next up previous contents
Next: OpenMP Implementation Up: Introduction Previous: Rosa   Contents

Concurrent development work

It should be noted that concurrently with this work, a new sparse matrix library has been introduced into CP2K - DBCSR (Distributed Block Compressed Sparse Row) - developed by Dr. Urban Borstnik at the University of Zurich. This library is designed from the outset to be highly scalable and particularly suited to density functional theory calculations, taking advantage of the blocked structure of the matrices. It also makes use of OpenMP to distribute work across a shared memory node, with MPI for inter-node communication. Since sparse matrix operations take up a significant fraction of the runtime for the types of jobs we have used for testing and benchmarking (typically 30% or more), the introduction of this library has a strong effect on the performance of the code, and is still under development. Where possible, the effects of this have been removed from the results reported, but the final benchmarks (section 4) do show some improvement due to DBCSR as well as the work of the project.

Iain Bethune