high performance computing on graphics processing units: hgpu.org

Nowadays, similarity search is becoming a field of increasing interest because these kinds of methods can be applied to different areas in computer science and engineering, such as voice and image recognition, text retrieval, and many others. However, when processing large volumes of data, query response time can be quite high. In this case, it is necessary to apply mechanisms in order to significantly reduce the average query response time. In this sense, the parallelization of the metric structures processing is an interesting field of research. Currently, most of the previous and current works developed in this area are carried out considering classical distributed or shared memory platforms. However, modern GPU/MultiGPU systems offer a very impressive cost/performance ratio as compared to multiprocessor or multicomputer platforms that are usually more expensive gaining in significance and popularity within the scientific computing community. More recently, GPUs have been proposed to evaluate similarity queries for indexes that remains statically stored in GPU’s memory. In this paper we propose two different pipelines to accelerate the process of similarity queries in datasets large enough not to fit in memory of the GPUs. The first pipeline makes use of CPU-cores and GPUs in a hybrid algorithm, and the second one is implemented into the GPU. The results show that the best performance is achieved with both pipelines at the same time.