high performance computing on graphics processing units: hgpu.org

The interactions between the components of many natural and artificial systems can be described using a graph. These graphs often have an irregular structure with non-trivial topological features. Complex system behaviour emerges on the macroscopic scale from a large number of relatively simple interactions on the microscopic scale. To better understand the observed behaviour of a complex system, the interactions among its basic elements are commonly described in a computational model. As long as the interactions are defined accurately and the number of elements is large enough for complex patterns to emerge, a simulation based on such a model is expected to produce the same behaviour as the system under investigation. The difficulty is often to simulate the model on a large enough scale to obtain scientifically meaningful results. Powerful computer systems are required to calculate the effects caused by the interactions of large numbers of elements. Supercomputers that are constructed from hundreds of thousands of processing units can be used to update many components of the system in parallel and thus reduce the overall simulation time, but these systems are expensive to buy and maintain. As the processor architectures used in workstations and regular desktop computers are becoming more powerful, a small cluster constructed from these systems can be a more viable option. In recent years, the highly data parallel architecture of commodity graphics processing units (GPUs) has received a growing amount of attention due to their high peak compute throughput compared to central processing units (CPUs). New software development tools that turn the GPU hardware into a general purpose compute accelerator have become available. This thesis describes how GPUs can be used to accelerate scientific simulations of complex systems that are based on irregular graph structures. New software development approaches and algorithms are needed to fully utilise the data parallel many-core architecture of today’s GPUs. Irregular graph structures are particularly challenging, as the hardware is based on the single instruction, multiple data (SIMD) processor design, where a group of processing elements receives the same instructions. The architecture also imposes strict requirements on memory access patterns, making the optimisation of the memory layout for the irregular data structures and associated information particularly important. Performance suffers dramatically when the algorithm does not comply with these design restrictions. The author proposes different software design strategies for a number of common graph problems and discusses the advantages and disadvantages of each approach. Two complex system models are used to demonstrate how the GPU can help to accelerate scientific simulations. The first model investigates how the phase transition from ordered to disordered system states in a computational ferromagnet is affected by distortions to the lattice substrate. The second model implements a large scale spiking neural network. The findings show that it is beneficial to utilise the GPU as accelerator to the CPU in almost all scenarios, as long as the project has a long enough run time to justify the more complex software development of the data parallel algorithms. When the model has some regularities in its structure or when some of the design decisions that influence the way memory is accessed can be made with the data parallel architecture in mind, then it is possible to achieve such high performance on the graphics device that it is best to leave the entire computational work to the GPU and use the CPU only to manage the execution of the program.