Multiscale Computing in the Exascale Era
📝 Abstract
We expect that multiscale simulations will be one of the main high performance computing workloads in the exascale era. We propose multiscale computing patterns as a generic vehicle to realise load balanced, fault tolerant and energy aware high performance multiscale computing. Multiscale computing patterns should lead to a separation of concerns, whereby application developers can compose multiscale models and execute multiscale simulations, while pattern software realises optimized, fault tolerant and energy aware multiscale computing. We introduce three multiscale computing patterns, present an example of the extreme scaling pattern, and discuss our vision of how this may shape multiscale computing in the exascale era.
💡 Analysis
We expect that multiscale simulations will be one of the main high performance computing workloads in the exascale era. We propose multiscale computing patterns as a generic vehicle to realise load balanced, fault tolerant and energy aware high performance multiscale computing. Multiscale computing patterns should lead to a separation of concerns, whereby application developers can compose multiscale models and execute multiscale simulations, while pattern software realises optimized, fault tolerant and energy aware multiscale computing. We introduce three multiscale computing patterns, present an example of the extreme scaling pattern, and discuss our vision of how this may shape multiscale computing in the exascale era.
📄 Content
Multiscale Computing in the Exascale Era Saad Alowayyed1,2, Derek Groen3, Peter V. Coveney4, Alfons G. Hoekstra1,5,*
1 Computational Science Lab, Institute for Informatics, University of Amsterdam, The Netherlands 2 King Abdulaziz City for Science and Technology (KACST), Riyadh, Saudi Arabia 3 Department of Computer Science, Brunel University London, United Kingdom 4 Centre for Computational Science, University College London, United Kingdom 5 ITMO University, Saint Petersburg, Russia * corresponding author, a.g.hoekstra@uva.nl
Keywords multiscale modelling, multiscale computing, high performance computing, exascale
Abstract We expect that multiscale simulations will be one of the main high performance computing workloads in the exascale era. We propose multiscale computing patterns as a generic vehicle to realise load balanced, fault tolerant and energy aware high performance multiscale computing. Multiscale computing patterns should lead to a separation of concerns, whereby application developers can compose multiscale models and execute multiscale simulations, while pattern software realises optimized, fault tolerant and energy aware multiscale computing. We introduce three multiscale computing patterns, present an example of the extreme scaling pattern, and discuss our vision of how this may shape multiscale computing in the exascale era.
1 Introduction Science is at its most powerful when it can not only convincingly explain the processes at work in natural phenomena but is able to predict what will occur before it does so. Predictions of real world events, such as weather forecasting, when a cement will set, the occurrence of an earthquake or what medical intervention to perform in order to save a person’s life, all require the bringing together of substantial quantities of data together with the performance of one or more likely many high-fidelity four (three space and one time) dimensional simulations before the event in question occurs. Such forms of calculation are among the most demanding known in computational science, as they need to be done rapidly, accurately, precisely and reliably, including quantification of the uncertainties associated with them. They are also multiscale in nature, as their accuracy and reliability depend on the correct representation of processes taking place on several length and time scales. Only now, as we move toward the exascale era in high performance computing (HPC) can we expect to be able to tackle such problems effectively and, eventually, in a routine manner.
Indeed, multiscale phenomena are everywhere around us [1–7]. If we study the origin and evolution of the universe [8] or properties of materials [9–13], if we try to understand health and disease [3,14– 21] or develop fusion as a potential energy source of the future [22], in all these cases and many more we find that processes on quite disparate length and time scales interact in strong and non-linear ways. In short, multiscale modelling is ubiquitous and progress in most of these cases is determined by our ability to design and implement multiscale models of the particular systems under study [1,6,23].
The sheer complexity of such multiscale phenomena still limits our capability to perform high-fidelity simulations that accurately and reliably predict the behaviour of a given system in all situations. Capturing even a few of those coupled processes in a multiscale simulation quickly reaches the limits of contemporary high performance computing at the petascale.
That the importance of multiscale modelling in many domains of science and engineering is still increasing is clearly demonstrated in numerous publications; see, for example, [1,24]. Therefore, we must anticipate that multiscale simulations will become an increasingly important form of scientific application on high end computing resources, necessitating the development of sustainable and reusable solutions for such emerging applications, that is, generic algorithms for multiscale computing. As we move into the exascale performance era we need to drastically change the way we use HPC for simulation based sciences [25].
For example, on current resources we can simulate composite nanomaterials using tens of millions of atoms, where the interaction potentials rest in turn on electronic structure and atomistic simulations of millions of particles but are performed at higher levels of particulate coarse-graining. Such simulations have already led to ground-breaking insights into chemically specific structural self- assembly and large scale materials property prediction [9]. However, they are still limited to at best volumes of matter whose linear dimensions are on the micrometre scale, well below the size one would need to study, e.g., for the formation of fractures on millimetre scales, let alone to predict performance of materials on macro scales such as arise in typical au
This content is AI-processed based on ArXiv data.