Towards Power Consumption Modeling for Servers at Scale

Centros de dados existentes hoje não estão prontos para apoiar as organizações de TI para atender às demandas de negócios em constante mudança. Assim, a próxima geração de centro de dados como Software Defined Data Center (SDDC) é explorada e deverá vir para ajudar. No entanto, o SDDC é relativamente novo desde a sua criação, em 2012, em que existem diferentes interpretações iniciais sobre a sua definição, os critérios, a arquitetura e os valores que SDDC traz referência. Há também limitada literatura e compartilhamento sobre como o SDDC funciona. O objetivo deste estudo é esclarecer a respeito da definição operacional do SDDC, critérios, arquitetura de referência, representação em obras como SDDC em três cenários, bem como padronizou os valores que ela traz. Além disso, alguns fatores para guiar as organizações de TI como adotar SDDC também são discutidos. Este estudo teve uma abordagem qualitativa em que a literatura SDDC é revista e alguns profissionais de TI SDDC são entrevistados. Por fim, as limitações do estudo, futuraspesquisas e conclusões também são fornecidos.

MuMMI (Multiple Metrics Modeling Infrastructure) environment is an infrastructure that facilitates systematic measurement, modeling, and prediction of performance, power consumption and performance-power tradeoffs for parallel systems. MuMMI builds upon three existing frameworks: Prophesy for performance modeling and prediction of parallel applications, PAPI for hardware performance counter monitoring, and PowerPack for power measurement and profiling. In this paper, we present the MuMMI framework, which consists of an Instrumentor, Databases and Analyzer. The MuMMI Instrumentor provides automatic performance and power data collection and storage with low overhead. The MuMMI Databases extend the databases of Prophesy to store power and energy consumption and hardware performance counters' data with different CPU frequency settings. The MuMMI Analyzer extends the data analysis component of Prophesy to support power consumption and hardware performance counters, and it entails performance and power modeling, performance-power tradeoff and optimizations, and web-based automated modeling system. Currently, our MuMMI online automated performance and power modeling system uses four modeling techniques: curve fitting, parameterization, kernel coupling and performance-counters-based, we discuss the effort to automate the process of developing performance and power models for scientific applications online, and focus on exploring performance-counters-based performance and power modeling. The MuMMI environment is able to aid in performance and power data measurement, storage, modeling and prediction of scientific applications on XSEDE resources in XSEDE community.

We investigate optimal power and performance management for heterogeneous and arbitrary cloud servers in a data center. In particular, we study the problems of power-constrained performance optimization and performance-constrained power optimization in a data center with multiple heterogeneous and arbitrary servers. These problems are essential to find optimal server speeds, such that: 1) the average task response time is minimized, and that the total power consumption does not exceed certain power constraint or 2) the total power consumption is minimized, and that the average task response time does not exceed certain performance constraint. Each server is treated as a G/G/1 queuing system, whose task interarrival times and task execution requirements can have arbitrary probability distributions. Furthermore, these servers are entirely heterogeneous in terms of task interarrival time, task execution requirement, and power consumption models. The main contributions of this paper are summarized as follows: 1) we formulate the average task response time as well as the total power consumption in a data center with multiple heterogeneous and arbitrary servers as the functions of server speeds; 2) we define our optimization problems by finding optimal server speeds, since the server speeds determine both the average task response time and total power consumption; 3) we develop algorithms to find the optimal solutions and demonstrate numerical data; and 4) we also develop several closed-form heuristic solutions and compare their quality with that of the optimal solution. Our approach provides an analytical way of studying the power-performance tradeoff at the data center level.

Finding the best energy-performance tradeoffs for High Performance Computing (HPC) applications is a major challenge for many modern supercomputing centers. With the increased focus on data center energy efficiency and the emergence of possible data center power constraints, making the right decision at a given time is becoming more important. A real-world situation like can a given 1000 compute node application be executed at a maximum of 2.7 GHz CPU frequency without going over the energy provider defined power band, or the available monthly energy limit? is just one example of the types of decisions HPC data centers will face. The previously developed Adaptive Energy and Power Consumption Prediction (AEPCP) model answers this question for the case of a fixed CPU frequency. This paper will extend the AEPCP process to enable the development of analytical models for estimating application execution time, power, and energy consumptions as functions of the number of compute nodes and maximum operating CPU frequency. Based on these analytical models, an adaptive model (Lightweight Adaptive Consumption Prediction (LACP)) is presented that implements the extended prediction process. This information allows for improved estimation of potential energy-performance costs and tradeoffs of applications and thus identifies the optimal resource configuration for specific data center boundary conditions.

The cost‐performance tradeoff is a fundamental issue in a data center for cloud computing, which is closely related to two key metrics that both cloud consumers and service providers care the most, that is, quality of service and cost of service. While there are different definitions of quality of service, the average response time is a common choice of performance metric. While there are various considerations in cost of service, the average power consumption is a common choice of cost metric. Hence, the cost‐performance tradeoff becomes the power‐performance tradeoff. In this article, we deal with the power‐performance tradeoff at the data center level. We study cost‐performance ratio optimization by using the techniques of workload management and server speed setting. In particular, we make the following tangible contributions. We solve three optimization problems, that is, (1) the workload management problem—to find a workload distribution, such that the cost‐performance ratio is minimized; (2) the server speed setting problem—to find a server speed setting, such that the cost‐performance ratio is minimized; (3) the workload management and server speed setting problem—to find a workload distribution and a server speed setting, such that the cost‐performance ratio is minimized. All the three optimization problems are analytically defined as multivariable optimization problems based on M/M/m queueing systems for multiple heterogeneous multiserver systems, together with two power consumption models, that is, the idle‐speed model and the constant‐speed model. Our approach makes it possible to quantitatively evaluate and optimize the cost‐performance ratio of a data center within a rigorously developed framework. Each multivariable optimization problem is transformed to a nonlinear system of equations. Due to the sophistication of these equations, they are solved algorithmically by a numerical procedure. Furthermore, we provide approximate, accurate, and analytical solutions to the first two problems. Performance data are demonstrated for each problem, and the accuracy of our approximate solutions are also discussed. To the best of the author's knowledge, this is the first paper which analytically and algorithmically minimizes the cost‐performance ratio of a data center with multiple heterogeneous multiserver systems using the techniques of workload management and server speed setting.

Reaming operation is a commonly used finishing phase for already drilled hole. Finishing is required because surface roughness of hole plays a significant role towards the functionality of the component. Surface roughness is a critical parameter for fatigue life of the component. Cutting forces are important indicator for power consumption required for cutting task. An artificial neural network (ANN) based surface roughness and power consumption model was established for Al 6061 under reaming operation. Back propagation neural networks were utilized for prediction of surface roughness and power consumption. Reaming test data was used to train and test the ANN network. In this presented study comparative investigation has been performed between the actual experimental values and neural network outputs to achieve good agreement.

Congestion is a major factor which affects the data center network traffic. Major applications of data center producing huge data traffic are social networks, online shopping portals, online banking portals, mobile application data, weather forecasting applications, Internet of Things applications and so on. Handling such large amount of data is a difficult task for data center architectures and mechanisms and always lead to congestion. The congestion control mechanisms already proposed in existing research works for Software Defined Data Centers lack behind in terms of scalability and throughput due to increase in data center traffic and centralized controller environment. So, in this paper, we are proposing a Simple Congestion Avoidance mechanism for Multipath Transmission Control Protocol traffic using Random Early Detection Algorithm to overcome the limitation of huge packet loss for a specific flow and detection of congested links in the early stage. In our proposed work, path generation among the particular source and destination has been carried out as Multipath Transmission Control Protocol traffic which can prevent congestion state due to the generation of sub flows between two hosts. Then, we bring up Random Early Detection mechanism as a congestion avoidance strategy to minimize overloaded links, link delay and to enhance throughput during packet transmission. This implementation is carried out in the network simulator, Mininet with OpenDaylight controller setup which runs on the top of Jellyfish data center topology. We finally compare the performance of proposed work with an existing congestion avoidance mechanism and show that our implementation offers better throughput by considering the link load, capacity for different node ranges.

We proposed a workload assignment policy for reducing power consumption by air conditioners in data centers. Power consumption was estimated by computational fluid dynamics for both the conventional equipment arrangement and the tandem arrangement newly developed to reuse the exhaust heat from servers. To reduce the air conditioner power consumption by raising the temperature set points of the air conditioners, the temperatures of all server back-planes were equalized by moving workload from the servers with the highest temperatures to the servers with the lowest temperatures. Consequently, the air conditioners' power consumption was reduced by 10.4% in the conventional arrangement. In the tandem arrangement, the air conditioners' power consumption was reduced by 53%, and also the total power consumption of the whole data center was exhibited to be reduced by 23% by reusing the exhaust heat from the servers.

SummaryPower consumption of current High Performance Computing systems has to be reduced by at least one order of magnitude before they can be scaled up towards ExaFLOP performance. While we can expect novel hardware technologies and architectures to contribute towards this goal, significant advances have to come also from software technologies such as proactive and power‐aware scheduling, resource allocation, and fault‐tolerant computing. Development of these software technologies in turn relies heavily on our ability to model and accurately predict power consumption in large computing systems. In this paper, we present a data‐driven model of power consumption for a hybrid supercomputer (which held the top spot in the Green500 ranking in June 2013) that combines CPU, GPU, and MIC technologies to achieve high levels of energy efficiency. Our model takes as input workload characteristics—the number and location of resources that are used by each job at a certain time—and calculates a predicted power consumption at the system level. The model is application‐code‐agnostic and is based solely on a data‐driven predictive approach, where log data describing the past jobs in the system are employed to estimate future power consumption. For this, three different model components are developed and integrated. The first employs support vector regression to predict power usage for jobs before these are started. The second uses a simple heuristic to predict the length of jobs, again before they start. The two predictions are then combined to estimate power consumption due to the job at all computational elements in the system. The third component is a linear model that takes as input the power consumption at the computing units and predicts system‐wide power consumption. Our method achieves highly‐accurate predictions starting solely from workload information and user histories. The model can be applied to power‐aware scheduling and power capping: alternative workload dispatching configurations can be evaluated from a power perspective and more efficient ones can be selected. The methodology outlined here can be easily adapted to other HPC systems where the same types of log data are available.

Due to the rapid advancement of mobile communication technology, mobile devices nowadays can support a variety of data services that are not traditionally available. With the growing popularity of mobile devices in the last few years, attacks targeting them are also surging. Existing mobile malware detection techniques, which are often borrowed from solutions to Internet malware detection, do not perform as effectively due to the limited computing resources on mobile devices. In this paper, we propose VirusMeter, a novel and general malware detection method, to detect anomalous behaviors on mobile devices. The rationale underlying VirusMeter is the fact that mobile devices are usually battery powered and any malicious activity would inevitably consume some battery power. By monitoring power consumption on a mobile device, VirusMeter catches misbehaviors that lead to abnormal power consumption. For this purpose, VirusMeter relies on a concise user-centric power model that characterizes power consumption of common user behaviors. In a real-time mode, VirusMeter can perform fast malware detection with trivial runtime overhead. When the battery is charging (referred to as a battery-charging mode), VirusMeter applies more sophisticated machine learning techniques to further improve the detection accuracy. To demonstrate its feasibility and effectiveness, we have implemented a VirusMeter prototype on Nokia 5500 Sport and used it to evaluate some real cellphone malware, including FlexiSPY and Cabir. Our experimental results show that VirusMeter can effectively detect these malware activities with less than 1.5% additional power consumption in real time.

Efficient energy optimization and scheduling in industrial factories depend on accurate, reasonable, and real-time monitoring of equipment power consumption. However, the power prediction of industrial equipment requires a large number of process data, which could be inevitably contaminated by some imperfect data, due to the harsh environments. Monitoring or predicting equipment power consumption is usually not feasible using data-driven black-box models with imperfect data. This paper proposes a prediction-based approach for power consumption monitoring using an interpretable data-driven model. First, a data preprocessing method is used to remove outliers and fill in missing values. Then, a Volterra polynomial basis function (VPBF) model is built to predict equipment power consumption. This model decomposes power values into a series of basis functions consisting of input parameters. Moreover, to compensate for data dropouts during the power consumption monitoring process, a networked predictive monitoring system is also proposed. Finally, this paper presents two case studies based on actual production equipment in an industrial manufacturing factory. The results demonstrate that the proposed approach can achieve satisfactory monitoring accuracy and adaptation ability. <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Note to Practitioners</i> —This paper is motivated by the problem of power consumption monitoring of industrial equipment in harsh production environments. A conventional solution is to predict the power consumption using data-driven black-box models, with limited feasibility and interpretability. This paper proposes a new power consumption monitoring approach, utilizing an interpretable data-driven model and a networked predictive method. This approach accurately reveals the transparent relationships between the output and input parameters. The power prediction result is consequently interpretable and more reasonable. Meanwhile, this approach actively compensates for data dropouts in the network, which can help operators real-time grasp the power consumption of equipment. Furthermore, this approach can be integrated into energy optimization systems as a basis of optimization and decision-making. Practical applications in an aluminum manufacturing company located in Guangdong Province demonstrate that this approach is feasible and applicable. Future research is to expand this approach further for energy optimization and scheduling.

Cloud Data Centers are becoming a more and more important reality nowadays, as more and more services are being offered in the form of Cloud resources. In order to accommodate as many requests as possible, Data Centers typically deploy oversubscription, which requires a thoughtfully calibrated usage of the computing and network resources of the system. We especially need to re-think the usage of network resources, as the kinds of demands from applications we see today have more stringent network requirements than in the past.The Software-Defined Network paradigm offers new tools to manage the network in agile, flexible ways, according to specific needs. It enables to build new customized functions in order, for example, to monitor the instantaneous load of traffic, slice the network, route each flow according to specific policies that can be modified accordingly with the needs of the system.In this paper we present an orchestrator for Software-Defined Data Center, and the strategies used to allocate Virtual Machines, basing our decisions on the current state of the system.

Power consumption of tracked and wheeled small mobile robots on deformable terrains–model and experimental validation

To improve the accuracy and reliability of load forecasting, this paper proposes a load forecasting method based on clustering and LSTM(long short term memory network). Firstly, the K-means clustering method is used to learn the regional power consumption model. Then the support vector machine algorithm is used to find the relationship between the climate conditions and the power consumption model. According to the climate conditions, the trained model is used to predict the power consumption model. Finally, the LSTM neural network is used to predict the total power consumption of the day to be forecasted, and the predicted load curve of the power consumption pattern is combined with the predicted total power consumption to obtain the predicted load at each time point of the day to be predicted. Matlab platform was used to verify the prediction. Six days were randomly selected for prediction. The MAPE (mean absolute percentage error) of this model kept between 1.6% and 2.5%, which is higher in prediction accuracy than traditional prediction methods.

Household power consumption pattern modeling through a single power sensor