Internship in data science for scientific computing: accelerating advanced stochastic methods on modern computer architectures
Start date : 2 March 2020
Mission duration : 6 months
Deadline for applications : 1 March 2020
Salary : 620€/month
Multilevel Monte Carlo (MLMC) sampling is an approach popularized in 2008 by Giles for the solution of stochastic differential equations. It has since been adapted to accelerate uncertainty quantification (UQ) sampling methods for partial differential equations (PDEs), with applications in various areas of engineering and applied mathematics for industrial problems. The underlying idea of MLMC is to take advantage of different levels of numerical discretization in such a way that many (cheap) evaluations are performed on the coarsest levels while fewer (expensive) computations are required on the finest levels, resulting in a reduced computational cost. In terms of error, many coarse evaluations help reduce the sampling error while fewer fine evaluations help reduce the discretization error. MLMC methods have a solid mathematical foundation and have demonstrated their superiority over single-level Monte Carlo sampling methods on a wide range of applications.
Modern computer architectures, especially upcoming exascale parallel systems, provide exciting opportunities to design efficient MLMC methods that make the most of mixed precision floating-point arithmetics offered by hardware accelerators such as graphical processing units (GPU) and tensor cores. This is indeed very well aligned with the MLMC core idea of balancing different sources of errors. The proposed internship project aims at developing a prototype MLMC algorithm exploiting these concepts.
- Having completed one parallel programming or HPC course
- Experience in C or C++ programming
- Experience with GPUs
- Experience in scientific computing
How to apply
Have a look at the full internship description.