Algorithmic Skeleton Framework for the Orchestration of GPU Computations

Abstract

perpétuo e sem limites geográficos, de arquivar e publicar esta dissertação através de exemplares impressos reproduzidos em papel ou de forma digital, ou por qualquer outro meio conhecido ou que venha a ser inventado, e de a divulgar através de repositórios científicos e de admitir a sua cópia e distribuição com objectivos educacionais ou de investigação, não comerciais, desde que seja dado crédito ao autor e editor. iv Eu dedico esta tese a todos aqueles que, de uma forma ou de outra, contribuiram para que ela se torna-se possível. Esta tese não é so minha, é também vossa. A huge Thank You to Soraia Assis, whose friendship got me by the rough times, and whose smile warmed my own. Without you this thesis would be but a dream.I thank my parents for all the sacrifices that they endured, so that I would be fortunate enough to study, and eventually reach this mark in my life.Thank you to all my family members and friends, whose names were not mention so far. They are the silent heroes, whose word is spoken throughout this thesis. Obrigado a todos! vii viii AbstractThe Graphics Processing Unit (GPU) is gaining popularity as a co-processor to the Central Processing Unit (CPU), due to its ability to surpass the latter's performance in certain application fields. Nonetheless, harnessing the GPU's capabilities is a non-trivial exercise that requires good knowledge of parallel programming. Thus, providing ways to extract such computational power has become an emerging research topic.In this context, there have been several proposals in the field of GPGPU (Generalpurpose Computation on Graphics Processing Unit) development. However, most of these still offer a low-level abstraction of the GPU computing model, forcing the developer to adapt application computations in accordance with the SPMD model, as well as to orchestrate the low-level details of the execution. On the other hand, the higher-level approaches have limitations that prevent the full exploitation of GPUs when the purpose goes beyond the simple offloading of a kernel.To this extent, our proposal builds on the recent trend of applying the notion of algorithmic patterns (skeletons) to GPU computing. We propose Marrow, a high-level algorithmic skeleton framework that expands the set of skeletons currently available in this field. Marrow's skeletons orchestrate the execution of OpenCL computations and introduce optimizations that overlap communication and computation, thus conjoining programming simplicity with performance gains in many application scenarios. Additionally, these skeletons can be combined (nested) to create more complex applications.

On the support of task-parallel algorithmic skeletons for multi-GPU computing

“…Productivity of Marrow against OpenCL for the programming of single-GPU systems has already been addressed in [8]. Naturally, given than the multi-GPU support on Marrow mostly impacts the runtime system, while on OpenCL it falls on the programmer, the gap between both platforms widens.…”

“…Our proposal builds on Marrow [8], a C++ ASkF for the orchestration of complex OpenCL computations. Among its most distinguishable features are a set skeletons previously unavailable in the GPU context, such as Pipeline and Loop, and the ability to compose these skeletons, through nesting, to create the aforementioned complex computations -a feature also new to this context.…”

SafeGPU: Contract- and library-based GPGPU for object-oriented languages

“…Arbitrary skeleton nesting is provided in FastFlow [7] (resp. Marrow [8]) for pipeline and farm (resp. pipeline, stream, loop), but concurrency and synchronization issues are exposed to the programmer.…”

Performance-aware component composition for GPU-based systems

“…Marrow [154] is a skeleton programming framework for systems containing a single GPU using OpenCL. It provides data (map) and task parallel (stream, pipeline) skeletons that can be composed together to model complex computations.…”