基于对端重叠网络的通用大规模计算资源共享环境的构造

The Construction of a Generic Large Scale Computation-Sharing Environment Based on Peer-to-Peer Overlay Network

大规模计算资源共享是网格计算和其它基于Internet的计算模式的首要目标之一 .作为计算网格的一个主要分支 ,自愿计算环境在大规模计算模式中正在扮演着越来越重要的角色 .目前的自愿计算环境大多是专用和集中控制的 ,因而存在着某些可用性和可扩展性的问题 .该文构造了一个通用和可扩展的自愿计算环境 :Paradrop per .该环境的可扩展性得益于一个自组织的对端重叠网络 ( peer to peeroverlaynetwork) ;通用性得益于统一的计算和任务描述手段以及建立在该网络之上的计算任务调度、容错机制和界面友好的使用平台 .该文首先对自愿计算的特点以及Paradropper计算平台做了简要介绍 ,然后重点阐述该环境下计算网络的构造和基于该网络的计算任务调度算法 ,对计算网络的若干特性 (诸如高聚合性和低搜索路径 )以及调度的效能进行了仿真分析 ,最后 ,通过若干示例应用对计算平台的功能进行了测试 .仿真结果说明 ,该文构造的计算网络表现出明显的自组织特性 ,具有较好的可扩展性和自组织能力 ,能为计算资源的调度提供很好的支持 .小规模测试结果初步说明 ,该文构造的计算环境为Internet环境下的单程序多数据 (SPMD)的易并行计算应用提供了一个可行的思路 .

Large scale resource sharing is the main object of gird computing and other Internet based computing modes. As a variant of computation grid, volunteer computing is playing an important role. Current volunteer computing systems are essentially center based, which leads to some scalability and accessibility problems. The goal of this paper is to build a generic and scalable volunteer computing environment, its scalability roots in a self organizing peer to peer overlay computing network, and its generity is based on a universal description mechanism of computing and task, a simple but efficient heuristic task scheduling algorithm, a fault tolerance mechanism and a user friendly GUI interface. This paper first briefly introduces the volunteer computing and its implementation, and then gives a detail explanation of the computing network constructing approach, including the simulations of small world network characteristics such as clustering coefficient and path length, and how these characteristics are selected in Paradropper computing task scheduling algorithm. Two small scale applications (Ray tracing and Mersenne prime number searching) are used to test Paradropper computing system. Simulation results show that, Paradropper computing network emerges outstanding self organizing characteristics and is very suitable for computing task scheduling. The preliminary small scale tests show that, compared to current center based global computing system such as Javelin, Paradropper environment probably leads to a new solution which could overcome the scalability and accessibility problems for internet based embarrassingly parallel applications.