than 8 cores. Although using openMP demonstrates modest improvement in performance, this study indicates that further parallelization is required to achieve model performance that will yield practical benefit.

Share

COinS
 
Oct 17th, 10:00 AM

Parallelization of a Three-Dimensional Full Multigrid Algorithm to Simulate Tumor Growth

We present the performance gains of an openMP implementation of a fully adaptive nonlinear full multigrid (FMG) algorithm to simulate three-dimensional multispecies desmoplastic tumor growth on computer systems of varying processing capabilities. The FMG algorithm is applied to solve a recently published thermodynamic mixture model that uses a diffuse interface approach with fourth-order reaction-advection-diffusion PDEs (Cahn-Hilliard-type equations) that are coupled, nonlinear, and numerically stiff. The model includes multiple cell species and extracellular matrix (ECM), with adhesive and elastic energy contributions in chemical potential terms, as well as including blood and lymphatic vessels represented as continuous vasculatures. Advection-reaction-diffusion PDEs are employed for the cell-ECM components, whereas reaction-diffusion/advection-reaction-diffusion PDEs are used for the cell substrate and vessel species. This desmoplastic tumor model exhibits an extracellular matrix rich tumor microenvironment and may be beneficial when applied to studying fibrotic tumors such as pancreatic adenocarcinoma.

After adding openMP to the FMG code, the program was run for a single time step on a i 7-4600U processor in single core and dual code configurations. Timing macros were used to determine how effective openMP improved the performance of the program from the single core to dual code execution. The model was then timed on a “FAT Node” of the Cardinal Research Cluster (CRC) for 1, 2, 4, 6, 8, 16, and 32 cores.

The resulting data indicate that, relative to a single core system, openMP applied to the FMG algorithm renders the initial time step 1. 7 times faster on a dual core system and approximately 3 times faster on a quad core system. However, overhead between processing cores overtakes the benefits of using openMP on CPUs with more than 8 cores. Although using openMP demonstrates modest improvement in performance, this study indicates that further parallelization is required to achieve model performance that will yield practical benefit.