Resource management in cloud data-centers

今天的数据中心应用包括Web搜索、金融服务、推荐系统和社交网络，这些应用要求在几分之一秒内完成请求，通常被称为“在线数据密集型应用”（OLDI）。这些要求也被称为软实时要求，对服务提供商有重大影响。例如，亚马逊报告说，延迟每增加100毫秒，业务收入就会减少1%。该提案的目的是开发算法，以减少在数据中心运行的应用程序/作业的响应时间。为了实现这一目标，需要深入了解数据中心中的多层协议交互。我们将为数据中心设计跨层协议栈，其中应用信息被传递到传输协议，使用SDN记录在交换机中，并用于调度流，以满足软实时应用边界。**我们的研究活动将包括设计基于MapReduce分区/聚合工作流的数据偏斜的流的优先级调度，以及设计具有虚拟机放置的应用程序准入控制。我们还将评估在交换机转发表中设置条目所需的时间，并制定改进方案。我们将在数据中心网络中部署SCTP（流控制传输协议）作为传输协议，并开发数据中心应用程序之间的安全隔离方案。*

Today's data center applications include web search, financial services, recommendation systems and social networking which require request completion within a fraction of a second and they are usually called "online data intensive applications" (OLDI). Those requirements are also known as soft real-time requirements and have significant impact on service providers. For example Amazon reported that every 100ms increase of latency reduces business revenue by 1%.***The aim of this proposal is to develop algorithms for decreasing response time of applications/jobs running in data centers. In order to achieve that, deep understanding of multi-layer protocol interactions in data centers is needed. We will design cross layer protocol stack for data centers where application information is passed to transport protocol, recorded in switches using SDN and used to schedule the flows in order to meet soft-real time application bounds.******Our research activities will include designing priority scheduling of flows based on data skew coming from MapReduce partition/aggregate workflow and designing application admission control with placement of virtual machines. We will also evaluate time needed to set entries in switch forwarding table, and develop schemes for their improvement. We will  deploy SCTP (stream control transmission protocol) as transport protocol in data center network and develop schemes for security isolation among applications in data center.***********