Workshop Program

Although a mobile cloud computing paradigm has obtained significant attentions from research community, we note that most of work is based on an ad hoc fashion. Furthermore, little work has shown a model-based cost optimization of offloading. Mobile cloud computing may instead be holistically analyzed and systematically designed as distributed solutions to some global optimization problems. Such a paradigm enables optimized code offloading of mobile applications, where mobile devices can be thought of a waypoint of powerful cloud resources. This paper tackles a suite of optimization subproblems: a program partitioning, network resource allocation, network selection, and cloud resource allocation problem. The key objective is to satisfy the mobile application’s quality of service requirements by quantifying the performance of each subsystem: mobile clients, wireless network medium, and cloud services. By extensive experiments, we present mobile clients can have up to 73.69x and 39.69x offloading benefit in terms of time and energy.

The past few years have been witnessing a surging demand for cloud computing services, resulting in a huge carbon footprint and making energy cost one of the top operational costs of data centers. Meanwhile, as sustainable computing has become increasingly important, data centers are constantly pressured to cap the long-term usage of their energy produced from carbon-intensive sources (a.k.a., “brown” energy). In this paper, we study energy budgeting and propose a novel online resource management algorithm, called ORM, to control the number of active servers for delay-sensitive workloads in a data center for minimizing the operational cost while satisfying the energy capping constraint. We rigorously prove that ORM achieves a close-to- minimum operational cost compared to the optimal offline algorithm with future information, while bounding the potential violation of energy capping, in an almost arbitrarily random environment. We also perform a trace-based simulation study to complement the analysis and validate the effectiveness of ORM.

Traffic to a Web site can vary dramatically. At the same time it is highly desirable that a Web site is reactive. To provide crisp interaction on thin clients, 150 milliseconds has been suggested as an upper bound on response time. Unfortunately, the popular Apache Web server is limited in its capabilities to be reactive under varying traffic. To address this problem, we design in this paper an adaptive controller for the Apache Web server. A modified recursive least squares algorithm is used to identify system dynamics and a minimum degree pole placement controller is implemented to adjust the maximum number of concurrent connections. Experimentally, we show that the controller effectively regulates the reply time of HTTP connection requests, and hence provides reactive response, by limiting the maximum number of connections accepted by an Apache Web server.

The performance of a database can significantly deteriorate due to the fragmentation of data/index files. Manual database defragmentation and performance optimization remain time consuming and even infeasible as it requires knowledge of the complicated behavior of fragmentation and its relationships with system parameters. We propose a model-based detection and management framework for the database fragmentation which can automatically optimize database performance, detect the fault existence, estimate its future impact on system performance and recover the system back to normal. A predictive controller is designed to take proper actions to guarantee the QoS and remedy faults. Experimental studies on a realistic test-bed show the applicability and effectiveness of our approach.

This paper articulates three main challenges for employing feedback control with humans in the loop. They are: (i) the need for a comprehensive understanding of the complete spectrum of the types of human-in-the-loop controls, (ii) the need for extensions to system identification or other techniques to derive models of human behaviors, and (iii) most importantly, determining how to incorporate human behavior models into the formal methodology of feedback control.

Today’s cloud data center infrastructures are not even near being able to cope with the enormous and rapidly varying capacity demands that will be reality in a near future. So far, very little is understood about how to transform today’s data centers (being large, power-hungry facilities, and operated through heroic efforts by numerous administrators) into a self-managed, dynamic, and dependable infrastructure, constantly delivering expected QoS with reasonable operation costs and acceptable carbon footprint for large-scale services with sometimes dramatic variations in capacity demands. In this paper, we discuss some of the major challenges for resource-optimized cloud data center. We propose a new research area called Cloud Control, which is a control theoretic approach to a range of cloud management problems, aiming to transform today's static and energy consuming cloud data centers into self-managed, dynamic, and dependable infrastructures, constantly delivering expected quality of service with acceptable operation costs and carbon footprint for large-scale services with varying capacity demands.

A generic linear mathematical model is proposed to represent the dynamics of bandwidth usage, network topology, and host processing power in large scale Peer-to-Peer (P2P) networks. Feedback control theory is employed to analyze system stability and query controllability, as well as deriving an explicit solution for the bandwidth usage. The proposed model and analysis methods can be applied to various P2P networks, such as broadcast based P2P network, super-peer based unstructured P2P network, and distributed hash table based structured P2P network. The synthesis of the model is also presented, which adaptively adjusts query rate in order to properly control the bandwidth usage.

An important challenge in the design and implementation of self-optimizing systems is that of finding a model that maps changes in a tunable parameter (or “knob”) into an effect on the performance, power, or energy, of the overall system. This paper describes ARPE (Analyzing the Relationship between Parameters and Effectors), an open source tool to analyze the effect of parameter changes on the behavior of applications in a complex system with interrelated knobs.

We evaluate ARPE in several case studies on real systems with different sensors and parameters. Our results show that ARPE can help determine the best sensors for a system designed to predict application execution time. For space limitations, only one case study is here shown, demonstrating that the error of modeling energy consumption is limited to the range 0.1–10% for previously unseen benchmarks.