CPU与GPU混合虚拟化资源高效调度策略

发布时间：2018-02-02 05:00

本文关键词： 虚拟化混合负载反馈控制服务等级协议资源调度　出处：《上海交通大学》2015年硕士论文　论文类型：学位论文

【摘要】：目前工业界,由多核CPU和拥有大量核数的GPU组成的系统架构已经成为云计算中处理混合负载的强有力的平台。然而多个虚拟机之间共享CPU和GPU资源依然存在亟待解决的难题。有效的GPU资源调度策略的缺乏造成GPU密集型负载的性能混乱和GPU资源的浪费。其次负载的异构性使CPU密集型的负载占用大量的CPU和IO资源,这会影响GPU密集型负载的性能。如果对每一个负载定义各自的服务等级协议(Service Level Agreement,简称SLA),那么在混合负载情况下它们的SLA是很难被保证的。最后,由于云计算数据中心需要和用户维持大量的网络连接,而有限的网络带宽资源成为影响用户体验的重要瓶颈。针对上述三个问题,本文提出了一种基于虚拟化的分布式拓扑结构,这是一个能够高效合理地处理GPU与CPU混合资源调度的轻量级框架。该框架由两部分组成,其中第一部分运行在单个物理节点中并负责GPU、CPU和网络带宽资源调度。它采用微软的钩(Hook)技术,通过拦截相关的应用程序编程接口(Application Programming Interface)对虚拟机内应用程序申请的资源进行调度,全过程不对客户操作系统、宿主操作系统、底层驱动程序以及负载程序进行任何修改。框架的第二部分是由独立的主控机器(master)运行,接受外来的用户请求并自动做出仲裁,向受控机器(slave)发出响应指令,最终该受控机器会自动启动新的虚拟机和用户建立起网络连接,响应用户请求。本框架集成了若干资源调度策略,即两个CPU资源调度算法和两个网络带宽资源调度算法。它们分别是开环控制调度策略、自适应控制调度策略、基于公平的资源调度策略、基于共享的资源调度策略。前两者针对CPU资源的调度,后两者针对网络带宽资源的调度。调度策略在实现中引入了自动控制技术,增强了负载运行时性能的稳定性和健壮性,有效控制混合负载的QoS水平。最后实验验证,本文所提出的框架适用于实际混合负载计算场景。主控机器有效地控制受控机器响应用户请求。框架集成的资源调度策略能有效地在运行多个GPU和CPU密集型负载的服务器上进行混合资源的调度,其性能符合设计目标。
[Abstract]:At present, industry. The system architecture composed of multi-core CPU and GPU with a large number of cores has become a powerful platform for handling mixed loads in cloud computing. However, the sharing of CPU and GPU resources among multiple virtual machines still exists. The lack of effective GPU resource scheduling strategy leads to the confusion of GPU intensive load performance and the waste of GPU resources. Secondly, the heterogeneity of the load makes CPU intensive load occupy a lot. For CPU and IO resources. This affects the performance of GPU intensive loads, if you define a service Level agreement for each load. In the case of mixed load, their SLA is difficult to guarantee. Finally, because the cloud computing data center needs to maintain a large number of network connections with users. Limited network bandwidth resources have become an important bottleneck to affect the user experience. In view of the above three problems, this paper proposes a distributed topology based on virtualization. This is a lightweight framework that can efficiently and reasonably handle the scheduling of GPU and CPU mixed resources. The framework consists of two parts, the first part of which runs in a single physical node and is responsible for GPU. CPU and network bandwidth resource scheduling. It uses Microsoft Hook technology. The application Programming interface is used to schedule the resources requested by the application in the virtual machine by intercepting the related application programming interface. The entire process does not make any changes to the client operating system, the host operating system, the underlying driver, and the load program. The second part of the framework is run by an independent master machine. Accept requests from outside users and automatically arbitrate and issue response instructions to the controlled machine, which will automatically start the new virtual machine and establish a network connection with the user. In response to user requests, this framework integrates several resource scheduling strategies, namely, two CPU resource scheduling algorithms and two network bandwidth resource scheduling algorithms, which are open-loop control scheduling strategies. Adaptive control scheduling strategy, based on fair resource scheduling strategy, based on shared resource scheduling strategy. The first two for CPU resources scheduling. The latter two are aimed at the scheduling of network bandwidth resources. The scheduling strategy introduces automatic control technology to enhance the stability and robustness of load runtime performance. Effectively control the QoS level of the mixed load. The framework proposed in this paper is suitable for the actual mixed load computing scenario. The master control machine can effectively control the controlled machine to respond to user requests. The integrated resource scheduling strategy of the framework can effectively run multiple GPU and CPU dense. Scheduling of mixed resources on a server with a centralized load. Its performance conforms to the design goal.
【学位授予单位】：上海交通大学
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：TP391.41;TP332

【相似文献】