Mr Farhan Rizwi1, Dr Mike Herzfeld1
1CSIRO, Hobart, Australia
Biography:
Farhan is one of the developers of the coastal ocean numerical code base, the Environment Modelling Suite (EMS). His interests include HPC optimisation, visualisation and frameworks for running large scale 4D ocean models.
Abstract:
Computers are getting faster, but the free ride of clock speed increases has long since ceased. Instead, the performance gains have largely materialised from massive data parallelism across distributed nodes, increasing cores per CPU, bigger caches, the return of hardware vectorisation (in fat registers) and use of GPU technology.
In this talk, I first present the MPI techniques used to optimise the domain decomposition, coarse-grain parallelisation of the unstructured ocean model COMPAS (Coastal Ocean Marine Prediction Across Scales). I then explore the challenges faced to fully utilise the power of the modern CPU/GPU. Often algorithms are memory bound and involve lots of indirect data access, which is usually a consequence of the straightforward implementation of mathematical equations. However, highly performant code necessitates a meticulous audit of every memory operation and as such the coding techniques needed may not be as intuitive. To address this challenge, exploration of sparse matrices and optimisation techniques of bandwidth reduction will be explored.