Titulo Estágio
High performance computing using GPU acceleration
Área Tecnológica
Informática Industrial
Local do Estágio
Ciengis, IPN, Rua Pedro Nunes, 3030-199 Coimbra
Enquadramento
Most scientific research and engineering software rely on High Performance
Computing (HPC) for sensitive calculations. Recent developments in the computer
graphics area allows the use of GPU acceleration for HPC and therefore is receiving much attention. However, there is currently a lack of a good sparse linear
algebra library that takes advantage of these recent developments.
The main goal of this internship will be the
development of such a library.
Objetivo
The student is expected to make an
assessment of the existing technologies/frameworks for GPU program
development such as OpenCL and CUDA.
The chosen framework will be
used to develop a C/C++ library for sparse linear algebra operations
taking advantage of GPU hardware acceleration.
An interface for
a commonly used Nonlinear and differential/algebraic equation solvers
suite will also be developed.
Plano de Trabalhos - Semestre 1
Task 1 (Sept-Oct): Assessment of the
different frameworks for GPU programming
The student will write a
small report with a comparison between the existing frameworks for
GPU programming which may be used for High Performance Computing.
Frameworks like OpenCL and CUDA will be reviewed.
Task 2
(Oct-Nov): Assessment of the existing strategies used in sparse linear
algebra
The intern will write a review on the several strategies
taken in existing sparse linear algebra libraries. A comparison
between data structures and parallelization strategies will be of key
significance in this overview.
Task 3 (Dec): Definition of
the platform specifications
Considering the results of two
previous tasks, the student will design a platform were he/she will
do the development of the library and later the testing/benchmarking
(hardware/framework).
Plano de Trabalhos - Semestre 2
Task 4 (Feb-Apr): Development of the
sparse linear algebra library
The high performance computing
library for sparse linear algebra will be written using the framework
chosen in task 3. The library should include a linear solver and be
able to perform a Cholesky decomposition.
Task 6(Apr):
SUNDIALs interface
The intern will develop an interface for a
Nonlinear and differential/algebraic equation solvers suite -
(SUNDIALS).
Task 5(May): Benchmarking
The student will
create examples for benchmarking and then perform a comparison
between GPU and CPU performance. Additionally the effect of several
configuration options may be studied.
Condições
The student will work at Ciengis, in
Instituto Pedro Nunes (near pólo II).
In the 1st semester the
student will work on a regular computer available at Ciengis. After
the completion of task 3, the student will work on the chosen
platform.
The intern will receive a monetary compensation in the
2nd semester.
Observações
All development will be performed in Linux.
Orientador
João Leal
joao.leal@ciengis.com 📩