Method for reporting multi-dimensional perception of edge computing-based computing network resource state

Abstract

The application discloses an edge-computing-based multi-dimensional perception and reporting method for computing network resource state, and relates to the technical field of computing network resource management. The method comprises the following steps: deploying a lightweight edge computing unit on each computing network node to collect computing power, storage, network, load and environment dimension running data nearby; performing abnormality elimination, standardization and hierarchical quantization to output a simplified resource state; calling a standard task library to match resource demand calculation adaptation degree and generate a task adaptation label; inputting the simplified state into a time series prediction model to identify abnormalities and divide early warnings, and generate a local pre-scheduling scheme; and encapsulating the simplified state, the adaptation label, the early warning and the scheme, and reporting them to a central server in an incremental manner through a lightweight protocol. The application solves the problems of large computing pressure, data redundancy and poor real-time performance of the central server in the prior art.

Description

Technical Field

[0001] This application relates to the field of computing network resource management technology, and in particular to a method for multi-dimensional perception and reporting of computing network resource status based on edge computing. Background Technology

[0002] Resource status awareness and reporting are the core foundation of computing network scheduling. Its core objective is to collect the operational status of computing, storage, and network resources in real time and accurately, and efficiently aggregate this data to the scheduling center, providing a reliable basis for global resource allocation and task scheduling. With the large-scale deployment of computing networks, the number of network nodes has surged, and resource heterogeneity has intensified. Traditional awareness and reporting models are gradually revealing their bottlenecks. Among existing related technologies, a typical solution is a centralized resource scheduling architecture, where a central server uniformly receives the raw resource data reported by each node, performs data cleaning, status analysis, and capability assessment, and then executes global scheduling decisions.

[0003] This centralized processing model can operate stably in small-scale computing network scenarios, but it has significant drawbacks when facing large-scale, highly dynamic computing networks. The main problem is the overload of the central server's computing pressure: on the one hand, each computing network node needs to report all raw sensing data, resulting in a massive influx of raw data into the center, causing a sharp increase in the center's data receiving and caching pressure; on the other hand, the center needs to perform calculations such as cleaning, deduplication, normalization, state assessment, and task adaptation for each stream of raw data. As the number of nodes increases, the computational complexity increases exponentially, easily leading to processing delays and response timeouts, and even causing system lag or crashes, failing to meet real-time scheduling requirements. At the same time, the high redundancy and large transmission volume of raw data occupy core network bandwidth, further exacerbating the overall performance bottleneck of the system. Therefore, there is an urgent need for a technology that can offload the pressure on the center and improve the efficiency of sensing reporting. Summary of the Invention

[0004] This application provides a multi-dimensional perception and reporting method for computing network resource status based on edge computing, which solves the problems of high computing pressure on the central server, high redundancy of raw data, and insufficient real-time perception and reporting in the prior art.

[0005] On the one hand, embodiments of this application provide a multi-dimensional perception and reporting method for the status of computing network resources based on edge computing, including: S1. Deploy lightweight edge computing units at each node of the computing network. The edge computing units collect operational data on computing power, storage, network, load, and environment of the nodes nearby. The collected operational data is then subjected to anomaly removal and standardization processing in sequence, and resource capacity is quantified according to hierarchical rules to obtain simplified resource status data. S2. Retrieve the standard task type library, extract the resource requirement conditions corresponding to each type of task, match the simplified resource status data with the resource requirement conditions of each type of task one by one, calculate the corresponding degree of adaptation, and generate task adaptation tags based on the ranking results of the degree of adaptation. S3. Input the simplified resource status data into the lightweight time series prediction model and output the prediction results of the resource status; based on the prediction results, identify abnormal resource trends and classify warning levels; combine the simplified resource status data and task adaptation tags to generate a local pre-scheduling scheme that includes scheduling operations, operation parameters and execution timing. S4. Encapsulate the simplified resource status data, task adaptation tags, early warning levels, and local pre-scheduling scheme into structured data, and report it to the central server using an incremental transmission method and a lightweight communication protocol.

[0006] On the other hand, this application also provides a multi-dimensional perception and reporting system for computing network resource status based on edge computing. This system includes lightweight edge computing units deployed at various nodes of the computing network. The edge computing units collect operational data from nearby nodes, considering dimensions such as computing power, storage, network, load, and environment. The collected operational data undergoes anomaly removal and standardization processing sequentially, and then resource capacity quantification is performed according to hierarchical rules to obtain simplified resource status data. The edge computing units retrieve standard task type libraries, extract resource requirement conditions corresponding to various tasks, and match the simplified resource status data with the resource requirement conditions of each task. The edge computing unit calculates the corresponding adaptation degree and generates task adaptation labels based on the ranking results of the adaptation degree. The edge computing unit inputs the simplified resource status data into the lightweight time series prediction model and outputs the prediction results of the resource status. Based on the prediction results, it identifies abnormal resource trends and classifies warning levels. Combining the simplified resource status data and task adaptation labels, it generates a local pre-scheduling scheme that includes scheduling operations, operation parameters, and execution timing. The edge computing unit encapsulates the simplified resource status data, task adaptation labels, warning levels, and local pre-scheduling scheme into structured data and reports it to the central server through a lightweight communication protocol using incremental transmission.

[0007] The edge computing-based multi-dimensional perception and reporting method for network resource status in this application has the following advantages: 1. Reduce the computational burden on the central computing center. This application offloads all computational tasks, such as data cleaning, status quantification, capability assessment, task adaptation, early warning analysis, and pre-scheduling scheme generation, to the edge computing unit. The central computing center only receives simplified results and does not need to process massive amounts of raw data, thus solving the problem of central computing overload and significantly improving scheduling response speed.

[0008] 2. Multi-dimensional perception + lightweight reporting reduces bandwidth overhead. This application constructs a perception system containing 13 indicators, ensuring comprehensive and accurate data. After noise reduction and incremental reporting by the edge computing unit, the amount of transmitted data is compressed by more than 70%, saving bandwidth in the core network and avoiding signaling storms.

[0009] 3. Real-time edge early warning + pre-schedule generation enhances scheduling foresight. In this application, the edge computing unit predicts resource status in the short term, identifies risks such as overload and bottlenecks in advance, and issues timely warnings; at the same time, it generates local pre-schedule plans, clarifying resource allocation and adaptation tasks, which can support pre-schedule and avoid congestion after scheduling. Attached Figure Description

[0010] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 A flowchart of a multi-dimensional perception and reporting method for the status of computing network resources based on edge computing provided in this application embodiment. Detailed Implementation

[0012] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0013] Figure 1 A flowchart illustrating a multi-dimensional perception and reporting method for the status of computing network resources based on edge computing, provided in this application embodiment. This application embodiment provides a multi-dimensional perception and reporting method for the status of computing network resources based on edge computing, including: S1. Deploy lightweight edge computing units at each node of the computing network. The edge computing units collect operational data on computing power, storage, network, load, and environment of the nodes nearby. The collected operational data is then subjected to anomaly removal and standardization processing in sequence, and resource capacity is quantified according to hierarchical rules to obtain simplified resource status data.

[0014] For example, the specific content of step S1 is as follows: S1.1 Construction of a Multi-Dimensional Perception Indicator System. The edge computing unit incorporates standardized perception templates, constructing a perception system encompassing five dimensions and a total of thirteen indicators. This system covers the hardware, performance, and environmental dimensions of the computing network resource operation. The specific classification and definitions of these indicators are as follows: Computing power dimension: Focusing on the operational status of computing, it includes four indicators: processor utilization, core load, cache hit rate, and clock speed utilization. Among them, processor utilization refers to the proportion of processors used per unit time, reflecting the level of computing activity; core load refers to the number of processor cores used, reflecting the scale of computing power supply; cache hit rate is the proportion of requests that are hit by the processor cache, reflecting the efficiency of data access; and clock speed utilization is the ratio of the processor's actual operating clock speed to its rated clock speed, reflecting the level of computing power performance.

[0015] Storage dimension: Focusing on the operational status of storage media, including four indicators: remaining capacity, read / write speed, IOPS, and storage utilization. Remaining capacity refers to the unused space on the storage media, reflecting storage supply capacity; read / write speed refers to the amount of data read / written per unit time, reflecting storage performance; IOPS is the number of input / output operations per second, reflecting storage concurrency capability; and storage utilization is the percentage of used space, reflecting the storage load level.

[0016] Network dimension: Focusing on data transmission quality, it includes three indicators: transmission latency, bandwidth utilization, and packet loss rate. Transmission latency refers to the end-to-end transmission time of data, reflecting network response speed; bandwidth utilization refers to the proportion of bandwidth used, reflecting the network's busyness; packet loss rate is the proportion of data packets lost due to transmission, reflecting network stability.

[0017] Load dimension: Focusing on the overall operating pressure of nodes, including the overall node load, it comprehensively reflects the overall resource consumption level of computing network nodes.

[0018] Environmental dimension: Focusing on the hardware operating environment, including the hardware temperature indicator, which reflects the operating temperature of core hardware such as chips and motherboards.

[0019] S1.2 Raw Data Acquisition. Edge computing units are deployed locally on each computing network node. They utilize three proximity-based acquisition channels: hardware sensors, out-of-band management interfaces, and host machine agents, to achieve latency-free and highly reliable raw data acquisition. Specific acquisition methods and frequencies are as follows: Hardware sensors directly collect low-level data such as hardware temperature, clock frequency utilization, cache hit rate, and core load; out-of-band management interfaces collect hardware operation data such as transmission latency, processor utilization, bandwidth utilization, packet loss rate, IOPS, and storage utilization; host agents collect system-level data such as storage read / write speed, remaining capacity, and overall node load. These three channels complement each other to ensure the integrity of data collection.

[0020] Based on the dynamic nature of the indicators, this application embodiment sets the following differentiated collection frequencies to balance data real-time performance and collection overhead: 1. High-frequency acquisition metrics: processor utilization, transmission latency, overall node load and packet loss rate, with an acquisition interval of 100ms, capable of capturing real-time dynamic changes.

[0021] 2. Intermediate frequency acquisition indicators: core load, bandwidth utilization, read / write speed, IOPS and main frequency utilization, with an acquisition interval of 500ms to balance dynamism and stability.

[0022] 3. Low-frequency acquisition indicators: remaining capacity, hardware temperature, cache hit rate and storage utilization. The acquisition interval is 5 seconds to adapt to the slow change characteristics of these indicators and reduce acquisition power consumption.

[0023] The collected data is stored directly in the local memory of the edge computing unit, without the need for cross-node transmission, thus avoiding raw data leakage and transmission delays and ensuring data integrity and timeliness.

[0024] S1.3 Anomaly Removal. The raw data collected is susceptible to hardware fluctuations and signal interference, leading to random outliers. Therefore, invalid data needs to be removed using a sliding window statistical method. The specific processing procedure is as follows: The algorithm employs a combination of sliding window and 3σ criterion, with the following formula: ; ; Where μ is the mean of all original data within the sliding window, reflecting the central tendency of the data within the window; w is the length of the sliding window, a preset fixed value, generally taken as 5, used to limit the range of statistical data and avoid the data volume being too large or too small; p is the index of the data within the sliding window, used to identify different original data within the window; x p σ represents the p-th original data point within the sliding window, i.e., the original index value collected; σ is the standard deviation of the data within the sliding window, reflecting the degree of dispersion of the data within the window.

[0025] After calculating the mean μ and standard deviation σ within the window, if a certain original data x p Satisfy |x p If -μ|>3σ, it is considered abnormal data and is directly removed; only the valid data within the window is retained to form a denoised data sequence.

[0026] S1.4 Standardization Processing. Valid data after anomaly removal may have dimensional differences (e.g., utilization rate as a percentage, capacity in GB). Normalization is required to eliminate these differences, unify the data value range, and provide standardized data for subsequent quantification and adaptation. The specific formula is as follows: Normalization formula: ; in, The normalized data has a fixed value range of [0, 1] to eliminate dimensional differences; The valid data after anomaly removal, i.e., the denoised index values; The historical minimum value of the corresponding indicator is obtained through local historical data statistics of the edge computing unit; The historical maximum value of the corresponding indicator is also obtained through local historical data statistics of the edge computing unit.

[0027] The standardization process is performed on each of the thirteen indicators, and all data are mapped to the same value range after processing, which facilitates subsequent hierarchical quantification and adaptation matching.

[0028] S1.5, Hierarchical Quantization. The standardized data is continuous numerical and needs to be transformed into discrete and unique simplified resource status data through preset hierarchical rules to reduce data complexity. The specific hierarchical rules are as follows: A five-level quantization rule is adopted, based on normalized data. The range is divided into five ability levels, and the level division criteria are as follows: Level 1: 0≤ <0.2 indicates extremely poor resource capabilities; Level 2: 0.2≤ A value less than 0.4 indicates poor resource capacity. Level 3: 0.4≤ <0.6 indicates moderate resource capacity; Level 4: 0.6≤ <0.8 indicates good resource capacity; Level 5: 0.8≤ ≤1.0 indicates excellent resource capabilities.

[0029] After grading and quantification, a unique grade identifier is generated for each of the thirteen indicators, which are then combined to form simplified resource status data. The simplified resource status data only contains grade identifiers and has no original values. The data volume is compressed by more than 80%, achieving the goal of data simplification and providing a low-redundancy data foundation for subsequent adaptation, prediction and reporting.

[0030] S2. Retrieve the standard task type library, extract the resource requirement conditions corresponding to each type of task, match the simplified resource status data with the resource requirement conditions of each type of task one by one, calculate the corresponding degree of adaptation, and generate task adaptation tags based on the ranking results of the degree of adaptation.

[0031] For example, the specific content of step S2 is as follows: S2.1 Standard Task Type Library Construction. The edge computing unit has a built-in standard task type library, covering the mainstream task types of computing power networks, totaling six categories. Each task type has clearly defined and quantifiable resource requirements, without ambiguity. The specific task types and requirements are as follows: AI (Artificial Intelligence) training task: The core requirements are high computing power, large storage, and low latency. The requirements are processor utilization ≤80%, core load ≥16 cores, remaining capacity ≥500GB, and transmission latency ≤20ms. AI inference task: The core requirements are low latency and light load, with the following conditions: processor utilization ≤70%, transmission latency ≤10%, and overall node load ≤75%; General computing tasks: The core requirements are stable computing power and moderate load. The requirements are: core load ≥ 8 cores, processor utilization ≤ 85%, and hardware temperature ≤ 85℃. Storage-intensive tasks: The core requirements are large capacity and high read / write performance, with the following conditions: remaining capacity ≥ 1TB, read / write speed ≥ 500MB / s, and storage utilization ≤ 80%; Network-intensive tasks: The core requirements are high bandwidth and low jitter, with the following conditions: bandwidth utilization ≤ 60%, transmission latency ≤ 15ms, and packet loss rate ≤ 0.5%; Lightweight edge tasks: The core requirements are low power consumption and low load, with the following conditions: processor utilization ≤50%, overall node load ≤60%, and hardware temperature ≤75℃.

[0032] The standard task type library is built into the edge computing unit locally, eliminating the need to download it from the center and adapting to offline edge operation scenarios. Resource requirements can be dynamically updated based on the needs of the computing network services, and the updates take effect automatically without requiring a restart of the edge computing unit.

[0033] S2.2 Adaptability Calculation. The edge computing unit retrieves the standard task type library and simplified resource status data, and calculates the adaptability between resources and each type of task using a weighted matching algorithm. The specific process is as follows: Formula for calculating compatibility: ; Among them, Score k λ represents the suitability of the k-th task type, with a value ranging from [0, 1]. A larger value indicates higher suitability. k is the index of the task type, corresponding to six task types, with a value from 1 to 6. n is the number of requirement indicators for the k-th task, i.e., the number of resource requirement conditions corresponding to this task. j is the index of the requirement indicator, used to identify different resource requirement conditions under the same task. j Let be the weight of the j-th demand indicator, taking values ​​[0, 1], satisfying The weights are dynamically adjusted based on task requirements; a jTo simplify the numerical values ​​corresponding to the level of the j-th indicator in the resource status data, level one corresponds to 1, level two to 2, ..., level five to 5; r k，j is the minimum level requirement corresponding to the j-th requirement condition of the k-th task; f is a matching function used to determine whether the indicator level meets the requirement condition. If it meets the requirement condition, it is assigned a value of 1; if it partially meets the requirement condition, it is assigned a value of 0.5; if it does not meet the requirement condition, it is assigned a value of 0.

[0034] The adaptation degree calculation process is as follows: For the six types of tasks, the requirement indicators, weights and level requirements are extracted one by one, and the adaptation degree is calculated by substituting them into the formula. This calculation is completed locally at the edge without uploading data to the center, thus avoiding cross-node calculation delays.

[0035] S2.3 Adaptation Tag Generation. After completing the adaptation degree calculation, the edge computing unit's adaptation degree score for the six types of tasks is generated. k Sort the tasks in descending order and select the task type that appears first in the sorted list as the unique task adaptation tag, according to the following rules: If the highest fit score max If the value is ≥0.6, then a task adaptation tag for the corresponding task type will be generated. If all compatibility scores k If the value is less than 0.6, a "No Adaptation Task" label will be generated.

[0036] Task adaptation tags are standardized text identifiers with a concise and unique format, such as "AI training" or "edge lightweight task." The data size is small, facilitating subsequent encapsulation and reporting. The generation of task adaptation tags is completed entirely locally at the edge, without central intervention, further simplifying the central decision-making process.

[0037] S3. Input the simplified resource status data into the lightweight time series prediction model and output the prediction results of the resource status; based on the prediction results, identify abnormal resource trends and classify warning levels; combine the simplified resource status data and task adaptation tags to generate a local pre-scheduling scheme that includes scheduling operations, operation parameters and execution timing.

[0038] For example, the specific content of step S3 is as follows: S3.1 Lightweight Time-Series Prediction Model Construction and Deployment. To adapt to the low computing power and low power consumption operating environment of edge computing units, this application embodiment constructs a lightweight improved recurrent neural network (Edge-Long Short-Term Memory Network Edge-LSTM) for short-term trend prediction of resource status. The model structure, parameters, and training process are as follows: S3.1.1 Model Structure. Edge-LSTM adopts a three-layer lightweight structure with a parameter size ≤500K and inference time ≤10ms, adapting to edge computing power. The specific structure is as follows: Input layer: The input dimension is 13, corresponding to thirteen simplified resource status data. The time window length is T=20, that is, the input is simplified data for 20 consecutive sampling periods. The input data format is [batch_size, T, 12], where batch_size is the amount of edge local small batch data.

[0039] Lightweight Hidden Layer: A single-layer LSTM hidden layer with 64 hidden units is used, simplifying the redundant structure of traditional LSTM, eliminating the bias optimization module, and introducing a sparse attention mechanism to focus only on the last 5 key data points within the time window, thus reducing computational cost. The LSTM calculation formula is as follows: i t =σ'(W ii x t +b ii +W hi h t-1 ); f t =σ'(W if x t +b if +W hf h t-1 ); o t =σ'(W io x t +b io +W ho h t-1 ); c t =f t ⊙c t-1 +i t ⊙tanh(W ic x t +b ic ); h t =o t ⊙tanh(c t ); Among them, i t The input gate state controls the flow of current input information into the cell; f t The forget gate state controls the proportion of historical cell states retained; t The output gate state controls the cell state output to a hidden state; c t The cell state is used to store long-term temporal information; h t For the hidden state, output short-term temporal features; σ'(·) is the Sigmoid activation function, with values ​​[0, 1]; ⊙ is the element-wise multiplication operation; W ii b ii W hi Wif b if W hf W io b io W ho W ic b ic These are all trainable weights and bias parameters of the model.

[0040] Output layer: A fully connected output layer is adopted, with an output dimension of 13, corresponding to the prediction levels of thirteen indicators. The activation function is linear activation, and the output is a resource status prediction sequence for the next 5 sampling periods.

[0041] S3.1.2 Model Training Process. Edge-LSTM adopts offline training and edge deployment mode. The training dataset, loss function, training parameters, and process are as follows: 1. Training Dataset Construction. Collect more than 3 months of historical simplified resource status data from large-scale computing network nodes. The sampling frequency is consistent with the edge collection frequency. The training sample format is "historical 20-step simplified data + future 5-step real level", with a total of 1 million training samples, 200,000 validation samples and 100,000 test samples. The data covers different computing network nodes and different operating scenarios.

[0042] 2. Loss Function Design. A mean squared error (MSE) + L2 regularized loss function is used to prevent overfitting of the model. The formula is as follows: ; Where Loss is the total loss of the model; N is the total number of training samples; n is the sample index; and t is the prediction time step. The predicted rank for the nth sample at step t; x n，t Let α be the true rank of the nth sample at step t; α is the regularization coefficient, preset to 10. -4 ; This is the L2 norm of the model weights, used to suppress excessively large weights.

[0043] 3. Training parameter settings. The optimizer used is Adam, with an initial learning rate of 10. -3 The batch size is 32, the number of training rounds is 50, and an early stopping mechanism is introduced, that is, if the validation set loss does not decrease for 5 consecutive rounds, the training is stopped to ensure that the model converges and has strong generalization ability.

[0044] 4. Model Deployment. After training, the model parameters are fixed, the model size is compressed, and it is deployed to various edge computing units. Furthermore, it will be fine-tuned monthly based on newly added data to maintain the model's predictive accuracy.

[0045] S3.1.3 Model Inference. The edge computing unit collects simplified resource status data in real time, inputs it into the Edge-LSTM model, and outputs the predicted level sequence for the next 5 sampling periods. The entire inference process is completed locally at the edge, with an inference time of ≤10ms, demonstrating high real-time performance.

[0046] S3.2 Early Warning Identification. The edge computing unit identifies abnormal resource trends based on the predicted level sequence output by Edge-LSTM, and then classifies them into multiple early warning levels, with the specific rules as follows: Level 1 Warning (Emergency): If any indicator is predicted to be Level 1 (extremely poor) within the next sampling period, an emergency warning is triggered, indicating a risk of resource overload, hardware failure, or service interruption. Key indicators include: processor utilization, transmission latency, packet loss rate, overall node load, hardware temperature, and storage utilization.

[0047] Level 2 Warning (Attention): If any metric is predicted to be Level 2 (poor) within the next 3 sampling periods, and no Level 1 warning is triggered, an attention warning will be issued, indicating a risk of performance degradation or task lag. These metrics include: core load, clock speed utilization, read / write speed, IOPS, and bandwidth utilization.

[0048] Level 3 Warning (Normal): If the prediction level of all thirteen indicators is ≥ Level 3 (medium or above) within the next 5 sampling periods, it indicates that there is no abnormal risk and the resource operation is stable.

[0049] S3.3 Local Pre-Scheduling Scheme Generation. The edge computing unit combines simplified resource status data, task adaptation tags, and alert levels to generate a local pre-scheduling scheme. The scheme includes scheduling operations, operation parameters, and execution timing. The specific generation rules are as follows: 1. Scheduling Operations: Based on the warning type and resource type, standardized scheduling operations are matched, which are divided into five categories: CPU (Central Processing Unit) core allocation, video memory expansion, bandwidth limiting, load balancing, and hardware frequency reduction.

[0050] 2. Operating parameters: Based on the simplified resource level, determine specific parameter values, such as the number of CPUs allocated, bandwidth rate limiting ratio, expansion capacity, etc.

[0051] 3. Execution timing: The execution time is determined according to the warning level. Level 1 warnings are executed immediately, Level 2 warnings are executed after a one-cycle delay, and Level 3 warnings do not require execution.

[0052] S4. Encapsulate the simplified resource status data, task adaptation tags, early warning levels, and local pre-scheduling scheme into structured data, and report it to the central server using an incremental transmission method and a lightweight communication protocol.

[0053] For example, step S4 is specifically as follows: S4.1 Structured Data Encapsulation. The edge computing unit encapsulates simplified resource status data, task adaptation tags, warning levels, and local pre-scheduling schemes into standardized structured data, using key-value pairs. Specific fields are as follows: Node identifier: Used to distinguish different computing network nodes; Streamlined resource status data: a combination of thirteen indicators; Task adaptation tags: Standardized task identifier or no adapted task; Warning levels: Level 1 / Level 2 / Level 3; Local pre-scheduling scheme: scheduling operations, parameters, and execution timing.

[0054] The encapsulated single data entry has a size of ≤1KB, resulting in a very small data volume, making it suitable for low-bandwidth transmission scenarios.

[0055] S4.2 Incremental Transmission Mechanism. Edge computing units employ an incremental transmission mechanism, reporting only changed data and avoiding the repeated transmission of unchanged data. Specific rules are as follows: Data reporting is triggered only when any one of the following changes: streamlined resource status data, task adaptation tags, alert levels, and local pre-scheduling scheme; Unchanged fields are not transmitted repeatedly; only changed fields are transmitted, resulting in a data transmission volume compression of over 90%. If there are no continuous changes, a fallback report will be triggered, with a reporting interval of 30 seconds. Only the heartbeat identifier will be reported to ensure link connectivity.

[0056] S4.3 Lightweight Communication Protocol. Utilizing the CoAP lightweight communication protocol, the message header is ≤4 bytes, supporting low-power, low-bandwidth transmission to adapt to unstable edge network scenarios. This protocol supports QoS classification; Level 1 warning data is transmitted with high priority to ensure real-time performance, while ordinary data is transmitted with low priority to balance bandwidth usage.

[0057] S4.4 Central Reception and Processing. After receiving the structured data reported by the edge, the central server directly integrates the data from all network nodes without the need for secondary calculation, cleaning, adaptation, and prediction. It then generates global scheduling instructions based on global task priorities and load balancing requirements, and sends them to the corresponding edge computing units to achieve efficient collaboration between "edge computing and central scheduling".

[0058] This application also provides a multi-dimensional perception and reporting system for computing network resource status based on edge computing. This system includes lightweight edge computing units deployed at various nodes of the computing network. Each edge computing unit collects operational data from nearby nodes, considering dimensions such as computing power, storage, network, load, and environment. The collected operational data undergoes anomaly removal and standardization processing, followed by resource capacity quantification according to hierarchical rules to obtain simplified resource status data. The edge computing unit retrieves a standard task type library, extracts the resource requirement conditions corresponding to various tasks, and matches the simplified resource status data with the resource requirement conditions of each task one by one to calculate... The corresponding adaptation level is used to generate task adaptation labels based on the ranking results of the adaptation level; the edge computing unit inputs the simplified resource status data into the lightweight time series prediction model and outputs the prediction results of the resource status; based on the prediction results, it identifies abnormal resource trends and classifies warning levels, and generates a local pre-scheduling scheme containing scheduling operations, operation parameters and execution timing by combining the simplified resource status data and task adaptation labels; the edge computing unit encapsulates the simplified resource status data, task adaptation labels, warning levels and local pre-scheduling scheme into structured data, and reports it to the central server through a lightweight communication protocol using incremental transmission.

[0059] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0060] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A multi-dimensional perception and reporting method for network resource status based on edge computing, characterized in that, include: S1. Deploy lightweight edge computing units at each node of the computing network. The edge computing units collect operational data on computing power, storage, network, load, and environment of the nodes nearby. The collected operational data is then subjected to anomaly removal and standardization processing in sequence, and resource capacity is quantified according to hierarchical rules to obtain simplified resource status data. S2. Retrieve the standard task type library, extract the resource requirement conditions corresponding to each type of task, match the simplified resource status data with the resource requirement conditions of each type of task one by one, calculate the corresponding degree of adaptation, and generate task adaptation tags based on the ranking result of the degree of adaptation. S3. Input the simplified resource status data into the lightweight time series prediction model and output the prediction result of the resource status; based on the prediction result, identify abnormal resource trends and classify early warning levels; combine the simplified resource status data and the task adaptation tags to generate a local pre-scheduling scheme that includes scheduling operations, operation parameters and execution timing. S4. Encapsulate the simplified resource status data, the task adaptation tag, the early warning level, and the local pre-scheduling scheme into structured data, and report them to the central server using an incremental transmission method and a lightweight communication protocol.

2. The multi-dimensional perception and reporting method for network resource status based on edge computing according to claim 1, characterized in that, In step S1, the computing power dimension includes processor utilization, core load, cache hit rate and clock frequency utilization metrics; the storage dimension includes remaining capacity, read / write speed, IOPS and storage utilization metrics; the network dimension includes transmission latency, bandwidth utilization and packet loss rate metrics; the load dimension includes overall node load metrics; and the environment dimension includes hardware temperature metrics.

3. The multi-dimensional perception and reporting method for network resource status based on edge computing according to claim 1, characterized in that, In step S1, the anomaly removal is achieved through sliding window statistics, using the following formula: ； ； In the formula, μ is the mean of the data within the sliding window, w is the length of the sliding window, p is the index of the data within the window, and x p Let σ be the p-th original data point within the window, and let σ be the standard deviation of the data.

4. The multi-dimensional perception and reporting method for network resource status based on edge computing according to claim 1, characterized in that, In step S1, the standardization process uses a normalization method, and the formula is: ； In the formula, For the normalized data, This is valid data after anomaly removal. and These are the historical minimum and maximum values ​​of the indicator, respectively.

5. The multi-dimensional perception and reporting method for network resource status based on edge computing according to claim 1, characterized in that, In step S2, the formula for calculating the degree of fit is: ； In the formula, Score k Let λ represent the fit degree for the k-th type of task, n be the number of task requirement indicators, j be the index of the indicator, and λ be the value of the indicator. j As the indicator weight, a j For the normalized index value, r k，j Let f be the requirement condition for the j-th indicator of the k-th task, and f be the matching function.

6. The multi-dimensional perception and reporting method for network resource status based on edge computing according to claim 1, characterized in that, In step S2, the standard task type library includes categories such as AI training, AI inference, general computing, storage-intensive, network-intensive, and edge lightweight tasks.

7. The multi-dimensional perception and reporting method for network resource status based on edge computing according to claim 1, characterized in that, In step S3, the lightweight time series prediction model adopts an improved recurrent neural network, which consists of an input layer, a hidden layer and an output layer.

8. The multi-dimensional perception and reporting method for network resource status based on edge computing according to claim 1, characterized in that, In step S4, the structured data adopts a key-value pair format and includes node identifiers, the simplified resource status data, the task adaptation tags, the warning level, and the local pre-scheduling scheme.

9. The multi-dimensional perception and reporting method for network resource status based on edge computing according to claim 1, characterized in that, In step S4, the incremental transmission only reports the changed data fields; the unchanged fields are not transmitted.

10. A system applying the multi-dimensional perception and reporting method for network resource status based on edge computing as described in any one of claims 1-9, characterized in that, It includes lightweight edge computing units deployed at various nodes of the computing network. The edge computing units collect operational data on computing power, storage, network, load, and environment dimensions of the nodes nearby. The collected operational data is then subjected to anomaly removal and standardization processing in sequence, and resource capacity is quantified according to hierarchical rules to obtain simplified resource status data. The edge computing unit retrieves the standard task type library, extracts the resource requirement conditions corresponding to each type of task, matches the simplified resource status data with the resource requirement conditions of each type of task one by one, calculates the corresponding degree of adaptation, and generates task adaptation tags based on the ranking result of the degree of adaptation. The edge computing unit inputs the simplified resource status data into the lightweight time-series prediction model and outputs the prediction results of the resource status; based on the prediction results, it identifies abnormal resource trends and classifies warning levels, and combines the simplified resource status data and the task adaptation tags to generate a local pre-scheduling scheme that includes scheduling operations, operation parameters and execution timing. The edge computing unit encapsulates the simplified resource status data, the task adaptation tags, the early warning level, and the local pre-scheduling scheme into structured data, and reports them to the central server using an incremental transmission method and a lightweight communication protocol.

Images