Research And Implementation Of Data Prefetch-based Vector Line Parallel Visualization Algorithm In Large Scale Flow Field

Vector field visualization is an important method of scientific visualization,and the main methods are icon,vector line and visualization features.Vector line are widely used in engineering practice for its advantages in vision.As accuracy of scientific simulation growing,the data size of the vector field increase fast,and traditional visualization methods can't handle massive vector field visualization efficiently.Parallel visualization and out-of-core methods become effective means to solve the problem that the main memory of single node cannot satisfy all the flow field data in streamline visualization of large scale data.Due to the data dependent in vector line computation,data I/O is the bottleneck.How to decrease the delay in data I/O become a hotspot of research.In this paper,we present a new method of data prefetching based the particle's flowing between data blocks,and design a hybrid parallel visualization combined multi-cores and multi-nodes based data prefetching.The main work in this paper lists as following:1.Presenting a data prefetching method in vector line computing based on the particle's flowing between data blocks.The method uses statistic results of particle flowing in data block and blocks boundary to construct particles in blocks flowing probability graph(BFPG).When computing vector line,we can pre-fetch data needed for computing using BFPG.After analyzing the effect of different blocks partition sizes on hit rate of data prefetching,improve a new prefetching method called Frontiers flowing probability graph(FFPG)based on BFPG.As the results showing,the hit rate of BFPG and FFPG are higher comparing to other prefetching methods,usually improving 15%-25%.2.Presenting a hybrid parallel visualization algorithm based on data prefetching.The hybrid parallel visualization using multi-cores parallelism based on OpenMP and multi-nodes parallelism based on MPI to accelerate the vector line computing.In order to solve bottleneck of IO,design a distributed Key-Value data store scheme and improve data transforming effectiveness.The results show that when combining data prefetching method based on BFPG and FFPG,hybrid parallel visualization can get better speedup and higher scalability.3.On the basis of the work above,a hybird parallel visualization software has been designed and implemented.The software consists of data preprocessing module,data prefetching module,thread scheduling and management modules,vector line growing module and graphics rendering module.We use multiple sets of data to verify the correctness and validity of the proposed methods.