The computer science community needs simpler mechanisms to achieve the performance potential of accelerators, such as graphics processing units (GPUs), field-programmable gate arrays (FPGAs), and co-processors (e.g., Intel Xeon Phi), due to their increasing use in state-of-the-art supercomputers. Over the past 10 years, we have seen a significant improvement in both computing power and memory connection bandwidth for accelerators. However, we also observe that the computation power has grown significantly faster than the interconnection bandwidth between the central processing unit (CPU) and the accelerator. Given that accelerators generally have their own discrete memory space, data needs to be copied from the CPU host memory to the accelerator (device) memory before computation starts on the accelerator. Moreover, programming models like CUDA, OpenMP, OpenACC, and OpenCL can efficiently offload compute-intensive workloads to these accelerators. However, achieving the overlapping of data transfers with computation in a kernel with these models is neither simple nor straightforward. Instead, codes copy data to or from the device without overlapping or requiring explicit user design and refactoring. Achieving performance can require extensive refactoring and hand-tuning to apply data transfer optimizations, and users must manually partition their dataset whenever its size is larger than device memory, which can be highly difficult when the device memory size is not exposed to the user. As the systems are becoming more and more complex in terms of heterogeneity, CPUs are responsible for handling many tasks related to other accelerators, computation and data movement tasks, task dependency checking, and task callbacks. Leaving all logic controls to the CPU not only costs extra communication delay over the PCI-e bus but also consumes the CPU resources, which may affect the performance of other CPU tasks. This thesis work aims to provide efficient directive-based data pipelining approaches for GPUs that tackle these issues and improve performance, programmability, and memory management. / Doctor of Philosophy / Over the past decade, parallel accelerators have become increasingly prominent in this emerging era of "big data, big compute, and artificial intelligence.'' In more recent supercomputers and datacenter clusters, we find multi-core central processing units (CPUs), many-core graphics processing units (GPUs), field-programmable gate arrays (FPGAs), and co-processors (e.g., Intel Xeon Phi) being used to accelerate many kinds of computation tasks. While many new programming models have been proposed to support these accelerators, scientists or developers without domain knowledge usually find existing programming models not efficient enough to port their code to accelerators. Due to the limited accelerator on-chip memory size, the data array size is often too large to fit in the on-chip memory, especially while dealing with deep learning tasks. The data need to be partitioned and managed properly, which requires more hand-tuning effort. Moreover, performance tuning is difficult for developers to achieve high performance for specific applications due to a lack of domain knowledge. To handle these problems, this dissertation aims to propose a general approach to provide better programmability, performance, and data management for the accelerators. Accelerator users often prefer to keep their existing verified C, C++, or Fortran code rather than grapple with the unfamiliar code. Since 2013, OpenMP has provided a straightforward way to adapt existing programs to accelerated systems. We propose multiple associated clauses to help developers easily partition and pipeline the accelerated code. Specifically, the proposed extension can overlap kernel computation and data transfer between host and device efficiently. The extension supports memory over-subscription, meaning the memory size required by the tasks could be larger than the GPU size. The internal scheduler guarantees that the data is swapped out correctly and efficiently. Machine learning methods are also leveraged to help with auto-tuning accelerator performance.

No description available.

The main focus of this project is usage of multi - GPU systems and usage of CUDA unified memory . Its goal is to accelerate computation of 2D and 3D FFT, which is the main part of simulations in k- Wave library .K- Wave is a C++/ Matlab library used for simulations of propagation of ultrasonic waves in 1D , 2D or 3D space . Acceleration of these functions is necessary , because the simulations are computationally intensive .