Research On GPU-based Parallel Computing On BitTorrent Protocol

With the fast develop of internet, Peer-to-peer (P2P) content distribution systems, throughthe cooperation orientation and distribution file-sharing, such as file-sharing (BitTorrent)、livestreaming (PPLive) and video-on-demand (Joost) systems have become enormously popular ontoday’s internet. These systems organize tens of thousands or even millions participating userhosts into a randomized mesh topology, where each user maintains connection with a subset ofother peer users, sharing the desired data. In a typical file-sharing scenario like BitTorrent, alarge file is broken down into K granularity blocks. The goal is to distribute all the blocks to allthe users by letting peers exchange these blocks in a decentralized fashion. To attain this goal,each peer has to decide which neighboring peer to transmit to and which block to transmit at agiven time. Based on these questions, distributed block scheduling and neighbor selectionprotocols in P2P networks have become a very active research area in recent years.At the present stage the implementation of distributed block scheduling and neighborselection algorithms in the BitTorrent file-sharing protocol are based on CPU. It’ll need arelatively long time to distribute a big file to all the peers in the large-scale network. This paperput forward and improves P2P algorithm based on GPU parallel computing. Because thealgorithm is suitable for the independence peers in the BitTorrent protocol, so the improvedalgorithms speed up the file-share efficiency in the system. In this paper, the study content andmain contribution as follows:(1) RUB Algorithm Parallel Computing Based on GPU. This algorithm adopted threeschemes, the implementation of RUB algorithm based on the global memory for single-threadedsingle-node、 the implementation of RUB algorithm based on the sharing memory forsingle-threaded single-node and the implementation of RUB algorithm based on the globalmemory for single-threaded single-datablock. Because accessing global memory will producehigher delay, leading to low efficiency of the RUB algorithm. In order to overcome the abovedisadvantages, in the second scheme improve to using sharing memory. In certain improved theefficiency of the RUB algorithm. But the sharing memory space is very small; it can’t store the big file and in a certain way affect the algorithm efficiency. Therefore in this paper give the thirdscheme, using the implementation of RUB algorithm based on the global memory forsingle-threaded single-datablock. Because in the third scheme reduced the amount of work ofevery thread, make the efficiency of RUB algorithm greatly improved.(2) LRF Algorithm Parallel Computing Based on GPU. This algorithm adopted twoschemes, the implementation of LRF algorithm based on the global memory for single-threadedsingle-node and the implementation of LRF algorithm based on the global memory forsingle-threaded single-datablock. Similar to the RUB algorithm implementation, In order to gethigher LRF algorithm efficiency, improved the implementation of LRF algorithm based on theglobal memory for single-threaded single-node，using implementation of LRF algorithm basedon the global memory for single-threaded single-datablock and greatly improved the efficiencyof this algorithm.Finally, the paper summarizes the research work, and points out the content and direction offuture research.