Research On GPU Based Data Stream Parallel Processing

The occurrence Multi-Core CPU and Multi-Core GPU means parallel computing system become mainstream processing unit. Further more, reference to Moore's law the parallelism of processing unit will continuous extend. This also means parallel computing will become mainstream in future science computing. Relative to very expensive vector processor and MPI parallel computing standard which has huge network cost. As a inexpensive, widespread, powerful computing horsepower and high bandwidth general purpose parallel computing device, Graphic Processing Unit draws more and more domestic and abroad scholars' attention.As a new type of data, data stream has the properties of fast, continuous arrival, potential immensity capacity witch challenge current data stream processing and mining algorithms. How to increase data stream system throughput and real time processing ability of data stream becomes one of key problems in data stream research area.This paper goes in for research on Graphic Processing Unit data stream parallel processing and mining. Unite characteristics of both GPU general parallel computing and data stream; present a GPU data stream parallel computing frame model. Research on both single and multiple GPU data stream processing methods separately.For Single data stream, present a parallel data stream quantiles computing method. This method takes data block as data exchange unit between host and device memory, use GPU as a co-processor to parallel compute data stream quantiles. This method increase data stream processing throughput remarkably.For multiple data stream present a four layers sliding window model which cross PCIE BUS and minimize data exchange between host and device memory. Implement a GPU based multiple data stream processing frame with Computing Unified Device Architecture. Issue a GPU based multiple data streams correlations computing algorithm with exact computing method. Experiment shows that GPU algorithm highly advanced in real time property compare to pure CPU algorithm in computing correlations for multiple data streams. This frame can be extend to other research problems in multiple data stream mining such as data streams PCA analysis or cluster.