Policy-Based Management of Cloud Systems

云基础设施提供的资源可以由来自不同组织的多个应用程序共享。根据峰值需求分配计算资源往往会导致计算资源利用率不足。解决这个问题的一种方法是允许超额认购。假设所有应用程序的峰值需求可能超过云基础设施提供的可用计算资源，但峰值需求通常不会同时出现。但是，单个服务器、机架或集群上的应用程序的峰值需求完全有可能超过其容量。这就要求动态地、及时地分配资源，以响应云基础设施任何部分的过载。这被称为动态资源管理。拟议的工作正在解决几个挑战。其中包括以下内容：(1)动态资源管理要求解决与目标冲突所产生的问题。例如，我们希望最小化功耗。但是，主要关注功耗可能会导致性能问题。对性能管理的关注可能会导致电力消耗的增加。(2)性能问题的检测与原因的识别是不一样的。重要的是要解决这个问题，以便在重新分配资源方面作出正确的决定。(3)越来越多的客户拥有具有多个软件组件的应用程序。在重新分配资源时，如果其他软件组件依赖于一个软件组件，我们必须考虑对它迁移的影响。(4)可以采取多种行动来解决超载问题。其中一个挑战是了解在特定情况下哪种行动是最好的。(5)实现资源管理算法的软件开发需要简化。(6)对于类似的过载情况，不同的云提供商会做出不同的决策。决策将基于云提供商的业务目标，这些目标因提供商而异。解决这些问题对于下一代云基础设施非常重要，因为下一代云基础设施需要为客户端应用程序保持预期的性能，同时又要节能。

The resources provided by a cloud infrastructure can be shared by multiple applications from different organizations. Allocating computing resources based on peak demand often results in underutilized computing resources. One approach to dealing with this is to allow for over-subscription. The assumption is that the peak demand of all applications may be more than available computing resources provided by the cloud infrastructure but the peak demand typically does not occur simultaneously. However, it is entirely possible that the peak demand of applications on a single server or a rack or a cluster may exceed the capacity. This requires that resources be allocated dynamically and in a timely fashion in response to overload on any part of the cloud infrastructure. This is referred to as dynamic resource management. There are several challenges being addressed with the proposed work. Some of these include the following:(1) Dynamic resource management requires that issues that arise with conflicting goals is addressed. For example, it is desirable to minimize power consumption. However, a primary focus on power consumption may lead to performance problems. A focus on performance management may lead to increased power consumption. (2) The detection of performance problems is not the same as the identification of the cause. It is important to address this in order to make the correct decisions with regard to re-allocation of resources.(3) Increasingly clients have applications with multiple software components. In re-allocating resources we must consider the impact on migrating a software component if other software components depend on it. (4) There are multiple actions that can be taken to address overload. One challenge is learning which action is best for a particular scenario.(5) The development of software to implement resource management algorithms needs to be made easier.(6) Different cloud providers will make different decisions for similar overload situations. The decisions would be based on the cloud provider’s business goals, which vary from provider to provider.Addressing these problems is important for the next-generation of cloud infrastructures that need to maintain expected performance for client applications and yet be energy efficient.