
|
Name: Panchatcharam Mariappan
Qualification: Ph. D in Mathematics, 2013, IIT Madras, Chennai, India
PhD Thesis Title: GPU accelerated finite point set method for fluid flow problems
PhD thesis is a collaborative work of IIT Madras, India and TU Kaiserslautern, Germany
Research Experience: Post doctoral Fellow, 2013-2014, Fraunhofer ITWM, Kaiserslautern, Germany
Professional Appointments:
Software Development Engineer, 2014-present, NUMA Engineering Services Ltd, Dundlk, Co. Louth, Ireland
Lecturer, 2006-2007 Vysya College, Salem, Tamilnadu, India
Online Tutor, 2007, TWWI, Chennai, India
Publications:
1. Panchatcharam M and Sundar S., Finite Pointset method for 2D-dambreak problem with GPU acceleration, International Journal of Applied Mathematics, 25 (2012) 4:545
2. Panchatcharam M., Sundar S., Vetrivel V., Klar A and Tiwari S., GPU computing for meshfree particle method, Accepted, International Journal of Numerical Analysis and Modeling, Series B, 2013.
3. Panchatcharam M., Sundar S and Klar A., GPU metrics for linear solver, Neural, Parallel and Scientific Computations, 21 (2013) 361-374.
Area of Interest: GPU Computing, Mesh Free Methods, CFD, Numerical Analysis, Numerical Linear Algebra
Awards and Honours:
(1) DAAD Fellow, 2010-2013,
(2) GATE Scholarship, 2007-2010,
(3) NBHM Fellow, 2004-2006
Graphics Processing Units (GPUs) are nowadays used for numerical computation , beyond their original purpose of graphics accelerators. Mature hardware and GPU software tools and libraries support double precision and memory correction. GPU accelerated computational fluid dynamics has gained attention in both academia and industry. In this article, we investigate the importance of GPUs as accelerators in the field of biomedical engineering. We developed a software tool to predict the lesion development in cancer patients after the radio frequency ablation cancer treatment. We use Penne’s bioheat model with appropriate boundary conditions and the finite element method for numerical discretization. From the finite element discretization of the bioheat equation, we observe that no explicit element integration is required. Since the problem domain is fixed, we find the neighbours of each node at the first time step and generate a compressed sparse row structured (CSR) matrix which can be used for the entire domain. After the CSR matrix is generated, we send the domain information such as nodes, elements and matrix information (e.g. the CSR matrix rows and columns) to the GPU. The Central Processing Unit (CPU) loads the initial data, finds the neighbours list, generates the CSR matrix and stores the results on the disk, whereas the GPU constructs the shape functions, assembles the local stiffness matrix into the global matrix in the CSR form and solves the sparse linear system with the help of the existing CUDA libraries such as CUBLAS and CUSPARSE. In order to solve the linear system, we employed the ILU preconditioned BiCGStab algorithm, one of the fastest solvers among Krylov subspace solvers. At each time step, the GPU generates the heat source term and solves the cell death model, while the CPU saves the results in vtu/vtp files. The heat source term generation is based on our in-house point source model for approximating the Joule heating effect, and the cell death model is an adapted evolution equation, predicting whether cells near the tumour are alive or dead. The tasks assigned to the GPU are the most time consuming parts of the finite element method and the GPU accelerates them with the desired speed-up and accuracy. The major steps involved in this work are receiving the segmented CT scans of the patient from the doctors, generating the mesh, obtaining the needle position from the CT scans (approximately the centre of the tumour) and simulating them using our software tool. Existing software tools working on multi-core CPUs (Intel i5’s) take 6 hours to predict a lesion for 26 minutes of real treatment time, for around 1 million elements. Our current work with the assistance of the GPU acceleration yields the result in approximately 3 minutes for the same number of elements, where the comparison is done with Intel Xeon CPU E5 – 2680 @ 2.8 GHz, and NVIDIA GeForce Titan Black GPU @ 3.5 GHz (2880 CUDA Cores).
|