Exascale ComputingExascale Computational Simulations

Fusion modelling, as with other scientific disciplines, heavily uses computational simulation to investigate and model the fusion plasma and predict the behaviour of plasmas in tokamaks.  This modelling can require significant amounts of computing resources even to model current fusion tokamaks.  The next generation of fusion tokamaks being developed, ITER and DEMO, are significantly larger than current machines.  There is also a growing requirement for full tokamak modelling (modelling all the  processes within the plasma) rather than the current practice which is generally to model one part of the plasma (i.e. the core plasma, or the edge plasma) by itself.  Both of these requirement mean that the computational resources required for fusion modelling in the future will be significant.

Exascale computing, moving beyond the current level of computing power in large computers (Petascale computers) to Exascale machines (Exascale = 1000 x Petascale), has significant challenge associated with it.  Indeed, successfully and efficiently exploiting the Petascale supercomputers that are available today is not straight-forward and requires significant effort in the design and development of the simulation programs.

Within the Nu-FuSE project we will be investigating the challenge of exploiting Peta- and Exascale systems for fusion simulations and creating tools, techniques, and tailoring fusion simulation codes to enable fusion research to exploit such systems.

Through the partners, the project has access to a range of Petascale systems which are today’s exemplars of tomorrow’s Exascale computers including traditional supercomputers such as a Cray XE6 and IBM BlueGene/Q as well as more novel GPGPU-based demonstrator systems. This give us unparalleled abilities to test and refine new algorithms.

The nature of the computational scaling requirements of fusion compuiting can be illustrated by the Plasma Physics investigations of turbulence using the GTC-P code – a highly scalable particle-in-cell code used for studying micro-turbulent transport in tokamaks. The parallel algorithm is implemented with MPI and OpenMP and was developed specifically for Blue Gene (BG/P) systems to allow simulation of very large fusion devices with unprecedented efficiency – a key feature that Nu-FuSE is studying is the examination of the key question of how plasma microturbulence is affected as the plasma size increases to the very large plasmas characteristic of the ITER reactor.

Traditionally, we have used methods such as MPI, OpenMP to express the parallelism explicitly, however this approach may not be scalable to exascale. Hence, the project is investigating  next generation parallel programming languages, such as Xcalable MP (XMP), which provided both implicit and explicit parallel descriptions for ease and use, whilst retaining performance tuning on ultra-large scale scientific codes.