Search

Today, there is an ever-increasing number of workloads pushed and executed on the Cloud. Data center operators and Cloud providers have embraced application co-location and multi-tenancy as first-class system design concerns to effectively serve and manage these huge computational demands. In addition, the continuous advancements in the computers' hardware technology have made it possible to seamlessly leverage heterogeneous pools of physical machines in data center environments. Even though current modern Cloud schedulers and orchestrators adopt application-aware policies to achieve automation of time-consuming management tasks at scale, e.g., resource provisioning, they still rely on coarse-grained system metrics, such as CPU and/or memory utilization to place incoming applications, thus, not considering (1) interference effects that are provoked by co-located tasks, and (2) the impact on performance caused by the diversity of heterogeneous systems' characteristics. The lack of such knowledge in existing state-of-the-art orchestration solutions results in their inability to perform efficient allocations, which negatively impacts the overall latency distribution delivered by the infrastructure. In this paper, to alleviate this inefficiency, we present a machine learning (ML) based Cloud orchestration extension that takes into account both resource interference and heterogeneity. The framework adequately schedules data-analytics applications on a pool of heterogeneous resources. We evaluate our proposed solution on different application mixes and co-location scenarios. We show that the proposed framework improves the tail latency of the distribution of the deployed applications by up to 3.6x compared to the state-of-the-art Kubernetes scheduler. [ABSTRACT FROM AUTHOR]

With technological advancements, much data is being captured by sensors, smartphones, wearable devices, and so forth. These vast datasets are stored in data centres and utilized to forge data‐driven models for the condition monitoring of infrastructures and systems through future data mining tasks. However, these datasets often surpass the processing capabilities of traditional information systems and methodologies due to their significant size. Additionally, not all samples within these datasets contribute valuable information during the model training phase, leading to inefficiencies. The processing and training of Machine Learning algorithms become time‐consuming, and storing all the data demands excessive space, contributing to the Big Data challenge. In this paper, we propose two novel techniques to reduce large time‐series datasets into more compact versions without undermining the predictive performance of the resulting models. These methods also aim to decrease the time required for training the models and the storage space needed for the condensed datasets. We evaluated our techniques on five public datasets, employing three Machine Learning algorithms: Holt‐Winters, SARIMA, and LSTM. The outcomes indicate that for most of the datasets examined, our techniques maintain, and in several instances enhance, the forecasting accuracy of the models. Moreover, we significantly reduced the time required to train the Machine Learning algorithms employed. [ABSTRACT FROM AUTHOR]

In the past twenty years, the IT industry has moved away from using physical servers for workload management to workloads consolidated via virtualization and, in the next iteration, further consolidated into containers. Later, container workloads based on Docker and Podman were orchestrated via Kubernetes or OpenShift. On the other hand, high-performance computing (HPC) environments have been lagging in this process, as much work is still needed to figure out how to apply containerization platforms for HPC. Containers have many advantages, as they tend to have less overhead while providing flexibility, modularity, and maintenance benefits. This makes them well-suited for tasks requiring a lot of computing power that are latency- or bandwidth-sensitive. But they are complex to manage, and many daily operations are based on command-line procedures that take years to master. This paper proposes a different architecture based on seamless hardware integration and a user-friendly UI (User Interface). It also offers dynamic workload placement based on real-time performance analysis and prediction and Machine Learning-based scheduling. This solves a prevalent issue in Kubernetes: the suboptimal placement of workloads without needing individual workload schedulers, as they are challenging to write and require much time to debug and test properly. It also enables us to focus on one of the key HPC issues—energy efficiency. Furthermore, the application we developed that implements this architecture helps with the Kubernetes installation process, which is fully automated, no matter which hardware platform we use—x86, ARM, and soon, RISC-V. The results we achieved using this architecture and application are very promising in two areas—the speed of workload scheduling and workload placement on a correct node. This also enables us to focus on one of the key HPC issues—energy efficiency. [ABSTRACT FROM AUTHOR]

Dissolved Organic Carbon (DOC) is the largest organic carbon pool in the ocean. Considering the biotic and abiotic factors controlling DOC processes, indirect satellite methods for open ocean DOC estimation can be developed, using conceptual, empirical or statistical models, driven by multiple satellite products. In this study, we infer a time series of global DOC from data of the European Space Agency's (ESA) Ocean Colour Climate Change Initiative (OC-CCI) in combination with a global database of in situ DOC observations. We tested empirical machine learning modelling approaches in which the available in situ data are used to train the models and to find empirical relationships between DOC and variables available from remote sensing. Of the tested methods, a random forest regression showed the best results, and the details of this model are further reported here. We present a time series of global open ocean DOC concentrations between 2010-2018 that is made freely available through the archive of the UK Centre for Environmental Data Analysis (CEDA). [ABSTRACT FROM AUTHOR]

Summary: The rapid expansion of IoT systems has caused network congestion and delays in task placement and resource provisioning as usually the tasks are executed at a far location in the cloud. Fog computing reduces the computing burden of cloud data centers as well as the communication burden of the internet as fog resources are placed near the data generation points. Within Fog computing, an important challenge is the optimal task placement which is an NP‐class problem. This work applies machine learning for task clustering and addresses the task placement problem in a fog computing environment using a hybrid of two recent metaheuristics; Jaya and gray wolf optimization (GWO). The hybrid method considers optimizing the total number of active fog nodes, load balancing in fog nodes, and average response time of the tasks. The performance of the proposed method is evaluated on a real‐time LCG dataset and is compared with reinforcement learning fog scheduling (RLFS), genetic algorithm (GA), dynamic resource allocation mechanism (DRAM), load balancing and scheduling algorithm (LBSSA), and particle swarm optimization with simulated annealing (PSO‐SA) algorithms. The results demonstrate the superiority of the suggested method over the baseline techniques in terms of average improvement of 51.04% in load balance variance, 30.25% in average response time, 24.16% in execution time, and 47.10% in the number of devices used. [ABSTRACT FROM AUTHOR]

Data centers are a major source of carbon emissions and are subsequently contributing to global carbon footprints. Keeping in view of providing a sustainable solution to society, we have analyzed various factors that can help to achieve carbon neutrality and maximum sustainability. Our study pointed towards the need to follow a sustainable approach for incoming workflow throughout the life-cycle of Data centers. We analyzed that workloads need to be segregated before assigning them to the data centers so that energy-efficient resource allocation could be done. This paper demonstrates unsupervised learning techniques to cluster the incoming cloud workloads. The heterogeneous workloads were characterized using machine learning approaches and appropriate clusters were crafted. For analysis, Google Cluster Dataset is used. In order to improve the accuracy, data were normalized, and random samples of data were selected for clustering. The machine learning algorithms applied were able to successfully determine the appropriate clusters that can further be used for energy-efficient resource scheduling. [ABSTRACT FROM AUTHOR]

Water is essential for humans as well as for all living organisms to sustain their lives. Therefore, any climate‐driven change in available resources has significant impacts on the environment and life. Global climate models (GCMs) are one of the most practical methods to evaluate climate change. Based on this, this research evaluated the capability of GCMs from the Coupled Model Intercomparison Project 6 (CMIP6) to reproduce the historical flow of climate prediction centre data for the Konya Closed basin and to project the climate of the basin using the selected GCMs. Global climate models based on the CMIP6 under the scenario of common socioeconomic pathways (SSP245 and SSP 585) were used to analyse the climate change effect on streamflow of the study area by Bias Correction of GCM Models using Long Short‐Term Memory (LSTM), Bidirectional LSTM (BiLSTM), AdaBoost, Gradient Boosting, Regression Tree, and Random Forest methods. The coefficient of determination (R2), mean square error (MSE), mean absolute error (MAE), root mean square error (RMSE) were used to assess the performance of the methods. Findings show that the Random Forest Model consistently outperformed other models in both the testing and training phases. A significant downward in the volume of water flowing through the region's rivers and streams in the next decades. It is critical to enhance climate‐resilient water infrastructure financing, establish an early warning system for drought, introduce best management practices, implement integrated water resource management, public awareness, and support water research to alleviate the negative consequences of drought and increase resilience against the effects of climate change on Turkey's water resources. [ABSTRACT FROM AUTHOR]

Continuous discovery and update of applications or their boundaries running in cloud environments in an automatic way is a highly required function of modern data center operation solutions. Prior attempts to address this problem within various products or projects were/are applying rule-driven approaches or machine learning techniques on specific types of data–network traffic as well as property/configuration data of infrastructure objects, which all have their drawbacks in effectively identifying roles of those resources. The current proposal (ADLog) leverages log data of sources, which contain incomparably richer contextual information, and demonstrates a reliable way of discriminating application objects. Specifically, using native constructs of VMware Aria Operations for Logs in terms of event types and their distributions, we group those entities, which then can be potentially enriched with indicative tags automatically and recommended for further management tasks and policies. Our methods differentiate not only diverse kinds of applications, but also their specific deployments, thus providing hierarchical representation of the applications in time and topology. For several applications under Aria Ops management in our experimental test bed, we discover those in terms of similarity behavior of their components with a high accuracy. The validation of the proposal paves the path for an AI-driven solution in cloud management scenarios. [ABSTRACT FROM AUTHOR]

In response to the global megatrends of digitalization and transportation automation, Turku University of Applied Sciences has developed a test platform to advance autonomous maritime operations. This platform includes the unmanned surface vessel eM/S Salama and a remote operations center, both of which are detailed in this article. The article highlights the importance of collecting and annotating multi-modal sensor data from the vessel. These data are vital for developing deep learning algorithms that enhance situational awareness and guide autonomous navigation. By securing relevant data from maritime environments, we aim to enhance the autonomous features of unmanned surface vessels using deep learning techniques. The annotated sensor data will be made available for further research through open access. An image dataset, which includes synthetically generated weather conditions, is published alongside this article. While existing maritime datasets predominantly rely on RGB cameras, our work underscores the need for multi-modal data to advance autonomous capabilities in maritime applications. [ABSTRACT FROM AUTHOR]

The proliferation of data centers is driving increased energy consumption, leading to environmentally unacceptable carbon emissions. As the use of Internet-of-Things (IoT) techniques for extensive data collection in data centers continues to grow, deep learning-based solutions have emerged as attractive alternatives to suboptimal traditional methods. However, existing approaches suffer from unsatisfactory performance, unrealistic assumptions, and an inability to address practical data center optimization. In this paper, we focus on power usage effectiveness (PUE) optimization in IoT-enabled data centers using deep learning algorithms. We first develop a deep learning-based PUE optimization framework tailored to IoT-enabled data centers. We then formulate the general PUE optimization problem, simplifying and specifying it for the minimization of long-term energy consumption in chiller cooling systems. Additionally, we introduce a transformer-based prediction network designed for energy consumption forecasting. Subsequently, we transform this formulation into a Markov decision process (MDP) and present the branching double dueling deep Q-network. This approach effectively tackles the challenges posed by enormous action spaces within MDP by branching actions into sub-actions. Extensive experiments conducted on real-world datasets demonstrate the exceptional performance of our algorithms, excelling in prediction precision, optimization convergence, and optimality while effectively managing a substantial number of actions on the order of 10 13 . [ABSTRACT FROM AUTHOR]

Long-term PM2.5 data are needed to study the atmospheric environment, human health, and climate change. PM2.5 measurements are sparsely distributed and of short duration. In this study, daily PM2.5 concentrations are estimated from 1959 to 2022 using a machine learning method at 4011 terrestrial sites in the Northern Hemisphere based on hourly atmospheric visibility data, which are extracted from the Meteorological Terminal Aviation Routine Weather Report (METAR). PM2.5 monitoring is the target of machine learning, and atmospheric visibility and other related variables are the inputs. The training results show that the slope between the estimated PM2.5 concentration and the monitored PM2.5 concentration is 0.946± 0.0002 within the 95 % confidence interval (CI), the coefficient of determination (R2) is 0.95, the root mean square error (RMSE) is 7.0 μg/m3, and the mean absolute error (MAE) is 3.1 μg/m3. The test results show that the slope between the predicted PM2.5 concentration and the monitored PM2.5 concentration is 0.862 ± 0.0010 within a 95 % CI, the R2 is 0.80, the RMSE is 13.5 μg/m3, and the MAE is 6.9 μg/m3. The multiyear mean PM2.5 concentrations from 1959 to 2022 in the United States, Canada, Europe, China, and India are 11.2 μg/m3, 8.2 μg/m3, 20.1 μg/m3, 51.3 μg/m3 and 88.6 μg/m3, respectively. PM2.5 is low and continues to decrease from 1959 to 2022. PM2.5 in the United States increases slightly at a rate of 0.38 μg/m3/decade from 1959 to 1990 and decreases at a rate of -1.32 μg/m3/decade from 1991 to 2022. Trends in Europe are positive (5.69 μg/m3/decade) from 1959 to 1972 and negative (-1.91 μg/m3/decade) from 1973 to 2022. Trends in China and India are increasing (3.04 and 3.35 μg/m3/decade, respectively) from 1959 to 2012 and decreasing (-38.82 and -42.84 μg/m3/decade, respectively) from 2013 to 2022. The dataset is available at National Tibetan Plateau / Third Pole Environment Data Center (https://doi.org/10.11888/Atmos.tpdc.301127) (Hao et al., 2024). [ABSTRACT FROM AUTHOR]

UAVs or Drones can be used to support wireless communication by acting as flying or mobile Base Stations for the accumulation of the different types of data to train the models. However, in traditional or DL-based UAVs, the raw data is sent from the devices to the centralized server, which causes problems in terms of the privacy of the devices and the UAVs' communication resources or limited processing. Therefore, the issue with DL-based UAVs is that sending the original data to the centralized body raises questions about security and privacy. The transmission of distributed, unprocessed data from the drones to the cloud, including interactive media information data types, requires a significant amount of network bandwidth and more energy, which has an enormous effect on several trade-offs, including communication rates and computation latencies. Data packet loss caused by asynchronous transmission, which doesn't prevent peer-to-peer communication, is a concern with AFL-based UAVs. Therefore, in order to address the aforementioned issues, we have introduced SFL-based UAVs that focus on creating algorithms in which the models simultaneously update the server as they wait for all of the chosen devices to communicate. The proposed framework enables a variety of devices, including mobile and UAV devices, to train or learn their algorithms for machine learning before updating the models and parameters simultaneously to servers or manned aerial data centers for model buildup without transferring their original private information. This decreases packet loss and privacy threats while also enhancing round effectiveness as well as model accuracy. The comparative analysis of AFL and SFL techniques in terms of accuracy, global rounds, and communication rounds are offered. Simulation findings suggest that the proposed methodology improves in terms of global rounds and accuracy. [ABSTRACT FROM AUTHOR]

Solar radio bursts are intense radio radiation sources that occur during the energy-release process and represent a hot topic in solar-physics and space-weather research. In this paper, we present a multimode prediction model for daily solar radio bursts. The model uses deep learning and machine learning to obtain data information from different dimensions and to establish the relationship between the characteristics of the solar active region on the solar surface and solar radio bursts. For this model, we use data from the Solar and Heliospheric Observatory (SOHO)/Michelson Doppler Imager (MDI) total solar magnetic map, the Royal Observatory of Belgium World Data Centre in Brussels, and NOAA sunspot parameters (including number, area, and type of sunspots) as inputs. The output results are then compared with the list of solar radio bursts recorded by the Radio Solar Telescope Network (RSTN) to determine whether solar radio bursts are present and to determine the key parameters for determining radio bursts. Based on 5449 days of observational data, we find that the prediction accuracy of the model is 0.898 ± 0.011, and that the number of sunspots is a key parameter in determining the occurrence of solar radio bursts. Specifically, when the number of sunspots is greater than 15, the probability of occurrence of solar radio bursts is greater than 90%. We have identified the key parameters and thresholds for determining solar radio bursts and highlighted the key parameters for space-weather prediction. In addition, the prediction model can also be used for predicting in other fields. [ABSTRACT FROM AUTHOR]

Cloud computing has revolutionized the way software, platforms, and infrastructure can be acquired by making them available as on-demand services that can be accessed from anywhere via a web browser. Due to its ubiquitous nature Cloud data centers continuously experience fluctuating workloads which demands for dynamic resource provisioning. These workloads are either placed on Virtual Machines (VMs) or containers which abstract the underlying physical resources deployed at the data center. A proactive or reactive method can be used to allot required resources to the workload. Reactive approaches tend to be inefficient as it takes a significant amount of time to configure the resources to meet the change in demands. A proactive approach for resource management is better in meeting workload demands as it makes an appropriate number of resources available in advance to cater to the fluctuations in workload. The success of such an approach relies on the ability of the resource management module of a data center to accurately predict future workloads. Machine Learning (ML) has already proven itself to be very effective in performing prediction in various domains. In this work, we propose an ML meta-classifier based on stacked generalization for predicting future workloads utilising the past workload trends which are recorded as event logs at Cloud data centers. The proposed model showed a prediction accuracy of 98.5% indicating its applicability for the Cloud environment where SLA requirements must be closely adhered to. [ABSTRACT FROM AUTHOR]

In data center networks, when facing challenges such as traffic volatility, low resource utilization, and the difficulty of a single traffic scheduling strategy to meet demands, it is necessary to introduce intelligent traffic scheduling mechanisms to improve network resource utilization, optimize network performance, and adapt to the traffic scheduling requirements in a dynamic environment. This paper proposes a fine-grained traffic scheduling scheme based on multi-agent deep reinforcement learning (MAFS). This approach utilizes In-Band Network Telemetry to collect real-time network states on the programmable data plane, establishes the mapping relationship between real-time network state information and the forwarding efficiency on the control plane, and designs a multi-agent deep reinforcement learning algorithm to calculate the optimal routing strategy under the current network state. The experimental results demonstrate that compared to other traffic scheduling methods, MAFS can effectively enhance network throughput. It achieves a 1.2× better average throughput and achieves a 1.4–1.7× lower packet loss rate. [ABSTRACT FROM AUTHOR]

Summary: In this paper, we study the task offloading optimization problem in satellite edge computing environments to reduce the whole communication latency and energy consumption so as to enhance the offloading success rate. A three‐tier machine learning framework consisting of collaborative edge devices, edge data centers, and cloud data centers has been proposed to ensure an efficient task execution. To accomplish this goal, we also propose a Q‐learning‐based reinforcement learning offloading strategy in which both the time‐sensitive constraints and data requirements of the computation‐intensive tasks are taken into account. It enables various types of tasks to select the most suitable satellite nodes for the computing deployment. Simulation results show that our algorithm outperforms other baseline algorithms in terms of latency, energy consumption, and successful execution efficiency. [ABSTRACT FROM AUTHOR]

Industrial Internet of Things (IIoT) technology, as a subset of the Internet of Things (IoT) in the concept of Industry 4.0 and, in the future, 5.0, will face the challenge of streamlining the way huge amounts of data are processed by the modules that collect the data and those that analyse the data. Given the key features of these analytics, such as reducing the cost of building massive data centres and finding the most efficient way to process data flowing from hundreds of nodes simultaneously, intermediary devices are increasingly being used in this process. Fog and edge devices are hardware devices designed to pre-analyse terabytes of data in a stream and decide in realtime which data to send for final analysis, without having to send the data to a central processing unit in huge local data centres or to an expensive cloud. As the number of nodes sending data for analysis via collection and processing devices increases, so does the risk of data streams being intercepted. There is also an increased risk of attacks on this sensitive infrastructure. Maintaining the integrity of this infrastructure is important, and the ability to analyse all data is a resource that must be protected. The aim of this paper is to address the problem of autonomous threat detection and response at the interface of sensors, edge devices, cloud devices with historical data, and finally during the data collection process in data centres. Ultimately, we would like to present a machine learning algorithm with reinforcements adapted to detect threats and immediately isolate infected nests. [ABSTRACT FROM AUTHOR]

Cloud computing technology provides access on demand to virtualized resources, services, and applications via a distributed network. In cloud data centers effective energy utilization is a critical concern in today's technology-driven world. Cloud data centers (CDC) are massive facilities that host and manage an enormous amount of data and computing resources. This article addresses the growing significance and energy-intensive nature of data centers. Due to the rapid growth of cloud computing, it offers on-demand access to resources globally and leads to substantial power consumption and carbon impact on the environment. They consume substantial amounts of energy, and optimizing their energy utilization is essential for reducing operational costs, minimizing environmental impact, and ensuring sustainable growth. To combat this, efficient energy-saving approaches using machine learning methods have been researched. ML methods hold great potential for enhancing energy efficiency in CDCs by analysing data, detecting patterns, and optimizing resource usage. The focus areas include CPU usage prediction, overload finding, underload estimation, selection, migration, and relocation of VMs to attain resources and improve energy utilization. The paper compares energy-saving results achieved by different machine learning techniques in data centers to minimize energy usage and meet service level agreements (SLA). The machine learning approaches reduce energy consumption from 1.6% to 88.5% compared to the benchmark approach mentioned, considering various settings and parameters. [ABSTRACT FROM AUTHOR]

Task scheduling based on temperature perception is beneficial for avoiding hotspots and optimizing the internal temperature distribution of data centers. However, the accuracy of task scheduling largely depends on the accuracy of temperature prediction. There are many features that affect the accuracy of temperature prediction in data centers, and the variation periods of these features vary greatly. Traditional machine learning models are difficult to accurately fit them. Therefore, this article proposes a step-by-step temperature prediction algorithm based on Gated Recurrent Unit (GRU). This algorithm establishes prediction models for important parameters such as CPU utilization and air conditioning temperature that affect temperature prediction, and uses the outputs of these two models as inputs for the server temperature prediction model to better fit the changes of feature values. The model combines the principle of thermal locality and integrates the temperature of upper and lower servers for joint modeling. Experiments show that our prediction model can accurately predict the inlet temperature evolution of the server with dynamic workload. RSME reaches 0.278 and the average prediction temperature difference is 0.633, which is much higher than the traditional model. In addition, this article also propose a minimum temperature difference scheduling algorithm based on temperature prediction model, which can effectively reduce the number of servers running at high temperature and low temperature in the data center, make the temperature of the data center more balanced and achieve better energy-saving compared with other baseline algorithms. [ABSTRACT FROM AUTHOR]

Purpose: In rotating machines, bearings are essential components, but these have a limited life and are subject to degradation. This paper discusses detecting anomalies in the bearing's various components, such as the inner ring, outer ring, and ball, by processing the bearing's vibration signal. Methodology: Self-Organizing Maps (SOMs) and Variational Mode Decomposition (VMD) are used to automate diagnosing bearing defects. The approach's performance was evaluated using the Bearing Data Center (BDC) database, which includes data collection, preprocessing, and SOM algorithm implementation. Results and Conclusions: The proposed method accurately classifies bearing fault types and can aid in early detection and proactive maintenance to prevent catastrophic equipment failure. The findings also show that VMD decomposition and envelope demodulation effectively detect various bearing faults. Therefore, using advanced machine learning techniques to accurately diagnose rotating machine failures is necessary. [ABSTRACT FROM AUTHOR]

Modern data centers suffer from immense power consumption. As a result, data center operators have heavily invested in capacity-scaling solutions, which dynamically deactivate servers if the demand is low and activate them again when the workload increases. We analyze a continuous-time model for capacity scaling, where the goal is to minimize the weighted sum of flow time, switching cost, and power consumption in an online fashion. We propose a novel algorithm, called adaptive balanced capacity scaling (ABCS), that has access to black-box machine learning predictions. ABCS aims to adapt to the predictions and is also robust against unpredictable surges in the workload. In particular, we prove that ABCS is (1+ε) competitive if the predictions are accurate, and yet, it has a uniformly bounded competitive ratio even if the predictions are completely inaccurate. Finally, we investigate the performance of this algorithm on a real-world data set and carry out extensive numerical experiments, which positively support the theoretical results. Funding: This work was partially supported by the Division of Computing and Communication Foundations [Grant 2113027]. The authors also acknowledge financial support for this project from the Algorithm and Randomness Center–Transdisciplinary Research Institute for Advancing Data Science Fellowship at Georgia Tech. [ABSTRACT FROM AUTHOR]

The 6th generation mobile networks (6G) network is a kind of multi-network interconnection and multi-scenario coexistence network, where multiple network domains break the original fixed boundaries to form connections and convergence. In this paper, with the optimization objective of maximizing network utility while ensuring flows performance-centric weighted fairness, this paper designs a reinforcement learning-based cloud-edge autonomous multi-domain data center network architecture that achieves single-domain autonomy and multidomain collaboration. Due to the conflict between the utility of different flows, the bandwidth fairness allocation problem for various types of flows is formulated by considering different defined reward functions. Regarding the tradeoff between fairness and utility, this paper deals with the corresponding reward functions for the cases where the flows undergo abrupt changes and smooth changes in the flows. In addition, to accommodate the Quality of Service (QoS) requirements for multiple types of flows, this paper proposes a multi-domain autonomous routing algorithmcalled LSTM+MADDPG. Introducing a Long Short-TermMemory (LSTM) layer in the actor and critic networks, more information about temporal continuity is added, further enhancing the adaptive ability changes in the dynamic network environment. The LSTM+MADDPG algorithm is compared with the latest reinforcement learning algorithm by conducting experiments on real network topology and traffic traces, and the experimental results show that LSTM+MADDPG improves the delay convergence speed by 14.6% and delays the start moment of packet loss by 18.2% compared with other algorithms. [ABSTRACT FROM AUTHOR]

Load balancing is an element that must exist for a cloud server to function properly. Without it, there would be substantial lag and the server won't be able to operate as intended. In a Cloud computing (CC) establishing, load balancing is the process of dividing workloads and computer resources. The distribution of assets from a data centre involves many different factors, including load balancing of workloads in cloud environment. To make best use each virtual machine's (VM) capabilities, load balancing needs to be done in a way that ensures that all VMs have balanced loads. Both overloading and underloading are examples of load unbalance, and both of these types of load unbalance could be fixed by using techniques created especially for load balancing. The research that has been done on the subject have not attempted to take into account the factors that affect the problem of load unbalancing. Results indicate that the LSTM and DForest-based load balancing approach significantly improves cloud resource utilization, reduces response times, and minimizes the occurrence of overloading or underloading scenarios. To effectively design those programmes, it is essential to first understand the advantages and disadvantages of current methodologies before developing efficient AI-based load balancing programmes. Compared to existing method the proposed method is high accuracy 0.98, KNN accuracy is 0.97, SVM accuracy is 0.99, and DForest accuracy is 0.987. [ABSTRACT FROM AUTHOR]

Aiming at the problems of low convective heat transfer coefficient and high energy consumption in the air-cooled data center of immersed liquid cooling, an improved deep learning algorithm is proposed for the data center system of immersed liquid cooling equipment room. By improving the design of the immersed liquid cooling system, heat exchange is carried out between the immersed liquid cooling system and heating components such as the central processing unit of the server. The insulation coolant and cooling water achieve server heat dissipation through energy exchange, achieving data management of the immersed liquid cooling room. The proposed algorithm improves data management efficiency while ensuring computational accuracy by conducting in-depth training and learning on the obtained immersed liquid cooling data, thus achieving the management of data in the immersed liquid cooling room. Through experiments, it has been proven that the immersed liquid cooling system in this study has high data management efficiency and low error, and can maintain server memory heat below 37 ° C, with a research accuracy of up to 92%. [ABSTRACT FROM AUTHOR]

Based on the fifth‐generation reanalysis (ERA5) data of the European Centre for Medium Range Weather Forecasts (ECMWF), spring Mongolian cyclones were examined from 1950 to 2023. Then, based on the feature parameters of activity path information, machine learning method (k‐means) is used to classify the trajectories of spring Mongolian cyclone. Subsequently, the atmospheric circulation configurations of the three categories of Mongolian cyclones and their influence on the weather in China verify the reasonability of the classification method. Specifically: (1) eastward moving type, affecting Inner Mongolia and Northeast China (Clus‐1). The circulation background of the Clus‐1 cyclone displays a broad negative geopotential height anomaly centre over East Asia; (2) northeastward moving type, affecting northeast part of Northeast China (Clus‐2). The circulation background of the Clus‐2 cyclones shows a "positive–negative–positive" distribution of circulation anomalies in the European Plain, Lake Baikal and the Sea of Japan; (3) southeastward moving type, this type can move to the south of China (Clus‐3). The circulation background of the Clus‐3 cyclones is opposite to that of the Clus‐2 cyclones. The three categories of Mongolian cyclones will cause dust weather and strong winds in different regions of China, mainly affecting Inner Mongolia and Northeast China. [ABSTRACT FROM AUTHOR]

The Global Association model is a crucial tool in seismic data analysis at the International Data Centre (IDC) of the Comprehensive Nuclear-Test-Ban Treaty Organization. However, it faces challenges due to its limitations in accurately associating seismic events on a global scale. Over the past years, attempts have been undertaken to tackle these issues by introducing the Network Processing Vertically Integrated Seismic Analysis (NET-VISA) algorithm, specifically designed to enhance seismic event association across the globe. NET-VISA uses a machine learning Bayesian approach to solve the automatic association problem. NET-VISA has been implemented in operation as an additional automatic event scanner tool since January 2018. In this study, we assess the effect of the NET-VISA automatic scanner on the IDC output REB and LEB bulletins. We used three distinct time periods to evaluate the NET-VISA performance. The results show a 4.6% increase in the number of LEB events after including the NET-VISA scanner in operation, with an average of 7 additional events per day, and an increase of 17.90% in the number of scanned events. A comparison between the different bulletins in distinct periods shows NET-VISA is beneficial to build more valid events, providing opportunities to improve nuclear-test-ban monitoring. However, NET-VISA exhibits slightly reduced performance for events occurring at depths exceeding 300 km. [ABSTRACT FROM AUTHOR]

While providing efficient and convenient cloud services, data center also brings great pressure to energy consumption and environment. The rise of server temperature not only increases the refrigeration cost, but also seriously affects the operation safety of the data center. Effective analysis and prediction of data center temperature is not only conducive to preventing server overheating and shutdown, but also crucial to data center task scheduling, resource allocation optimization and energy efficiency improvement of data center. Therefore, this article proposes a Two-stage Gated Recurrent Unit (GRU) temperature prediction algorithm with self-healing mechanism. The algorithm establishes a prediction model for the important parameters affecting temperature prediction - CPU utilization, and takes the output of the model as the input parameter of the server temperature prediction model, which fits the changes of each parameter more accurate. To avoid the decrease in prediction accuracy caused by new operating conditions that have not been learned before and changes in physical environmental factors during the operation of the model, a self-healing mechanism is proposed to ensure the prediction accuracy of the model. Experiments show that our prediction model can accurately predict the inlet temperature evolution of the server with dynamic workload. It reduces the prediction error (RSME) to 0.280, and the average prediction temperature difference is only 0.675, which is 10 % higher than the single stage prediction accuracy. The use of Two-stage prediction methods in other machine learning methods can also improve prediction accuracy. Based on the prediction model, this paper proposes a task scheduling algorithm that minimizes temperature difference. The algorithm can make the temperature between servers more balanced after task allocation, effectively reducing the number of servers running at high and low temperatures in the data center, avoiding refrigeration waste, and achieving energy conservation in the data center. [ABSTRACT FROM AUTHOR]

Failure prediction in a population of hard drives remains challenging due to the extreme class imbalance. Many existing failure prediction methods rely on problematic resampling approaches. Besides, existing deep learning methods in the failure prediction domain typically employ the sigmoid cross-entropy loss function which is not capable of solving the class imbalance problem. To address these challenges, this study proposes long-short term memory networks that leverage the sequential time series self monitoring analysis and reporting technology sensors data provided by the Backblaze data center. These models employ a new, modified focal loss function combined with weighted cross-entropy loss to tackle the negative influence of the class imbalance during training the long-short term memory networks. The long-short term memory networks models are compared to two traditional machine learning algorithms: random forest and logistic regression. This study focuses on daily snapshot data of the Seagate ST4000DM000 hard drive model from the Backblaze data center to maintain the homogeneity of the SMART attributes. Long-short term memory networks with modified focal loss function model provides the highest geometric mean score of 0. 786 ± 0. 011 while keeping the failure detection rate higher without compromising the false alarm rate. Hence, it demonstrates that the proposed focal loss function in long-short term memory networks leads to improved performance compared to other approaches under the class imbalance. [ABSTRACT FROM AUTHOR]