动态时变约束下的赛百平台资源优化理论与算法研究

项目名称： 动态时变约束下的赛百平台资源优化理论与算法研究

项目编号： No.60803017

项目类型： 青年科学基金项目

立项/批准年度： 2009

项目学科： 金属学与金属工艺

项目作者： 曹军威

作者单位： 清华大学

项目金额： 18万元

中文摘要： （1）致力于研究虚拟资源自适应供给问题。将弹性云计算平台建模成多层次，内部并联外部串联的队列，提出针对任意结构模式下表达用户请求到达率和平均响应时间的逻辑关系，并因此灵活自适应和弹性提供虚拟资源。实验结果证实仿真模型和真实实验结果比较一致； （2）针对队列模型过于简化运行期的随机性因素问题，引入仿真优化粗糙模型的思想，提出了一种低开销仿真优化调度方法从而实现虚拟资源的敏捷供给。实验结果证实了本低开销调度方法能够减少仿真优化阶段的调度时间，并能获得具有良好性能的满意调度策略； （3）针对弹性云计算平台的多阶段、实时性和动态性情况，充分利用低开销的调度优势，构建基于连续仿真优化方法的多阶段迭代调度模型，理论和实验均证实了方法的低开销特性。基于此优势，在更短时间能做出的决策方案能够更加细粒度的捕获工作负载强度的局部特性，整体上提升系统的调度性能； （4）针对弹性云计算平台多任务负载的局部特性，分析了相邻阶段多任务负载的相似性特点从而自适应的动态划分决策阶段，将迭代仿真优化方法拓展成一种进化仿真优化方法，从而自适应的在多任务负载变化剧烈和和平缓处分割和合并调度区间，实现整体仿真调度性能的提升。

中文关键词： 赛百平台；弹性云计算；序优化；虚拟资源管理；自适应供给

英文摘要： (1) Adaptive provisioning of virtual resources is firstly discussed. It models the Elastic Computing Platform (ECP) as a multi-tier, internally parallel and externally sequentially connected queue, and provides the logical relationship between the arrival rate of user requests and average response time, and provisions virtual resource adaptively, flexibly and elastically. Experiments verify the queuing model matches with real benchmarking results. Furthermore, it is substantiated that the method outperforms traditional utilization based method, both in terms of the satisfaction of SLA and cost minimization; (2) Due to the over simplification of the runtime conditions of the queuing model, it brings in the concept of simulation based optimization, proposes a Low Overhead Scheduling (LOS) based method to provision virtual resources agily and elasticly. Simulation results show this method narrows down the simulation time in scheduling stage, and delivers good-enough schedules to apply; (3) Multi-stage, real-time and dynamic provisioning as it is in ECP, it utilizes the loosely coupled multitasking essence, and constructs a multi-phase iterative simulation based scheduling model, and proves the LOS both in theoretical and experimental verifications. Based on this advantage, it captures the fine-grained characteristics of workload and the scheduling performance is overall improved. (4) Characterizing the local patterns in multitaskingworkload and analyzing the similarities between inter-phase and intra-phase, it generates decision stage automatically, thereby extending the iterative LOS into an evolutionary LOS method. It adaptively partitions or merges scheduling phase when workload similarity becomes large or small to improve performance to a higher level.

英文关键词： Cyberinfrastructure; Elastic Cloud Computing; Ordinal Optimization; Virtual Resource Management; Adaptive Provisioning