high performance computing on graphics processing units: hgpu.org

Considerable research has been conducted recently on near-data processing techniques as real-world tasks increasingly involve large-scale and high-dimensional data sets. The advent of solid-state drives (SSDs) has spurred further research because of their processing capability and high internal bandwidth. However, the data processing capability of conventional SSD systems have not been impressive. In particular, they lack the parallel processing abilities that are crucial for data-centric workloads and that are needed to exploit the high internal bandwidth of the SSD. To overcome these shortcomings, we propose a new SSD architecture that integrates a graphics processing unit (GPU). We provide API sets based on the MapReduce framework that allow users to express parallelism in their application, and that exploit the parallelism provided by the embedded GPU. For better performance and utilization, we present optimization strategies to overcome challenges inherent in the SSD architecture. A performance model is also developed that provides an analytical way to tune the SSD design. Our experimental results show that the proposed XSD is approximately 25 times faster compared to an SSD model incorporating a high-performance embedded CPU and up to 4 times faster than a model incorporating a discrete GPU.