Performance evaluation of data migration methods between the host and the device in CUDA-based programming

• Autor(a) principal: Santos, Rafael Silva [UNESP]
• Data de Publicação: 2016
• Outros Autores: Eler, Danilo Medeiros [UNESP], Garcia, Rogério Eduardo [UNESP]
• Tipo de documento: Capítulo de livro
• Idioma: eng
• Título da fonte: Repositório Institucional da UNESP
• Texto Completo: http://dx.doi.org/10.1007/978-3-319-32467-8_60; http://hdl.handle.net/11449/177938

CUDA-based programming model is heterogeneous – composed of two components: host (CPU) and device (GPU). Both components have separated memory spaces and processing units. A great challenge to increase GPU-based application performance is the data migration between these memory spaces. Currently, the CUDA platform supports the following data migration methods: UMA, zero-copy, pageable and pinned memory. In this paper, we compare the zero-copy performance method with the other methods by considering the overall application runtime. Additionally, we investigated the aspects of data migration process to enunciate causes of the performance variations. The obtained results demonstrated in some cases the zero-copy memory can provide an average performance on 19% higher than the pinned memory transfer. In the studied situation, this method was the second most efficient. Finally, we present limitations of zero-copy memory as a resource for improving performance of CUDA applications.