Doctoral Dissertation Defense: Xuan Huang
Advisor: Dr. Gobbert
Friday, April 17, 2015 · 9:30 - 11:30 AM
Title: An MPI-CUDA Implementation of a Model for Calcium Induced Calcium Release in a Three-Dimensional Heart Cell on a Hybrid CPU/GPU Cluster
Abstract:
A model for Calcium Induced Calcium Release (CICR) in a heart cell describes a physiological process where calcium is able to activate calcium release from the sarcoplasmic reticulum into the cytosol, which is crucial for excitation-contraction coupling in the cardiac muscle. It is modeled by a system of coupled, non-linear, time-dependent advection-diffusion-reaction equations that can be solved by a method of lines approach. The finite element method solves the equation system without advection, while the finite volume method captures the simulation with advection. Through numerical convergence study we show that the finite volume method has the same convergence rate as the finite element method when there is no advection. We also discuss appropriate discretizations of the advection term for different source terms. We present parallel performance studies for two parallel implementations, one using MPI and running on CPU only nodes, the other using CUDA and MPI together and running on hybrid CPU/GPU nodes. We first establish strong and weak scalability of the implementation using MPI. Then with an extended implementation using CUDA and MPI, we show how to combine several hybrid CPU/GPU nodes successfully in a multi-node distributed-memory compute cluster with high performance interconnect. We present results for a combination of different spatial discretizations and different linear solvers, all showing good speedup and outperforming the CPU only implementation.
Abstract:
A model for Calcium Induced Calcium Release (CICR) in a heart cell describes a physiological process where calcium is able to activate calcium release from the sarcoplasmic reticulum into the cytosol, which is crucial for excitation-contraction coupling in the cardiac muscle. It is modeled by a system of coupled, non-linear, time-dependent advection-diffusion-reaction equations that can be solved by a method of lines approach. The finite element method solves the equation system without advection, while the finite volume method captures the simulation with advection. Through numerical convergence study we show that the finite volume method has the same convergence rate as the finite element method when there is no advection. We also discuss appropriate discretizations of the advection term for different source terms. We present parallel performance studies for two parallel implementations, one using MPI and running on CPU only nodes, the other using CUDA and MPI together and running on hybrid CPU/GPU nodes. We first establish strong and weak scalability of the implementation using MPI. Then with an extended implementation using CUDA and MPI, we show how to combine several hybrid CPU/GPU nodes successfully in a multi-node distributed-memory compute cluster with high performance interconnect. We present results for a combination of different spatial discretizations and different linear solvers, all showing good speedup and outperforming the CPU only implementation.