A SYSTEM AND METHODS FOR ACHIEVING ENERGY SAVINGS IN DATA CENTERS USING TFT.

Abstract

The invention relates to a system and method for providing prediction and resource management based on a Temporal Fusion Transformer (TFT).

Description

1 TARIFF SYSTEM THAT PROVIDES ENERGY SAVINGS IN DATA CENTERS WITH TFT. METHOD Technical Area The invention provides energy savings in the fields of cloud computing and data center management. It is related to the system and method. The invention specifically concerns Temporal Fusion Transformer (TFT) based estimation and resource management. It is related to a system and method that provides this. State of the Art Today, companies are looking for 10 ways to reduce energy costs and decrease their carbon footprint. They are turning to dynamic resource management systems (DRS). However, the fundamentals in these systems... The problem is that sudden increases in workload require restarting physical hosts and virtual ones. These are delays that occur during server migration. These delays, especially in real-world sectors such as finance, e-commerce, online services, or IoT (Internet of Things). Timely performance leads to unacceptable levels of slowdowns in critical sectors. 15 This opens up opportunities. For example, an e-commerce site might see a tenfold increase in traffic on Black Friday. In this situation, the company may lose a customer or incur a financial loss due to the hosts not being ready. The institution may experience processing errors with millisecond delays. Current practices and algorithms generally involve migrating virtual machines and hosting them. Technologies like DRS (Distributed Resource Scheduler) used for energy efficiency 20 It relies on. In VMware and similar virtualization platforms, the DRS feature is unused. The goal is to shut down the hosts and thus save energy. However, this The biggest problem encountered during system operation is sudden load increases. The problem is that it takes time to bring the hosts back online. This process involves the hosts... Because it involves accessing services and migrating virtual machines, there is a certain delay of 25 days. This happens and can lead to slowdowns or even interruptions in the environment. It is usually instantaneous. It makes decisions based on resource usage. That is, it detects changes in system load. They only consider the current situation, which makes them resistant to sudden load increases. This leaves them unprepared. Another significant drawback of such algorithms is energy. They fail to consider performance losses while achieving cost savings. However, the reality is 30 Globally, a balance needs to be struck between energy efficiency and high performance. 2 These shortcomings are particularly relevant in environments with high availability and high performance requirements. This leads to major problems in various sectors. For example, in the financial sector, data... Temporary performance losses at these centers lead to significant financial losses. It can open up. Similarly, in the healthcare sector, sudden increases in load or system slowdowns. This can prevent patients from quickly accessing their data, which could have serious consequences. These 5 In this context, the shortcomings of existing algorithms, especially in large-scale data centers, This constitutes a significant technical problem affecting the speed and efficiency of applications. As a result of the research conducted on this subject, it was determined that the "Data center" with the number CN114511208A. energy consumption optimization control method based on deep reinforcement An application titled "learning" was found. The system is a data 10 based on deep learning. This refers to a central energy optimization solution and proactive sourcing. It does not include a solution regarding its activation. In conclusion, due to the negative aspects described above and the current solutions being the subject of discussion... Due to its shortcomings, an improvement is needed in the relevant technical field. It has been made. 15 The purpose of the invention The invention was created by drawing inspiration from existing situations and overcoming the aforementioned drawbacks. It aims to solve the problem. The main purpose of the invention is to provide prediction and sourcing based on Temporal Fusion Transformer (TFT). The goal is to develop a system and method for managing. 20 Another aim of the invention is to strike an optimal balance between energy efficiency and performance. by establishing both large-scale data centers (Google Cloud, AWS, Azure, etc.) and The goal is to offer a system and method targeting organizations with private cloud infrastructures. It can also be applied in the field of edge computing. For example, smart IoT devices in factories or autonomous vehicles require real-time data processing, thus consuming 25% energy. It needs infrastructure that saves energy but also anticipates increased load. Similarly, hybrid cloud systems, telecommunication networks (5G core networks), and even government The demands of these institutions fall within the scope of this invention. Specifically, artificial intelligence-based prediction. thanks to its mechanism, seasonality, trends and external factors (marketing Increases in load (considering campaigns, social media interactions, etc.) are 30%. This is predictable. This also means hospitals will manage patient data flow, video streaming. 3 platforms providing uninterrupted service during peak hours or game companies This comes into play in scenarios such as dynamically scaling server capacity. In addition, in data centers with renewable energy integration, energy consumption is reduced. It can also be used to optimize for solar / wind power generation. Another objective of the invention is to use TFT deep learning architecture to analyze 5-bit time series data. multiple time scales (minute, hourly, seasonal), heterogeneous data sources (IoT) sensors, user logs, external factors) and contextual factors (marketing Combining campaigns (hardware status) to achieve high accuracy in load forecasting. The goal is to provide a system and method that enables TFT to improve self-attention. mechanisms and variable choice networks, both long-term dependencies of the model 10 (for example, annual increasing traffic trends) as well as short-term sudden fluctuations (for example, This allows for the simultaneous analysis of the immediate impact of social media interactions. In this way, future workload increases can be predicted in advance, reducing the load on the hosts. It is made ready before they arrive. Another aim of the invention is proactive resource activation and dynamic energy optimization. 15 The goal is to design a system and method that provides this. The invention uses TFT predictions. It activates the hosts as needed and for as long as needed. For example, a traffic increase of 5 If it is predicted to happen in minutes, the hosts are woken up after 4.5 minutes. Migration time (30 seconds - 1 minute) and service startup time (1-2 minutes) It is integrated into the slice. Thus, when the load comes, the hosts are ready and the latency is 20 This is not possible. Dynamic sleep modes are used for energy saving: the TFT's low load By accurately predicting their cycles, the number of unnecessary active hosts is minimized. At the algorithmic level, TFT's multi-head attention layers are different It improves forecast quality by highlighting key features across time scales. Static and Dynamic feature encoders, with fixed parameters (e.g., server location) 25 It analyzes parameters that change over time (e.g., CPU usage). It also analyzes the system. It updates itself with real-time feedback loops: Prediction errors or sudden changes Hardware failures are immediately integrated into the model, allowing for adaptation. As a result... This invention radically solves the delayed scaling problem while achieving high energy efficiency. It maximizes performance simultaneously. TFT's superior predictive ability and dynamic 30 Optimization mechanisms eliminate the shortcomings of existing systems by up to 30- It provides 40% energy savings and millisecond-level SLA compliance. 4 To achieve the objectives described above, the invention enables proactive measures in data centers. It is a system that enables resource activation and energy saving. Accordingly, the system;  physically and / or virtually characterized as located in a data center presenter,  Raw data collected from servers located in the data center can be analyzed. 5 To achieve this, noisy, incomplete, or inconsistent recordings need to be cleaned up and... Data preprocessing module that enables normalization,  The deep learning model to be implemented will use continuously updated data and training with real-time feedback and static and dynamic data The relationships between them are established via static feature encoders and dynamic feature encoders. 10 training module that enables differentiation,  centralized storage, management, and rapid access to all data database that enables access  Data received from the aforementioned servers and processed through pre-processing steps In light of this, proactive 15 by running pre-trained deep learning algorithms Instantaneous, enabling resource activation and dynamic energy optimization. By looking at performance data, deviations from predictions or any sudden changes Decision support takes action by identifying whether a change has occurred. system It includes. 20 The invention also enables proactive resource activation and energy savings in data centers. It also includes the method that provides this. Accordingly, the method is:  physically and / or virtually characterized as located in a data center Data collected via the server, data that can be processed by a data 25 Collection via preprocessing module,  The aforementioned preprocessing module cleans and processes the collected data. normalization processes are carried out,  The deep learning model to be implemented will use continuously updated data and Training via a training module with real-time feedback and a fixed 30 and the relationships between dynamic data using static feature encoders and dynamic features. parsing via encoder,  all data is centrally stored, managed, and... providing quick access,  Data received from the aforementioned servers and subjected to preprocessing steps In light of this, a decision support system capable of performing analysis using deep learning models. the system's proactive resource activation and dynamic energy optimization ensuring,  By looking at the real-time performance data of the aforementioned decision support system 5 whether there have been any deviations from the predictions or any sudden changes. by identifying and taking action It includes the steps involved in the process. The structural and characteristic features and all the advantages of the invention are given in the figures below and 10 This becomes clearer thanks to the detailed explanation written with references to these figures. This will be understood as such, and therefore the evaluation will also take these forms and detailed explanations into account. This should be done taking that into consideration. Ways to Help Understand the Discovery Figure 1 shows a schematic representation of the system that is the subject of the invention. 15 Figure 2 shows the flowchart of the method that is the subject of the invention. Explanation of Part References 1. Server 2. Data preprocessing module 3. Training module 20 4. Database 5. Decision support system 1001. Physical and / or virtual characteristics found in a data center data collected from the server, data preprocessor capable of processing the collected data. Collection via the module. 25 1002. Preprocessing module: Cleaning and normalization of collected data. to carry out the operations. 6 1003. The deep learning model to be implemented will use continuously updated data and training through a training module with real-time feedback and fixed and Relationships between dynamic data using static feature encoder and dynamic feature encoder separation by means of 1004. All data should be centrally stored, managed, and organized in a database. 5 providing quick access. 1005. Based on the data received from the servers and processed through pre-processing steps, deep a decision support system that can perform analysis using learning models, proactive It enables resource activation and dynamic energy optimization. 1006. Deviation in predictions based on real-time performance data of the decision support system is 10. or taking action by identifying whether there are any sudden changes. TH: Prediction errors PV: Performance data Detailed Description of Find This detailed explanation describes the preferred 15 systems and methods that are the subject of the invention. Their structures are explained solely to facilitate a better understanding of the subject. The invention enables proactive resource activation and energy savings in data centers. It is a system. Figure 1 shows a schematic representation of the system that is the subject of the invention. According to the system, a data center is characterized by its physical and / or virtual characteristics. 20 raw data collected from servers (1) located in the data center, which can be accessed from the server (1). noisy, incomplete or inconsistent recordings are analyzed in order to make them analyzable. data preprocessing module (2) that enables cleaning and normalization of data. The deep learning model to be implemented will be based on continuously updated data and real-world conditions. training with timely feedback and relationships between static and dynamic data 25 that enables separation via static feature encoder and dynamic feature encoder. training module (3) centrally stores, manages and quickly stores all data. database (4) which enables access to the aforementioned servers (1) Based on the data received and pre-processed, pre-trained deep-seated systems were developed. Proactive resource activation and dynamic energy through the operation of learning algorithms. Optimization is achieved by looking at real-time performance data (PV), with a deviation of 30 in predictions. 7 or taking action by detecting whether any sudden changes have occurred It includes a decision support system (5). The system works on the following principle: In the real-time data collection step, running on the server (1) in the data center Real-time performance metrics of virtual servers (VMs) are accessed: CPU usage, memory Parameters such as consumption, storage I / O speeds, and network traffic are monitored at second-by-second intervals. Physical host energy consumption, temperature readings and hardware health data (fan Speeds, power supply status, and RAID system performance are all aspects of IoT 10. Data is collected through sensors. The collected data includes processing loads, CPU usage, memory status, network traffic, error logs, measurements from IoT devices, and user data. It includes a wide variety of information, such as activities. This rich dataset allows the system to... being able to monitor its performance in real time and quickly react to instantaneous changes. It enables adaptation. Especially the dynamic 15 of virtual servers on the data center. monitoring situations such as sudden traffic spikes or unexpected increases in load Data plays a critical role in enabling early detection and taking necessary precautions. The collection process is not limited to analyzing the current situation; it also early detection of potential anomalies, hardware failures, or software errors. It also provides important data for diagnosis. Thus, other process steps such as data preprocessing, 20 Robust for load forecasting, proactive resource activation, and dynamic energy optimization. It forms a foundation. Continuous data flow provides real-time information about the system's status and environmental factors. by evaluating and enabling the rapid restructuring of resources when necessary. He / She / It does. In the data preprocessing phase, the raw data collected in the real-time data collection step is processed. data coming from different sources, varying in format and quality The data is processed by the data preprocessing module (2). Provided via the server (1). Noise, missing or inconsistent records contained in the data are removed using special algorithms. The data is cleaned and normalized. In this process, entries containing errors are removed from the data. During the data retrieval process, predictive completion methods are employed for missing data. Thus, the collected data is standardized, and the following stages are carried out: It provides a solid foundation for load forecasting and proactive resource activation. Furthermore, the data... preprocessing stage, virtual server performance data (PV), CPU and memory 8 usage, network traffic, and other important parameters require data to be analyzed accurately. It ensures consistency. During the load forecasting phase, previously collected and pre-processed data is used over time. With the Temporal Fusion Transformer (TFT) model, which demonstrated superior performance in series analysis, 5 TFT is trained via training module (3). TFT can handle multiple time scales and heterogeneous data. by bringing together its resources, it analyzes sudden fluctuations, trends and seasonal changes in the system. It has the ability to analyze changes simultaneously. The model is particularly effective at self-attention. Thanks to self-attention mechanisms and multiple layers of attention, both long-term It is able to detect both long-term dependencies and short-term sudden increases, and accordingly 10 It can generate predictions for the future. These predictions are based on data centers and virtual reality. in infrastructures where servers are located (for example, virtual machines running on server (1)) taking proactive steps against sudden traffic increases that may occur (from machines) It provides. The load estimation phase is continuously updated during the model training process. It is supported by data and real-time feedback mechanisms; thus, the system 15 It can adapt dynamically, and prediction accuracy improves over time. TFT training During this process, relationships between static and dynamic data are established using static feature encoders and This is achieved by parsing the model using dynamic feature encoders; this allows the model to process all kinds of external and external features. It allows him to take the internal factor into account. In the proactive resource activation phase, in the previous steps (especially load estimation) Data obtained during the stage) are securely stored on the database (4). It is stored and constantly updated. This ensures efficient resource usage within the system. Status, performance metrics, error logs, and other critical data are displayed in real time. It can be monitored. In the proactive resource activation process, the database (4) 25 Based on current and accurate data provided, the decision support system (5) TFT deep learning algorithms are run on virtual servers and other computing systems. Server resources are activated in advance to prepare for expected future load increases. After the raw data obtained from (1) is cleaned in the data preprocessing stage, The database is saved to (4) and is ready for use by other modules of the system. This integration allows for quick response to sudden traffic increases or unexpected situations. By enabling intervention, it minimizes potential delays in the system and It guarantees service continuity. Thanks to proactive resource activation, virtual Server migration and restart processes are carried out before load increases. This completes the process, thus preventing performance degradation and interruptions. At the same time, 35 9 The continuously updated data stream provided by the database (4) improves the energy efficiency of the system. By integrating it into these strategies, it also helps reduce unnecessary energy consumption. In the dynamic energy optimization step, based on Temporal Fusion Transformer (TFT) Thanks to the DRS - decision support system (5) which works integrated with the prediction mechanism, 5 Current and future increases in workload within the system are determined in advance and adjusted accordingly. According to this, energy consumption is optimized. TFT-supported decision support system (5), obtained Using the generated load forecast data, the activity of physical servers on the data center can be determined. or dynamically regulates their transition to sleep mode. Thus, intensity Unnecessary energy consumption is prevented during off-peak periods, while momentary load increases are minimized. The necessary resources are made available in advance. This mechanism is the server (1) and other By working integrated with components, it improves the performance of virtual servers and physical hosts. It develops strategies to reduce energy costs while protecting energy. The TFT model Its predictive capabilities across multiple time scales enable dynamic energy optimization. It plays a key role; the system takes into account short-term and long-term energy consumption data, 15 It ensures the efficient allocation of resources in line with the defined strategies. Furthermore, in reality... Thanks to the timely feedback mechanism, prediction errors (TH) and sudden changes are minimized. By being detected immediately, the system's adaptation process is continuously updated to ensure effectiveness. It meets the needs in this way. Real-time feedback and adaptation process, TFT-supported decision support system. By working in integration with (5), the system is constantly updating and optimizing itself. This is the stage that enables it. In this stage, all the dynamics that occur in the system... Changes, real-time data collection, data preprocessing, load forecasting, proactive sourcing. Data obtained from previous steps such as activation and dynamic energy optimization 25 It is analyzed in the light of the TFT-supported decision support system (5), which is taken instantly. By evaluating performance data (PV), deviations or sudden changes in predictions can be identified. It detects that this has occurred. Thus, the virtual servers and physical servers within the system... Based on the current status of the hosts, resource allocation and energy consumption strategies are developed. It makes revisions. This feedback mechanism is especially useful for unexpected traffic increases, 30 It plays a critical role in situations such as hardware failures or unexpected software errors; Because the system instantly integrates these malfunctions into the model, addressing both existing problems and other issues. It also provides adaptive solutions to prevent similar problems in the future. It produces TFT-powered decision support through continuous monitoring and real-time updates. The system (5) optimizes system performance while maximizing energy efficiency. Furthermore, the feedback process ensures coordination among all components of the system. This allows the server (1) and other modules to achieve the specified goals. By working in harmony, it provides high accessibility and uninterrupted service. The adaptation mechanism increases the system's learning ability, becoming more efficient over time. It contributes to making accurate predictions and managing resource utilization more effectively. It is found. The invention offers high energy efficiency and superior performance in the fields of cloud computing and data center management. It presents an innovative system aimed at balancing performance. The invention, Temporal By offering a Fusion Transformer (TFT) based forecasting and resource management system, these 10 It provides solutions to problems. TFTs analyze multiple time intervals in time series data. by combining scales, heterogeneous data sources and contextual factors load It provides high accuracy in predictions. This allows for anticipation of future workload increases. By anticipating the load, the system ensures that the hosts are ready before the load arrives. It is equipped with proactive resource activation and dynamic energy optimization elements. 15 Using TFT's predictions, hosts are activated when and for as long as needed. For example, if a traffic surge is predicted to occur in 5 minutes, hosts should adjust the settings to 4.5 By being woken up minutes earlier, migration and service startup times fall within this time frame. This approach ensures that hosts are ready when the load arrives and reduces delays. This prevents unnecessary active hosts from occurring. Additionally, during low load periods, the number of unnecessary active hosts is reduced to 20. Energy savings are achieved by minimizing [the process]. Real-time feedback and Thanks to the adaptation mechanism, the system avoids prediction errors (TH) or sudden hardware failures. It can react instantly to situations such as malfunctions. This flexibility allows the system to continuously This enables improvement and rapid adaptation to changing conditions. As a result... Our invention radically solves the delayed scaling problem while achieving 25% energy efficiency. It simultaneously maximizes high performance. This innovative approach, existing Eliminating system deficiencies results in 30-40% energy savings and millisecond efficiency. It ensures compliance with the Service Level Agreement (SLA) at this level.

Claims

11 REQUESTS 1. Proactive resource activation and energy saving in data centers. It is a system, and its characteristic is;  physically and / or virtually characterized as located in a data center server (1), 5  Analysis of raw data collected from servers (1) located in the data center noisy, incomplete or inconsistent recordings to make them usable data preprocessing module (2) that enables cleaning and normalization of data.  The deep learning model to be implemented will use continuously updated data and Training with real-time feedback and static and dynamic data 10 relationships between them via static feature encoder and dynamic feature encoder training module that enables separation with (3),  centralized storage, management, and rapid access to all data database that enables access (4),  Data received from the mentioned servers (1) and processed through preprocessing steps 15 In light of this, proactively by running pre-trained deep learning algorithms. Instantaneous, enabling resource activation and dynamic energy optimization. Deviations in predictions or any sudden changes by looking at performance data (PV) Decision support takes action by identifying whether a change has occurred. system (5) 20 It includes.

2. Proactive resource activation and energy saving in data centers. It is a method, and its characteristic is;  physically and / or virtually characterized as located in a data center Data collected via server (1) can be processed by a data 25 Collection via preprocessing module (2) (1001),  Cleaning and processing of the collected data by the aforementioned preprocessing module (2) normalization processes (1002),  The deep learning model to be implemented will use continuously updated data and training through a training module (3) with real-time feedback and 30 relationships between static and dynamic data using static feature encoders and dynamic parsing via feature encoder (1003),  central storage of all data on a database (4), management and quick access (1004), 12  Data received from the mentioned servers (1) and processed through preprocessing steps In light of this, a decision support system capable of performing analysis using deep learning models. proactive resource activation and dynamic energy optimization of the system (5) (1005),  Looking at the instantaneous performance data (PV) of the mentioned decision support system (5) whether there have been any deviations from the predictions or any sudden changes. taking action by identifying (1006) It includes the steps of the process.

3. This method complies with Request-2 and its characteristics are; CPU usage, memory consumption, 10 Storage I / O speeds and network traffic parameters, physical host power consumption, Temperature readings and fan speeds, power supply status, RAID systems. the process of collecting work performance data via servers (1) It includes the step.

4. This method complies with Claim-2 and its characteristic is that the TFT deep learning model is continuous. A training module with updated data and real-time feedback. (3) Training through and static relationships between fixed and dynamic data parsing process using feature encoder and dynamic feature encoder It includes step 20.

5. The method is compliant with Request-2 and its feature is that it is received from the mentioned servers (1) and Based on data that has undergone preprocessing steps, TFT learning models a decision support system (5) that can perform analysis using proactive resources Ensuring activation and dynamic energy optimization takes process step 25. It includes.

Images