Method of Resource Scaling for Container-based Digital Twin Simulation and System for executing the same

The DLinear model-based resource scaling method addresses resource interference and inefficiencies in DT simulations by predicting future workloads and dynamically allocating resources, enhancing efficiency and cost optimization.

KR103047459B1Active Publication Date: 2026-11-11DONGGUK UNIVERSITY INDUSTRY ACADEMIC COOPERATION FOUNDATION
View PDF 4 Cites 0 Cited by

Patent Information

Application Number
KR1020240039896
Authority / Receiving Office
KR · KR
Patent Type
Patents
Current Assignee / Owner
Filing Date
2024-03-22
Publication Date
2026-11-11
Estimated Expiration
2044-03-22

AI Technical Summary

Technical Problem

Existing digital twin (DT) simulations in high-performance cloud computing environments face resource interference and inefficiencies due to the use of virtual machines with guest OSes, leading to slow startup times and increased memory consumption, and existing resource scaling methods do not effectively manage frequent workload fluctuations.

Method used

A resource scaling method using DLinear models to predict future workloads by training on historical computing resource usage trends and residuals, decomposing resource metrics into trend and residual sequences, and dynamically allocating computing resources based on predicted requirements.

Benefits of technology

The method efficiently manages resource utilization, minimizes waste, and optimizes costs by rapidly responding to workload fluctuations, improving scaling latency and simulation performance.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure 112024032581267-PAT00018_ABST
    Figure 112024032581267-PAT00018_ABST
Patent Text Reader

Abstract

A resource scaling method according to a disclosed embodiment comprises: a step of training a first DLinear model related to a trend of historical computing resource usage and a second DLinear model related to a residual of computing resource usage; a step of converting resource metrics received from a CI (Continuous Integration) server into a vector; a step of decomposing the converted vector of the resource metrics into a trend sequence and a residual sequence, and generating a computing resource requirement by combining a first output sequence inferred by inputting the trend sequence into the first DLinear model and a second output sequence inferred by inputting the residual sequence into the second DLinear model; a step of calculating the number of future replicas based on a container computing resource request received from the CI server and the resource requirement, and providing computing resources based on the calculated number of future replicas.
Need to check novelty before this filing date? Find Prior Art

Description

Technology Field

[0001] The present invention relates to a resource scaling method using a DLinear model applied in a container-based DT simulation environment and a system for executing the same. Background Technology

[0002] IoT (Internet of Things) technology connects various consumer electronics, such as refrigerators, smartphones, and smartwatches, to collect and share data. The advancement of such smart connected technology is gradually increasing the complexity of connected electronic devices over time.

[0003] In order to manufacture smart electronic devices using IoT technology, information communication technology (ICT) using IoT AI (IoT Artificial Intelligence) or big data technology is also being researched recently, and Industry 4.0, which applies the consumer electronics manufacture process, is rapidly emerging.

[0004] The core technology of Industry 4.0 lies in the Digital Twin (DT), and its implementation is crucial for realizing Digital Transformation (DX) across the full lifecycle of smart electronic devices. Here, a Digital Twin refers to a virtual model designed based on data collected from IoT sensors installed on actual physical objects, such as products or manufacturing equipment. Simulations using such Digital Twins analyze changes in the state of actual physical objects caused by various internal and external factors and predict the likelihood of failure. Furthermore, by monitoring the full lifecycle of the manufacturing process of smart electronic devices through Digital Twins, manufacturers can efficiently produce high-quality smart electronic devices.

[0005] However, because digital twins require a massive amount of computation during processing, it is difficult to run simulations on personal desktops with limited computing power. Consequently, manufacturers are deploying digital twins in high-performance cloud computing environments to perform simulations. Nevertheless, operating multiple digital twins with different conditions in parallel on the same physical hardware can lead to resource interference, which results in a problem of degraded simulation performance.

[0006] To address these issues, research is being conducted on DT simulation utilizing virtual machines (VMs). Virtual machine-based DT simulation prevents conflicts and resource interference between DT instances by running each DT instance in a dedicated virtual machine. Additionally, DT simulation efficiently utilizes resources by coordinating computing resources allocated to DT instances through provisioning. However, since virtual machines use a hypervisor, they include a separate guest OS (Operating System) and related libraries, which leads to the problem of the virtual machine itself becoming heavy. Furthermore, there were issues such as slow startup times due to guest OS booting and increased memory consumption to run the guest OS. A prior art document is Korean Registered Patent Publication No. 10-2561320 (July 27, 2023). The problem to be solved

[0007] One embodiment disclosed to solve the aforementioned problem proposes a resource scaling method for efficiently performing DT simulation in a high-performance cloud computing environment and a system for executing the same. means of solving the problem

[0008] A resource scaling method according to a disclosed embodiment comprises: a step of training a first DLinear model related to a trend of historical computing resource usage and a second DLinear model related to a residual of computing resource usage; a step of converting resource metrics received from a CI (Continuous Integration) server into a vector; a step of decomposing the converted vector of the resource metrics into a trend sequence and a residual sequence, and generating a computing resource requirement by combining a first output sequence inferred by inputting the trend sequence into the first DLinear model and a second output sequence inferred by inputting the residual sequence into the second DLinear model; a step of calculating the number of future replicas based on a container computing resource request received from the CI server and the resource requirement, and providing computing resources based on the calculated number of future replicas.

[0009] The learning step may include calculating the trend sequence by performing a moving average on the past computing resource usage based on the kernel size and stride size, and calculating the residual sequence as the difference between the past computing resource usage and the trend sequence.

[0010] The first DLinear model and the second DLinear model may include linear layers including weights and biases.

[0011] The step provided above can calculate the number of replicas based on the buffer size and scaling reference value.

[0012] The method may further include the step of comparing the predicted amount of computing resources with the actual usage of computing resources, and updating the first DLinear model and the second DLinear model based on the comparison result and a preset prediction error threshold value.

[0013] The step of providing the above may include comparing the number of replicas calculated and the number of replicas pre-allocated at a predetermined future point in time, and providing the computing resources based on the result of the comparison.

[0014] A system according to another disclosed embodiment comprises: an object; a cloud server that receives data collected from the object and performs a simulation of a digital twin based on the data; wherein the cloud server learns a first DLinear model related to a trend of past computing resource usage and a second DLinear model related to a residual of computing resource usage, converts resource metrics received from a CI server into a vector, decomposes the converted vector of the resource metrics into a trend sequence and a residual sequence, generates a predicted amount of computing resources by combining a first output sequence inferred by inputting the trend sequence into the first DLinear model and a second output sequence inferred by inputting the residual sequence into the second DLinear model, calculates the number of future replicas based on the container computing resource request amount received from the CI server and the predicted amount of resources, and provides computing resources to the object based on the calculated number of future replicas.

[0015] The cloud server can calculate the trend sequence by performing a moving average on the past computing resource usage based on the kernel size and stride size, and can calculate the residual sequence as the difference between the past computing resource usage and the trend sequence.

[0016] The first DLinear model and the second DLinear model may include linear layers including weights and biases.

[0017] The above cloud server can calculate the number of replicas based on the buffer size and scaling criteria.

[0018] The cloud server can compare the predicted amount of computing resources with the actual usage of computing resources, and update the first DLinear model and the second DLinear model based on the comparison result and a preset prediction error threshold value.

[0019] The cloud server can compare the number of replicas calculated and the number of replicas allocated in advance at a predetermined future point in time, and provide the computing resources based on the result of the comparison.

[0020] A cloud server according to another disclosed embodiment comprises: a physical layer that stores data collected from an object; a digital layer that stores data necessary for the simulation of a digital twin; and a deployment layer that includes a CI server and generates a container image. and a DT simulation layer that performs a simulation of a digital twin; wherein the DT simulation layer learns a first DLinear model related to a trend of past computing resource usage and a second DLinear model related to a residual of computing resource usage, collects resource metrics from the container image received from the CI server, converts the resource metrics into a vector, decomposes the converted vector of the resource metrics into a trend sequence and a residual sequence, generates a computing resource prediction amount by combining a first output sequence inferred by inputting the trend sequence into the first DLinear model and a second output sequence inferred by inputting the residual sequence into the second DLinear model, calculates the number of future replicas based on the container computing resource request amount received from the CI server and the resource prediction amount, and provides computing resources to the digital layer based on the calculated number of future replicas. Effects of the invention

[0021] The resource scaling method disclosed here and the system executing it can monitor the workload of a DT simulation running in a container and train multiple DLinear models with the workload, thereby predicting the workload at a specific point in time and effectively providing computing resources.

[0022] In addition, the disclosed resource scaling method and the system executing it utilize a lightweight DLinear model to rapidly predict future workloads, thereby improving scaling latency issues and supporting sustainable DT simulation.

[0023] In addition, the disclosed resource scaling method and the system implementing it can respond to frequent workload fluctuations in DT simulations, increase resource utilization, and minimize waste of computing resources.

[0024] The disclosed resource scaling method and the system implementing it can manage resources more efficiently than general DT simulation techniques that do not perform autoscaling, and can optimize the resource cost required for DT simulation. Brief explanation of the drawing

[0025] Figure 1 is a diagram illustrating a cloud computing environment for running DT simulations. Figures 2 and 3 are drawings for explaining the architecture of a disclosed cloud server. FIG. 4 is a flowchart for specifically explaining the disclosed resource scaling method. Figure 5 is a diagram illustrating an artificial intelligence model used in the disclosed resource scaling method. Figure 6 is a flowchart illustrating a method for training a disclosed artificial intelligence model. Figure 7 is a flowchart illustrating an update method performed during the process of providing computing resources by applying a resource scaling method. Specific details for implementing the invention

[0026] Throughout the specification, the same reference numerals refer to the same components. This specification does not describe all elements of the embodiments, and general content in the art to which the invention pertains or content that overlaps between embodiments is omitted.

[0027] Throughout the specification, when a part is described as being "connected" to another part, this includes not only cases where they are directly connected but also cases where they are indirectly connected, and indirect connections include connections made via a wireless communication network.

[0028] Furthermore, when it is stated that a part "includes" a certain component, this means that, unless specifically stated otherwise, it does not exclude other components but may include additional components.

[0029] Singular expressions include plural expressions unless there is an obvious exception in the context.

[0030] In addition, terms such as "~part," "~unit," "~block," "~part," and "~module" may refer to a unit that processes at least one function or operation. For example, the above terms may refer to at least one piece of hardware such as an FPGA (field-programmable gate array) or ASIC (application specific integrated circuit), at least one piece of software stored in memory, or at least one process processed by a processor.

[0031] The symbols attached to each step are used to identify each step and do not indicate the order of the steps relative to one another; the steps may be performed differently from the specified order unless a specific order is clearly indicated in the context.

[0032] Hereinafter, an embodiment relating to an electronic device and a control method according to one aspect will be described in detail with reference to the attached drawings.

[0033] Figure 1 is a diagram illustrating a cloud computing environment for running DT simulations.

[0034] Referring to FIG. 1, a cloud server (10) generates a digital twin (DT, 1-1) through the digital transformation (DX) of a physical object (1) and predicts the actual behavior of the object (1) through this.

[0035] The physical object (1) may be a robot arm used in the manufacturing field, as in the embodiment illustrated in FIG. 1. However, the disclosed embodiment is not necessarily limited to a robot arm and may include all physical objects that are reproduced in various industries and environments as well as manufacturing fields, including various mechanical devices.

[0036] A cloud server (10) can generate a virtual digital twin (1-1) based on a physical object (1). Here, a digital twin (DT) is a virtual model designed to accurately reflect the object (1). For example, a research subject such as a wind turbine must be equipped with various sensors in critical areas, and these sensors can generate data such as energy output, temperature, and weather conditions. Data collected from the sensors is transmitted to the cloud server (10) and applied to the digital twin (1-1). Based on the collected data, the cloud server (10) performs a simulation of the digital twin (1-1) and, through the simulation, can determine the optimal method of operation considering potential problems and cost aspects that may arise when applied to the physical object (1). The cloud server (10) provides the determined method of operation to the physical object (1).

[0037] A cloud server (10) is a physical or virtual infrastructure that provides applications, handles maintenance, or provides data storage. The cloud server (10) can be accessed by users on demand and can be built using virtualization software that divides into multiple virtual servers via a network, as well as a centralized server. A detailed description of the cloud server (10) will be provided later through the drawings below.

[0038] FIGS. 2 and 3 are diagrams illustrating the architecture of a disclosed cloud server. To avoid redundant descriptions, they are described together below.

[0039] FIGS. 2 and 3 are diagrams illustrating the architecture of a disclosed cloud server. To avoid redundant descriptions, they are described together below.

[0040] First, referring to FIG. 2, the cloud server (10) includes a physical layer (110), a digital layer (120), a deployment layer (130), and a DT simulation layer (140).

[0041] Specifically, the physical layer (110) includes various data that can be collected through the object (1). For example, the physical layer (110) may include various data such as sensor data, geometric data, operational data, and manufacturing data. Here, operational data and manufacturing data refer to a set of various data necessary for comprehensively managing various objects (1) in addition to sensor data, and may include all various data without being limited to specific names.

[0042] Various data from the physical layer (110) form the basis of the digital twin (1-1).

[0043] The digital layer (120) includes the results of processing and analyzing the data included in the physical layer (110), and for example, the digital layer (120) includes data such as computational data, DT configuration, DT simulation result, historical resource requirements, etc.

[0044] The deployment layer (130) containers the artificial intelligence model used in the DT application and the disclosed resource scaling method built based on the data of the physical layer (110) and the digital layer (120), and performs deployment that makes the compiled code executable.

[0045] Here, a container is a portable digital compartment that maintains a bundle of application files in a single run-time environment in the cloud. In other words, a container is a virtual space that isolates the environment for running a specific application within the cloud environment, and in the disclosed resource scaling method, this means distinguishing each digital twin among multiple digital twins (1-1) or distinguishing the act of a single digital twin operating itself.

[0046] The deployment layer (130) may be composed of a DT application Repository that collects and stores data stored by the physical layer (110) and the digital layer (120), a Prediction model Repository that stores artificial intelligence models required for the disclosed resource scaling method, a CI server (Continuous Integration server) that generates container images based on the collected data, and a Container Image Registry that stores and manages container images generated by the CI server.

[0047] The CI server compiles the data delivered from the DT application Repository and Prediction model Repository, tests the data, and builds a container image.

[0048] The DT simulation layer (140) simulates the container image generated from the deployment layer (130) and predicts and optimizes the performance of the object (1) through the simulation results. The DT simulation layer (140) consists of a master node (141), a model update node (142), and a simulation node (143). Specifically, the master node (141) operates the container by considering the usage status of computing resources. The model update node (142) and the simulation node (143) update the artificial intelligence model described below and execute the container image through the learned artificial intelligence model.

[0049] Referring to FIG. 3, the master node (141) of the DT simulation server (140) is composed of a kube-controller-manager that monitors various resources of the cluster, a kube-scheduler that deploys pods containing containers to optimal work nodes, etcd that stores information about cluster components, a kube-apiserver that enables communication between all components within the cluster, and an IReS (Intelligence Resource Scaling, 100) that executes an initiated resource scaling method.

[0050] Here, a pod is a Kubernetes Pod, which means a collection of one or more containers and is the smallest unit of a Kubernetes application. That is, the disclosed DT simulation layer (140) shares the same computing resources among containers belonging to the same pod, creates a cluster by pulling computing resources into Kubernetes, and implements a distributed cloud application execution system based on this.

[0051] The model update node (142) and the simulation node (143) may be composed of a container runtime responsible for running the container, a kubelet managing the execution state of the container, a kube-proxy ensuring mutual network connectivity between pods, and a Prometheus that collects and stores time-series metric data.

[0052] The model update node (142) and the simulation node (143) perform different roles in executing the container. Specifically, the model update node (142) contains a container that updates the artificial intelligence model of IReS (100), whereas the simulation node (143) contains a container that simulates the digital twin. In particular, the model update node (142) aims to update the parameters of the artificial intelligence model running in IReS (100). Therefore, it requires relatively less computing resources compared to the simulation node (143), which continuously processes computationally intensive tasks such as DT simulation.

[0053] To efficiently provide the computing resources required for DT simulation, the Prometheus of the worker node transmits historical computing resource usage collected via the kube-apiserver of the master node (141) to the IReS (100). The IReS (100) calculates the number of replicas required in the future based on the predicted future computing resource usage, and then transmits a request signal to execute computing resource scaling to the kube-controller-manager via the kube-apiserver. The kube-controller-manager adjusts the number of replicas according to the aforementioned request.

[0054] When the prediction performance of IReS (100) deteriorates, the model update node (142) updates the parameters of the artificial intelligence model (prediction model) and then transmits them to IReS (100) via kube-apiserver to reduce the prediction error. A detailed explanation regarding this will be provided later through FIG. 7, etc.

[0055] The simulation results generated from the DT simulation layer (140) are transmitted to the physical layer (110) via the kube-apiserver of the master node (141), and the physical layer (110) updates sensor data, operation parameters, and other physical attributes. Additionally, the simulation results are transmitted to the digital layer (120) to update the computational model, virtual representation, etc., of the physical object system. The updated physical layer (110) and digital layer (120) are fed back to the DT simulation layer (140) to improve the accuracy of the DT simulation.

[0056] FIG. 4 is a flowchart for specifically explaining the disclosed resource scaling method. FIG. 5 is a diagram for explaining the artificial intelligence model used in the disclosed resource scaling method. To avoid redundancy, they are described together below.

[0057] The disclosed resource scaling method dynamically scales the computing resources of a container to improve resource utilization for DT simulation.

[0058] The disclosed resource scaling method performs a resource metric transformation that converts resource metrics received from a CI server into a vector (step 210), and performs a resource requirements prediction that predicts future workloads from computing resource metrics at a past point in time (steps 220 to 250). Finally, the disclosed resource scaling method performs a pod replica calculation step to predict resource usage (steps 260 to 270).

[0059] Referring to FIG. 4, the disclosed resource scaling method trains a first DLinear model and a second DLinear model (200).

[0060] The disclosed resource scaling method trains an artificial intelligence model in advance before predicting resource usage. The training of the artificial intelligence model is handled by the model update node (142) described in FIG. 3.

[0061] The artificial intelligence model used in the disclosed resource scaling method includes a lightweight deep learning model, the DLinear Model, and distinguishes between a first DLinear Model related to the trend of past computing resource usage and a second DLinear Model related to the residual of computing resource usage.

[0062] Referring to FIG. 5, the DLinear model (300) is formed with a simple structure consisting only of linear layers (301, 302). Specifically, when an input sequence (time series data) is input to the DLinear model (300), the DLinear model (300) outputs an output sequence (time series data). Because the DLinear model (300) has a small number of parameters and a fast computation speed, it has lower computational complexity and consumes fewer computing resources than existing time series prediction models.

[0063] Referring again to FIG. 4, the disclosed resource scaling method converts resource metrics into a vector (210).

[0064] Specifically, the disclosed resource scaling method collects core metrics of a container in which a DT simulation is running through a full monitoring pipeline. The disclosed resource scaling method extracts CPU usage metrics and memory usage metrics from the core metrics to account for CPU usage and memory usage, which have a direct impact on the performance degradation of the DT simulation. Hereinafter, CPU usage metrics and memory usage metrics are referred to as computing resource metrics.

[0065] The extracted resource metrics can be structured in JavaScript Object Notation (JSON) format. Therefore, the disclosed resource scaling method converts the resource metrics into a vector.

[0066] The disclosed resource scaling method performs decomposition of the transformation vector of the resource matrix (220).

[0067] The disclosed resource scaling method performs decomposition to analyze the underlying patterns of computing resource usage. That is, historical computing resource usage is separated into a trend sequence and a residual sequence. Here, the trend refers to the long-term pattern of computing resource usage, and the residual refers to the remaining values ​​after subtracting the trend from the historical computing resource usage.

[0068] The past computing resource usage decomposed by the disclosed resource scaling method can be expressed as Equation 1 below.

[0069] [Mathematical Formula 1]

[0070]

[0071] Here represents the input sequence of computing resources (r) at a time stamp (t).

[0072] The disclosed resource scaling method calculates a trend sequence by performing a moving average on past computing resource usage based on the kernel size and stride size. The aforementioned method for calculating the trend sequence can be expressed as shown in Equation 2 below.

[0073] [Mathematical Formula 2]

[0074]

[0075] Here is kernel size and means stride size, and represents the input sequence, i.e., the trend sequence. The residual sequence is unpredictable noise that is not captured by the basic pattern, and is calculated through Equation 3 below.

[0076] [Mathematical Formula 3]

[0077]

[0078] The disclosed resource scaling method outputs a first output sequence inferred by inputting a trend sequence into a first DLinear model (231, 232). At the same time, the disclosed resource scaling method outputs a second output sequence inferred by inputting a residual sequence into a second DLinear model (233, 234).

[0079] In order to preserve the temporal information of the trend sequence and residual sequence decomposed at step 200, the first DLinear model and the second DLinear model have each learned the unique features of the trend sequence and residual sequence as shown in Equation 4.

[0080] [Mathematical Formula 4]

[0081]

[0082] Here, L tr is the first DLinear model, L re represents the second DLinear model. α and β represent the input data, namely the trend sequence and the residual sequence, and represents the weight, and ε represents the bias.

[0083] The disclosed resource scaling method combines a first output sequence inferred by inputting a trend sequence into a first DLinear model and a second output sequence inferred by inputting it into a second DLinear model (250).

[0084] The method of combining the first output sequence and the second output sequence can be expressed as shown in Equation 5 below.

[0085] [Mathematical Formula 5]

[0086]

[0087] The disclosed resource scaling method generates a computing resource requirement by combining a first output sequence and a second output sequence. The computing resource requirement is a computing resource predicted to be requested at a future point in time. However, the disclosed resource scaling method provides computing resources by taking into account a prediction error through the following steps.

[0088] The disclosed resource scaling method calculates the number of future replicas (260) and provides computing resources (270).

[0089] The disclosed resource scaling method calculates the number of future replicas based on the amount of container computing resource requests received from the CI server and the predicted amount of computing resources calculated in step 250. Specifically, the disclosed resource scaling method calculates the number of replicas (as shown in Equation 6 below) Produces ).

[0090] [Mathematical Formula 6]

[0091]

[0092] Here is the buffer size, represents the scaling threshold for computing resources.

[0093] The disclosed resource scaling method calculates computing resources based on the calculated number of future replicas and can provide computing resources thereby. The method of calculating computing resources based on the number of replicas is the same as a general scaling method.

[0094] Figure 6 is a flowchart illustrating a method for training a disclosed artificial intelligence model.

[0095] Referring to FIG. 6, a method for training an artificial intelligence model converts resource metrics into vectors (201) and decomposes the converted vectors of resource metrics (202).

[0096] Here, the resource metrics are resource metrics generated in the target DT simulation environment, which may be based on past computing resource usage or may have been manipulated and modified by the user. The vector transformation and decomposition method of the resource metrics is the same as that described in Fig. 4.

[0097] A method for training an artificial intelligence model outputs a first output sequence inferred by inputting a decomposed trend sequence into a first DLinear model (203, 204). At the same time, outputs a second output sequence inferred by inputting a residual sequence into a second DLinear model (205, 206).

[0098] A method for training an artificial intelligence model combines the first output sequence and the second output sequence that are output (207).

[0099] As described above in FIG. 4, the result of combining the first output sequence and the second output sequence is a predicted amount of computing resources. Accordingly, a method for training an artificial intelligence model compares the combined computing resource usage with the next timestamp of past computing resources (208).

[0100] For example, the resource metrics used for training the artificial intelligence model may be computing resources generated at the time stamp of time t-2. Accordingly, the training of the initiated artificial intelligence model evaluates the performance of the DLinear model (300) by comparing the computing resource usage output from the DLinear model (300) with the computing resources generated at the time stamp of time t-1.

[0101] Based on the evaluation results, the artificial intelligence model's training method trains the DLinear model (300) by adjusting the parameters. There are various ways to adjust the parameters, and this is done by adjusting the loss function of Equation 4.

[0102] Figure 7 is a flowchart illustrating an update method performed during the process of providing computing resources by applying a resource scaling method.

[0103] The disclosed resource scaling method undergoes the vector transformation step described in FIG. 4 and the step of generating a computer resource prediction amount from the DLinear model (300). However, additional updates are performed in the process of calculating a replica and providing computing resources based thereon.

[0104] Referring to FIG. 7, the disclosed resource scaling method pre-trains a first DLinear model and a second DLinear model (271), converts resource metrics into vectors (272), and generates a computing resource prediction amount (273).

[0105] The disclosed resource scaling method compares the number of replicas calculated at a future point in time with the number of replicas allocated (274).

[0106] The fields in which DT simulations are applied can be diverse, and accordingly, simulations may be executed under various conditions. Therefore, the error between the predicted amount of computing resources predicted by the disclosed resource scaling method and the actual predicted amount of computing resources may gradually increase. To address this, the disclosed resource scaling method utilizes a predict error threshold.

[0107] Specifically, the disclosed resource scaling method compares the prediction error with a reference value, i.e., the prediction error reference value (275).

[0108] If the prediction error exceeds the prediction error threshold value (Yes of 275), the DLinear model (300) is updated by additionally learning new past computing resource usage.

[0109] If the prediction error is less than the prediction error threshold value (No. 275), the disclosed resource scaling method provides computing resources based on the calculated number of replicas (276).

[0110] The disclosed resource scaling method can predict the workload at a specific point in time during a DT simulation running in a container and effectively provide computing resources. Additionally, the disclosed resource scaling method utilizes a lightweight DLinear model to rapidly predict future workloads, thereby improving scaling latency issues and supporting sustainable DT simulation. Furthermore, the disclosed resource scaling method can respond to frequent workload fluctuations in DT simulation, increase resource utilization, minimize waste of computing resources, manage resources more efficiently than general DT simulation techniques that do not perform auto-scaling, and optimize the resource costs required for DT simulation. Explanation of the symbols

[0111] 1: Object 1-1: Digital Twin 10: Cloud Server

Claims

Claim 1 The method comprises the steps of: training a first DLinear model related to the trend of historical computing resource usage and a second DLinear model related to the residual of computing resource usage; converting resource metrics received from a CI (Continuous Integration) server into vectors; decomposing the transformed vector of the converted resource metrics into a trend sequence and a residual sequence, and generating a computing resource requirement by combining a first output sequence inferred by inputting the trend sequence into the first DLinear model and a second output sequence inferred by inputting the residual sequence into the second DLinear model; calculating the number of future replicas based on the container computing resource request received from the CI server and the resource requirement, and providing computing resources based on the calculated number of future replicas; wherein the step of providing comprises calculating the number of replicas based on a buffer size and a scaling reference value, A resource scaling method that calculates a value based on dividing the value reflecting the buffer size in the above resource prediction amount by the product of the above container computing resource request amount and the above scaling reference value. Claim 2 A resource scaling method according to claim 1, wherein the learning step comprises calculating the trend sequence by performing a moving average on the past computing resource usage based on a kernel size and a stride size, and calculating the residual sequence as the difference between the past computing resource usage and the trend sequence. Claim 3 In claim 1, the first DLinear model and the second DLinear model are a resource scaling method comprising a linear layer including weights and biases. Claim 4 delete Claim 5 A resource scaling method according to claim 1, further comprising the step of comparing the predicted amount of computing resources with the actual usage of computing resources, and updating the first DLinear model and the second DLinear model based on the comparison result and a preset prediction error reference value. Claim 6 A resource scaling method according to claim 1, wherein the step of providing comprises comparing the number of replicas calculated and the number of replicas allocated in advance at a predetermined future point in time, and providing the computing resources based on the result of the comparison. Claim 7 Object: A cloud server that receives data collected from the object and performs a simulation of a digital twin based on the data; wherein the cloud server learns a first DLinear model related to a trend of past computing resource usage and a second DLinear model related to a residual of computing resource usage, converts a resource matrix received from a CI server into a vector, decomposes the converted vector of the converted resource matrix into a trend sequence and a residual sequence, generates a predicted computing resource amount by combining a first output sequence inferred by inputting the trend sequence into the first DLinear model and a second output sequence inferred by inputting the residual sequence into the second DLinear model, calculates the number of future replicas based on the container computing resource request amount received from the CI server and the predicted resource amount, provides computing resources to the object based on the calculated number of future replicas, and calculates the number of replicas based on a buffer size and a scaling reference value, wherein the value reflecting the buffer size in the predicted resource amount is calculated based on the value obtained by dividing the value by the product of the container computing resource request amount and the scaling reference value. Claim 8 In claim 7, the cloud server calculates the trend sequence by performing a moving average on the past computing resource usage based on the kernel size and stride size, and calculates the residual sequence as the difference between the past computing resource usage and the trend sequence. Claim 9 In claim 7, the first DLinear model and the second DLinear model are a system comprising linear layers including weights and biases. Claim 10 delete Claim 11 In claim 7, the cloud server is a system that compares the predicted amount of computing resources with the actual usage of computing resources, and updates the first DLinear model and the second DLinear model based on the comparison result and a preset prediction error reference value. Claim 12 In claim 7, the cloud server is a system that compares the calculated number of replicas with the pre-allocated number of replicas at a predetermined future point in time and provides the computing resources based on the result of the comparison. Claim 13 A physical layer that stores data collected from objects; a digital layer that stores data required for the simulation of a digital twin; a deployment layer that includes a CI server and generates a container image; and a DT simulation layer that performs a simulation of a digital twin; wherein the DT simulation layer learns a first DLinear model related to the trend of past computing resource usage and a second DLinear model related to the residual of computing resource usage, collects resource metrics from the container image received from the CI server, converts the resource metrics into a vector, decomposes the converted vector of the converted resource metrics into a trend sequence and a residual sequence, generates a computing resource prediction amount by combining a first output sequence inferred by inputting the trend sequence into the first DLinear model and a second output sequence inferred by inputting the residual sequence into the second DLinear model, calculates the number of future replicas based on the container computing resource request amount received from the CI server and the resource prediction amount, provides computing resources to the digital layer based on the calculated number of future replicas, and calculates the number of replicas based on a buffer size and a scaling reference value, wherein the value reflecting the buffer size in the resource prediction amount is calculated based on the value obtained by dividing the value by the product of the container computing resource request amount and the scaling reference value. Cloud server.

Citation Information

Patent Citations

  • Cloud computing apparatus for supporting resource allocation based on container and cloud computing method for the same

    KR1020190070659A

  • Container scheduling method and device and non-volatile computer-readable storage medium

    KR1020220100982A

  • A container replica recommendation system through resource trend prediction and a recommendation method

    KR102561320B1

  • Method and apparatus for structured pruning with linear transform parameters in for deep neural networks

    KR102574489B1