Computational Fluid Dynamics (CFD) : CFD Simulation involves solution of
Initial Value problems or Boundary Value Problems in which the solution
of governing partial differential equations (PDEs) by Finite Difference
Method (FDM) and Finite Element Method (FEM) is obtained. The generation of
structured /unstructured grids in complex three-dimensional regions, followed by
solution of matrix system of linear equations involves data-parallel computations.
These involve solution of matrix system of linear equations in which matrix is banded
or sparse and the respective equations are solved on a HPC GPU cluster.
-
Finite Difference Method : Parallelisation : We consider solution of
Poisson System of PDEs with essential boundary conditions
in two- & three- dimensional regions in our computations.
To parallelize FDM algorithm, the physical domain (Structured Grid Generation) is sliced into slabs
(One-dimensional partitioning) or blocks (two-dimensional partitioning).
One of the task is deciding how to assign processes to each part
of the decomposed computational domain. The elements of the array that are used to
hold data from other processes are
called ghost points. The next task is deciding how to assign processes to each part
of the decomposed computational domain. The computational domain, with ghost points
for each process is generated and the resulting block matrix system of equations
are solved by Iterative Method. The MPI implementation and MPI-GPU implementation
based on CUDA enabled NVIDIA GPUs will be implemented to solve the Poisson equation
in which Jacobi /conjugate gradient iterative method will be employed.
Finite Element Method : Parallelisation : We consider solution of
Poisson System of PDEs with essential boundary conditions
in two- & three- dimensional regions in our computations.
To parallelize FEM algorithm, the physical domain (Unstructured Mesh Generation) is partitioned
into sub-domains using Open source software such as METIS (Graph Partitioning Software) and other
mesh partitioning algorithms.
Each MPI process gets partitioned sub-domain and Off-the Processor Communication is
generated. The resulting banded or Sparse matrix system of equations
are solved by Iterative Method. The MPI implementation and MPI-GPU implementation
based on CUDA enabled NVIDIA GPUs will be implemented to solve the Poisson equation
in which conjugate gradent iterative methods will be employed.
HPC GPU Cluster :
In hyPACK-2013 workshop, a prototype HPC GPU cluster (CUDA /OpenCL enabled NVIDIA GPUs
& AMD-ATI OpenCL Prog. env) is used to solve
application kernels, that are based on Heterogenous
programming model
In this workshop, programming and performance issues for applications on
HPC GPU Clusters will be discussed.
In laboratory session, a prototype Hybrid Heterogneous HPC GPU Cluster is made available,
which can address some of the heterogeneous computing workloads.
The HPC GPU Cluster can be made "adaptive" to the
application it is running, assigning the most effective resources in real-time as
per application demands, without requiring modifications to the application. One of the objectives of HPC GPU Cluster (hybrid computing system) is
to allocate resources of CPUs & GPUs in an optimal way to solve applications of different
characteristics.