Computer Science Day
High Performance Scientific Computing at the Exascale
LMS Computer Science Colloquium
Tuesday 11th October 2011
The aims of the colloquium are to showcase state of the art scientific computing applications, to identify some of the challenges posed by next generation high performance machines and indicate promising approaches to tackling the problem.
Researchers and students from both the mathematics and computer science communities are especially welcome to participate in this colloquium.
To register, please contact Elizabeth Fisher (computerscience [at] lms [dot] ac [dot] uk). The day is free for students and £5 for all others which is payable on the day. A sandwich lunch will be provided.
11.00-12.00: Marc Snir (Illinois)
Exascale Computing: The Last Rehearsal Before the Post-Moore Era
The expected slow-down in the trend described by Moore's Law implies that it will hard to develop an exascale computer, and practically impossible to develop a zettascale computer, using silicon echnology. Research and Development must shift from a focus on leveraging an increasing number of transistors to doing more with a fixed power and transistor budget. The talk will describe these trends and propose a research agenda for increased compute efficiency. This agenda is immediately relevant to work on exascale computing since the power requirements of a conventional exascale system will be prohibitive.
12.00-13.00: Ulrich Ruede (Erlangen)
Towards Exascale Computing: Multilevel Methods and Flow Solvers for Millions of Cores
Ulrich Ruede will report on his experiences implementing PDE solvers on existing Peta-Scale computers, such as the 290 000 core IBM Blue Gene system in the Jülich Supercomputing Center.
The talk will have two parts, the first one presenting his Hierarchical Hybrid Grid (HHG) method, a prototype Finite Element Multigrid Solver scaling up to a trillion (10^12) degrees of freedom on a tetrahedral mesh by using a carefully designed matrix-free implementation. The second part of the talk will present his work on simulating complex flow phenomena using the Lattice-Boltzmann method. His software includes parallel algorithms for treating free surfaces with the goal of simulating fully resolved bubbly flows and foams. Additionally, he will report on a parallel fluid-structure-interaction technique with many moving rigid objects. This work is directed towards the modelling of particulate flows that he can represent using fully resolved geometric models of each individual particle embedded in the flow. The talk will end with some remarks on the challenges that algorithm developers will be facing on the path to exascale in the coming decade.
13.00 – 14.00: Lunch
14.00 – 15.00: Oubay Hassan (Swansea)
Towards Fully Parallel Aerospace Simulations on Unstructured Meshes
Aerospace industries are employing unstructured meshbased solution techniques for the simulation of complex phenomena such as steady and unsteady compressible flows and the scattering of electromagnetic wave over complex geometries. Implementations of the approach, on a wide variety of computer platforms, have demonstrated that acceptable accuracy can often be achieved, for these class of problem, by employing meshes consisting of the order of a tens of million elements.
However, the computer time requirements of the approach have prevented its extension to simulations which require the use of significantly larger meshes. This is particularly apparent when problems involving turbulent flows, or even transient inviscid flows in the presence of moving boundaries, are considered. The situation is even worse when electromagnetic scattering by aerospace vehicles at realistic wave frequencies is simulated, as the available numerical algorithms will demand the use of meshes consisting of the order of billions of elements.
With this background, and given the current developments in computer technology, we have directed research into techniques which are based upon the use of parallel processing for the solution of computationally large problems. However, for such an approach to be successful in the current context, it will require not only a parallel implementation of the basic equation solver which is being employed but also a new approach to the pre--processing and post--processing stages of mesh generation, domain decomposition and data visualisation.
In this presentation, a general philosophy is presented in which all the modules within the computational cycle are parallelised and executed on parallel computer hardware, thereby avoiding the creation of computational bottlenecks. In particular, unstructured mesh generation with adaptation, computational fluid dynamics and computational electromagnetic solvers and the visualisation of grid and solution data are all performed in parallel.
An overview of these developments will be provided. In particular, details of the parallel mesh generator, which has been used to generate meshes in excess of 1 billion elements, will be given. A brief overview will be presented of the approach used to parallelise the solvers and how large data sets are interrogated and visualised on distributed computer platforms. Details of the parallel adaptation algorithm will be presented. Several examples are given of the approach applied to the simulation of large aerospace calculations in the field of aerodynamics and electromagnetics.
15.00-16.00: Paul Kelly (Imperial College)
Software Abstractions for many-core software engineering
What is the right code to generate, for a given hardware platform? How does this change as problem parameters change? This talk presents some recent work-in-progress exploring domain-specific languages and active libraries as a way to automate code generation for multicore and manycore platforms, and to capture the space of alternative implementation choices at a higher level than a compiler for a general-purpose language can. Paul Kelly will illustrate the potential for this idea by looking at our recent work on unstructured-mesh fluid dynamics, and finite element methods. By choosing the abstraction carefully, we can capture design choices far beyond what a conventional compiler can do - and, in the extreme, engage the users in selecting algorithms and numerical methods that match the capabilities of the underlying hardware to meet end-to-end objectives for solution quality.
16.00-16.30: Tea
Limited funds are available to help with students with travel costs. Further details are available from Elizabeth Fisher at the Society (computerscience [at] lms [dot] ac [dot] uk)
A short poster for the event is available (PDF)
Submitted by Donald Eastwood on 12 September, 2011 10:28