Research On Performance Acceleration Of DEFLATE Algorithm With Collaborative Computing Of CPU And GPU

Heterogeneous computing is the main trend of high performance computing. Followingthis trend, GPU general purpose computing has been given the best developing opportunityever. GPU general purpose computing is based on the features of graphics processing hard-ware, using its High FLOPS, suitability for large-scale and high-density parallel data pro-cessing to fulfill the general purpose computing task in non-graphical field. With the devel-opment of GPU programming technology, using GPU in cooperation with CPU to buildhigh-performance and low-cost computing platform has gained great prospects.Today is the times of the information technology and mobile communication; thelarge-data processing brings too much storage and transmission problems. The role which da-ta compression play, and its social, economic efficiency will become increasingly. If there isno data compression, it’s hard for data storage and transmission. The advantages of using datacompression: data compression is not only for the purpose of saving storage space. However,another very important meaning of it is reducing the communication delay in data transmitting;furthermore, data compression also play a significant role in saving communication band-width and resource consumption.There have been many popular lossless compression algorithms: DEFLATE, BZIP2,LZMA, LZMA2, etc. DEFLATE has the fastest compression speed of them. Based on testingon DEFLATE algorithms, however, the compression speed of it is still not so satisfactory.This paper is for optimizing the DEFLATE algorithm and researching the method to use GPUfor improving the performance of software.We choose the DEFLATE implementation in GZIP software in this paper, because thisversion of DEFLATE implementation is the closest to the description to the RFC1951docu-ment. As with optimization design, we’ll use CPU in collaboration with GPU to accelerate thisDEFLATE, which mainly include designing a new pipeline system and distribute workloadsthat are suitable for GPU computing to GPU. According to testing result, implementation ofthis solution can obviously accelerate some of the test cases.In this paper, while realizing an implementation of CPU and GPU computing design toaccelerate the DEFLATE algorithm, it also includes relatively deep research and analysis ofthe following contents:1. It’s the research work of GPU architecture and GPU programming techniques. Wedid a systematic analysis of the features of the GPU architectures of both NVIDIAand AMD. Then we described the history of GPU programming techniques and how to program with OpenCL.2. Following is the description of data compression techniques. Based on BZIP2andDEFLATE algorithms, we analyzed data compression theories and common com-pression techniques, which include a detailed description about the Huffman entropyencoding in the LZ77compression encoding part of DEFLATE algorithm.3. Analysis of GZIP source code. It’s mostly about its code architecture, implementationdetails of its key functions and the optimization possibility analysis and etc. In theend, we built a reasonable optimization solution based on this analysis and imple-ment our code in the GZIP.