A PLatform for GPGPU Investigations

The evolution of the HPC industry is at a pivotal point: GPU accelerated architectures are now appearing in mainstream facilities around the world. Such systems incorporate those fundamental characteristics important for allowing an approach to the Exascale. However, constraints of the hardware present challenges for the programmer: efficient utilisilisation of the available processing capability is typically only possible for real applications after significant adaptations are made. Access to a testbed GPU accelerated resource will allow researchers to gain vital experience with this disruptive architecture and prepare their applications for the future.

The GPU architecture is suitable for data-parallel problems, where each of the many cores can perform the same operation on a different piece of data. Most of those applications active on large-scale services such as HECToR already operate in this manner: in principle there are wide-ranging potential benefits to the UK research community from the processing and memory bandwidth capabilities on offer from GPUs. The current standard mechanism for programming GPUs involves extensions to traditional languages, and programming challenges arise due to the fact that the programmer is generally responsible for tasks such as identification of those sections which should be accelerated, GPU/GPU data transfer, decomposition of the problem in to a hierarchical thread model etc. Furthermore, typically just porting a code does not achieve anywhere near the available performance, so comprehensive GPU-specific tuning is required to fully exploit the resource. Overall, the amount of effort required to run efficiently on the GPU architecture depends of the nature of the application, not just which algorithms are being invoked but also, in no small part, specifically how they have been implemented. Immediate benefits will be available for those users who have (or have access to) codes which have already been ported and tuned to the GPU architecture. NVIDIA report that they have already ported a number of standard packages popular on HECToR (mostly in the area of Materials Chemistry): AMBER, DLPOLY, GAMESS, LAMMPS, NAMD, VMD, GROMACS. An inspection of the literature indicates that researchers have had success in independently porting codes across the spectrum of scientific and engineering areas, including those areas prominent on HECToR. For those users whose applications are not currently enabled for the GPU architecture, this testbed service will provide the opportunity to understand and evaluate the architecture and associated programming models and in turn provide a platform for development, adaptation and optimization, to prepare for deployment on this and future resources.