Research Of Parallel Method For GPU-based Multiple Sequence Relevance Analysis

Multiple sequence relevance analysis is a method based on Multiple SequenceAlignment (MSA) which is used to make deep analysis for the associated relevance amongall the sequences. The available techniques on CPU can not be able to fulfill the need ofpractical applications with the ever-increasing of database, mainly including the operationalcomplexity, as well as the low efficiency. As the fast evolution of Graphics Processing Unit(GPU), it has become an efficient platform for high performance computing with itsparallel operating model and strong parallel capacity, which is widely used for low efficienthandles, including parallel optimization of multiple sequence relevance analysis.Benefiting from the powerful computation capacity, a parallel method for multiplesequence relevance analysis is proposed and implemented based on GPU. Paralleltechniques with three different aspects have been put forward to parallelize the process ofmultiple sequence relationship analysis on GPU, among which a modified algorithm isproposed for relationship guiding tree by adjusting the execution paths. While the secondparallel technique is used to release the I/O load and synchronous processing by thedivision of input distance matrix, which has realized asynchronous collaborative process.Besides, the instruction stream based on GPU has achieves the dynamic multiplegranularity mechanism with a threshold to resolve the resources congestion and light.Moreover, data segmentation with asynchronous processing model has efficiently resolvedthe quick positioning of the minimums which is the most time-consuming process for theleast-distance pair of nodes.Based on Linux operating system and CUDA platform, with C、C++language, ourexperimental results show that an average speedup of25.1has been achieved with massivesequences compared with the CPU implementations. Besides, the parallel implementationscan achieve a steadier and faster performance.