Design And Research About Network Parallel Computing On ICT Image Reconstruction

Traditional parallel computing depends on the special-purpose and large-scale parallel computer. It adopts the memory and shares way, so it has high computational efficiency; Because of complicated programming, bad expansibility and expensive costs, the large-scale traditional parallel computer mainly applies to the professional fields such as weather forecast, etc. With the development at full speed of micro-processing and network technology, common network, such as the Local Area Network (LAN), is utilized to construct Cluster of Workstations (COW) by joining the high-performance workstations or top-grade PCs according to a certain structure. Base on COW, integrating modern communication technology and traditional parallel technology, a new knid of parallel technology—network paprallel computing is forming. Because it has some conspicuous advantages such as high systematic cost performance, easy programming, fine expansibility, fully utilizing existing resource and so on, it widely used in the field of parallel computation. In the field of network parallel computing, a lot of research of the algorithm (include parallel computing flow and task scheduling according to the construction of netware COW, etc.) and software (data communication in COW, Congestion processing,etc.) are still at the stage of exploring, and in technology they are still immature. There are huge difficulties to study the general parallel algorithm, and in practical use, it can't be used directly. So to practical application of concrete system structure, the research of the parallel algorithm is expected to get better development. In the field of Computed Tomography (CT), high energy X ray industrial CT that can measure the large scale industrial components has obvious importance to such fields as aviation and spaceflight industry. When the detection target of industry CT is large scale industry component and high-resolution is required, the reconstruction image matrix is enormous, the calculating amount increases greatly, thus the time of image reconstruction is too long. On the other hand, at the situation of incomplete projection data, the reconstruction image quality of Filter Back Projection (FBP) algorithm is difficult to up to people's requisition. At that moment, we generally select algebra algorithms such as Simultaneous Algebraic Reconstruction Technique (SART), etc. But compareing with FBP, the reconstruction time of SART increases greatly. In the real-time case, it is very difficult to get practical application. For this reason, this thesis leads new developing network parallel computing into the industry CT image reconstruction. This thesis regards new developing COW as the main research object, and has discussed the parallel computing and industry CT technology in detail; It has discussed the parallel algorithm, load balanced tactics and a design model of the parallel algorithm--C3 (Computation, Communication, Congestion) model, and the theoretical foundation and scanning measure way of industry CT especially. On the basis of key research on serial SART algorithm, parallel SART algorithm has been designed; By discussing the most common used load balancing tactics recently in parallel computing in detail, and the problem to the practical application of this thesis, the adaptive load beforehand-distribute tactics has been designed. Through discussion on the problem to the practical application of the thesis, the parallel computing platform has been built on the base of local area network; On this platform, the software structure of this application problem has been designed by adopting the mode of Master/Slave; In VC ++6.0 environment, it has exploited the visual manage platform of network parallel computing and parallel SART algorithm software; Finally, it analyses and discusses the accelerated ratio and efficiency of the parallel SART algorithm with the real test data. Through theoretical analysis and the real test data, it has been proved that the parallel SART algorithm, the adpative load beforehand-distribute tactics, and the network parallel computing platform can effectively improve the time efficiency of SART, so there is significant meaning in improving the practical application of SART in industry CT; It also plays an important role in other similar application problems.