Julius C. B. Leite

Power optimization for dynamic configuration in heterogeneous web server clusters

To reduce the environmental impact, it is essential to make data centers green, by turning off servers and tuning their speeds for the instantaneous load offered, that is, determining the dynamic configuration in web server clusters. We model the problem of selecting the servers that will be on and finding their speeds through mixed integer programming; we also show how to combine such solutions with control theory. For proof of concept, we implemented this dynamic configuration scheme in a web server cluster running Linux, with soft real-time requirements and QoS control, in order to guarantee both energy-efficiency and good user experience. In this paper, we show the performance of our scheme compared to other schemes, a comparison of a centralized and a distributed approach for QoS control, and a comparison of schemes for choosing speeds of servers.

Intelligent context-aware monitoring of hypertensive patients

We present a decision-level data fusion technique for monitoring and reporting critical health conditions of a hypertensive patient at home. Variables associated to the patient (physiological and behavioral) and to the living environment are considered in the solution, contributing to improve the confidence on the system outputs. In the paper, we model the problem variables as fuzzy, aiming to capture their intrinsic essence, and draw rules based on medical recommendations to identify the health condition of the patient. This initiative move towards to build an abstract framework for context-aware telemonitoring applications. We also describe the relevant components of the framework and provide an initial evaluation of its decision component. Our results demonstrate that a principled choice of rules and variables may lead to a consistent identification of critical patients conditions.

Statistical QoS Guarantee and Energy-Efficiency in Web Server Clusters

In this paper we study the soft real-time web cluster architecture needed to support e-commerce and related applications. Our testbed is based on an industry standard, which defines a set of Web interactions and database transactions with their deadlines, for generating real workload and bench-marking e-commerce applications. In these soft real-time systems, the quality of service (QoS) is usually defined as the fraction of requests that meet the deadlines. When this QoS is measured directly, regardless of whether the request missed the deadline by an epsilon amount of time or by a large difference, the result is always the same. For this reason, only counting the number of missed requests in a period avoids the observation of the real state of the system. Our contributions are theoretical propositions of how to control the QoS, not measuring the QoS directly, but based on the probability distribution of the tardiness in the completion time of the requests. We call this new QoS metric tardiness quantile metric (TQM). The proposed method provides fine-grained control over the QoS so that we can make a closer examination of the relation between QoS and energy efficiency. We validate the theoretical results showing experiments in a multi-tiered e-commerce web cluster implemented using only open-source software solutions.

Attaining soft real-time constraint and energy-efficiency in web servers

M/M/1 queues have been traditionally used to model several systems like phone calls at a call center, banking services and so on. However, recent studies showed that it does not model properly a web server system. In this paper we investigate the impact of this assumption in providing timeliness constraint in an energy-efficient web server. Although energy efficiency is a key issue, it should not be attained at the expense of a poor quality of service. The work proposed here describes a technique that uses queueing theory results to balance energy consumption and adequate application response times in heterogeneous CPU-intensive server clusters. Moreover, we investigate the I/O impact on a purely CPU-oriented energy saving strategy. This proposal shows that the assumption of a Poisson process is a good approximation to model a web server.

A decision-making mechanism for context inference in pervasive healthcare environments

This paper presents a Fuzzy approach to health-monitoring of patients in pervasive computing environments. A decision model considers three classes of variables that represent the context information being collected: environmental, physiological, and behavioral. A case study of blood pressure monitoring was developed to identify critical situations based on medical knowledge. The solution maintains the interpretability of the decision rules, even after a learning phase which may propose adjustments in these rules. In this phase, the Fuzzy c-Means clustering was chosen to adjust membership functions, using the cluster centers. A medical team evaluated data from 24-h monitoring of 30 patients and the rating was compared with the results of the system. The proposed approach proved to be individualized, identifying critical events in patients with different levels of blood pressure with an accuracy of 90% and low number of false negatives.

Generalized Tardiness Quantile Metric: Distributed DVS for Soft Real-Time Web Clusters

Performing QoS (Quality of Service) control in large computing systems requires an on line metric that is representative of the real state of the system. The Tardiness Quantile Metric (TQM) introduced earlier allows control of QoS by measuring efficiently how close to the specified QoS the system is, assuming specific distributions. In this paper we generalize this idea and propose the Generalized Tardiness Quantile Metric (GTQM). By using an online convergent sequential process, defined from a Markov chain, we derive quantile estimations that do not depend on the shape of the workload probability distribution. We then use GTQM to keep QoS controlled in a fine grain manner, saving energy in soft real-time web clusters. To evaluate the new metric, we show practical results in a real web cluster running Linux, Apache, and MySQL, with our QoS control and for both a deterministic workload and an e-commerce workload. The results show that the GTQM method has excellent workload prediction capabilities, which immediately translates in more accurate QoS control, allowing for slower speeds and larger energy savings than the state-of-the-art in soft real-time web cluster systems.

Treating uncertainty in distributed scheduling

Abstract In distributed systems the scheduler of an overloaded node may choose to transfer the execution of one or more tasks to other less busy nodes, in order to minimize their expected service times, or to increase the number of tasks that meet their deadlines, among other criteria. One solution makes use of Bayesian theory to infer the load state of the system and, based on this information, the scheduler of a busy node chooses an “appropriate” node to transfer a task too. The meaning of “appropriate” will be a function of the objectives established in the adopted location policy. In the Bayesian decision method, objectives are represented by a utility function. However, the development of a utility function can be a tricky and somewhat subjective task. In this paper, we describe a new approach that easily maps transfer objectives into useful mathematical expressions, representing location policy objectives by fuzzy sets. The proposed approach was successfully employed to add objectives to a Bayesian decision-based algorithm improving the number of tasks that are executed over time in a distributed real-time system.

Load forecasting applied to soft real-time web clusters

Dynamic configuration techniques such as DVFS (Dynamic Voltage and Frequency Scaling) and turning on/off computers are well known ways to promote energy consumption reduction in web server clusters. This paper demonstrates how the application of forecasting methods improves energy savings in a soft real-time application, and compares it with other energy aware methods. Instead of a synthetic workload, a real traffic pattern was used to make the experiments more realistic. Our system promotes energy reduction while maintaining users satisfaction with respect to deadlines being met. The results obtained show that prediction capabilities increase the QoS of the system, while maintaining or improving the energy savings over state-of-the-art power management mechanisms.

On the Integration of Configuration and Meta-level Programming Approaches

Configuration Programming, based on Architecture Description Languages, and Meta-Level Programming are considered promising approaches in the software engineering field. This paper shows that there is an immediate correspondence between some key concepts of Configuration and Meta-Level Programming approaches and that some of the main issues to be solved for their deployment infreal systems are quite similar. The main result is that the integration of both approaches in a single configuration programming framework can assist in employing meta-level programming in order to achieve separation of concerns and improve software reuse. In addition, the capability of supporting dynamic configuration and flexibility on component programming language choice are potentially improved. A prototype of a configuration programming centered environment and some application examples are presented in order to demonstrate the useful features of the combined approach.

Experience with a New Architecture Review Process Using a Globally Distributed Architecture Review Team

We present in this paper our experience with applying a new architecture review process that uses a globally distributed review team to assess architecture risk of a complex mission critical system. The new architecture review process uses aspects of the checklist-based architecture review process and the operational scenario-based architecture review process. We present the architecture review process approach, a summary of the architecture under review and the detailed analysis of the most important operational scenarios. We conclude by presenting a summary of the lessons we learned using the new process.