String Kernel Parallel Implementations Based On GPU

String similarity matching is one of the most basic technologies in computer field, which is widely used in network security, information retrieval, biological information processing and text root recognition. However, with the ever-changing network development, network traffic is doubled and matching ranges from an exact match to the fuzzy matching. String matching processing is too slow to meet the requirements of today’s large-scale data sets. So there is an urgent need for a new efficient and higher quality algorithm.Nowadays GPU is in a continuous upgrading, so it is more and more powerful. NVIDIA Corporation launched CUDA architecture in2007, using the graphics card to solve the task outside of the image calculation. CUDA can be used to solve the complex problems in industrial, commercial and scientific computing. CUDA has many advantages:cheaper price, efficient parallel, compute-intensive, long pipeline. CPU plus GPU heterogeneous processing platform has gradually become a mainstream parallel solution.This article aims to take advantage of efficient parallelism of CPU and GPU to process large-scale data set more effectively. This paper first outlines the wide application range of SVM based on string kernels, discusses the characteristics of GPU architecture and CUDA proprietary platform. It gives CUDA thread scheduling, memory types, memory allocation in detail, and recalls the history of CUDA-based string processing at home and abroad. Furthermore, this article optimizes three classic string kernel algorithms-edit distance kernel, p-spectrum kernel, gap-weighted subsequence kernel on GPU. We assess the proposed system using the Reuters-21578and SpamAssassin public corpus and get the speedup from7to33.