About Pieter Ghysels
Pieter Ghysels is a research scientist in the Scalable Solvers Group of the Computational Research Division at Lawrence Berkeley National Laboratory, in Berkeley, California. His main interests are in High Performance Computing (HPC) and linear algebra. Pieter has expertise in both iterative methods and direct methods for the solution of systems of linear equations. On the topic of iterative methods, Pieter has done pioneering work on hiding communication latency in Krylov algorithms. He presented a class of algorithms, called pipelined Krylov solvers, in which global communication latency can be overlapped with local work. This idea has been applied to several Krylov algorithms: the Generalized Minimal Residual Method (GMRes), Conjugate Gradients (CG), and Conjugate Residuals (CR). Pieter has also published results about performance modeling of (geometric) multigrid on novel multi/many-core architectures.
Recently Pieter started working on factorization based (direct) solvers for linear systems. He is now the main author (together with Xiaoye Sherry Li) of the STRUMPACK software library which offers a direct solver and a preconditioner for large sparse linear systems as well as memory effient representation of structured dense matrices. STRUMPACK leverages hierarchical matrices, using low-rank approximations, to get low complexity solvers and efficient preconditioners for large sparse systems derived from discretization of a variety of partial differential equations. Currently, STRUMPACK makes use of the Hierarchically Semi-Separable (HSS) matrix format and randomized sampling to efficiently find low-rank representations for matrix sub-blocks.
Pieter Ghysels received an Engineering degree in 2006 from KULeuven, the (Flemmish) Catholic University in Leuven, Belgium. He completed a PhD in engineering Sciences under the supervision of Dirk Roose and Giovanni Samaey titled "Coupling of Fine and Coarse Scale Models for the Simulation of Viscoelastic Plant Tissue - From Cellular Structure to Homogeneous Material", also at the KULeuven in 2010. From 2010-2013, Pieter worked at the Universiteit Antwerpen (University of Antwerp, Belgium) and at the Intel Exascience Lab Flanders on scalable iterative algoritms for highly parallel machines. During that time Pieter also worked on fast methods for algebraic tomographic reconstruction.
- Performance Computing (HPC), parallel algorithm design
- Hierarchical matrices (HSS, H, H2, HODLR, BLR)
- Iterative methods (Krylov methods, algebraic/geometric multigrid) for linear systems
- Direct, factorization based solver for sparse linear algebra, see STRUMPACK
- Stencil computations and memory-aware code optimizations
- Communication-avoiding and communication/synchronization hiding numerical algorithms
- Code optimization, code profiling