Autonomic Performance and Power Control for Co-Located Web Applications in Virtualized Datacenters

Reviewer: Balint Molnar

The proliferation of cloud computing and virtualized environments raises several performance issues regarding load balancing and power consumption. In order to create a manageable structure, an appropriate architecture should be defined. The paper proposes a multitier and multiservice application architecture for representing the components and a distributed controller solution in the form of middleware for virtualized computing environments. The proposed solution is a distributed control approach that optimizes the resource allocations, performance parameters as response time, central processing unit (CPU), and memory utilization. The deep-rooted property of a virtualized environment that is present as a natural part of the virtualized computing surroundings is that the virtual machines, applications, and services can freely move among or above computing resources that are shared among several participants that are co-located in cyberspace; that is, they are placed near to each other to utilize the same resources. The paper describes an architecture approach that helps organize the components in a virtualized environment into such a structure that lays the foundation for creating a set of equations that can represent the resource utilization. The proposed set of equations steps beyond the traditional approaches that reflect the interrelationships between applications and resource utilization. One of the vectors is a regression vector for applications that represents the recent and previous outputs of applications and the effect on the application output in the next control interval in discrete time. Naturally, there is a matrix containing the necessary regression parameters. This basic approach is extended with a fuzzy model and machine learning algorithm. The objective of the fuzzy model is to grasp the facet of the nonlinearity of the system. A machine learning algorithm (a kind of clustering) is used for determining the fuzzy rules and for tuning the fuzzy parameters. Moreover, the paper defines a forgetting factor for fine-tuning that yields exponentially less weight to older error samples. This tuning has importance in the case of multiservice architectures because it shows more complex behavior considering performance and resource allocation. The optimization problem can be formulated as a quadratic programming problem that should be computed at a certain point of time with the frozen parameters for the given interval. The paper concludes with measurements, statistical data, and graphs to prove that the proposed distributed controller and the algorithm outperform other solutions. Because the research is published in a peer-reviewed journal, the results seem to be convincing and feasible. The paper is interesting for researchers who are involved in the most recent research on performance issues, especially in virtualized and cloud environments. Online Computing Reviews Service