GPU'lu kümelenmiş sistemlerde yapısal çözüm alanı problemlerine yönelik bloklamaya dayalı programlama modeli

Currently, more than 25\% of supercomputers in TOP500 list employ GPUs due to their massively parallel and power-efficient architectures. However, programming GPUs efficiently in a large-scale system is a demanding task not only for computational scientists but also for programming experts because it requires generating GPU-specific code, managing distinct address spaces and handling communication. While these requirements reduce productivity, they also limit the portability of the application. There are pragma-based programming models such as OpenACC offering portability and productivity with code annotations. However, they require performance tuning and lack support for programming GPU clusters. Handling communication efficiently is essential in the cluster environment. There are related works such as Cluster-SkePU handling communication and offering skeleton-based abstractions to ease GPU programming, but such abstractions arguably limit programming flexibility. Moreover, most of the works in the literature do not consider the bandwidth bottleneck of the interconnect that links GPUs to hosts. Some task-based approaches provide overlapping data transfers with computation as a solution to the interconnect bandwidth limit, but they hold programmer responsible for GPU-specific code and task scheduling. To ease the programming effort, increase the portability and optimize communication, we propose a tiling-based programming model for structured grid problems running on a GPU cluster. The model implicitly applies data decomposition with TiDA and manages distinct address spaces with CUDA. It automatically generates GPU-specific code itself by leveraging OpenACC annotations behind a uniform interface for CPUs and GPUs. Furthermore, it handles data transfers and communication and overlaps them with computation by exploiting CUDA streams and non-blocking MPI routines. We demonstrate the effectiveness of the programming model on a heat simulation and a real-life cardiac modeling. The results show that it successfully overlaps communication and achieves good speedup.