Scene-based computing power task arrangement method and device

Abstract

The embodiment of the application provides a kind of scene-based computing power task arrangement method and device, pass through scene classification model inference scene type identification and read multi-dimensional resource weight vector from scene weight mapping table, combine time series prediction model driving dynamic resource view construction and multi-dimensional weighted node score screening obtain candidate node set, by transmission path quality index screening optimal path and task specification drive resource combination matching obtain target node and resource quota after encapsulation arrangement decision instruction issue execution, effectively solve the deficiency of traditional technology in scene perception adaptive resource weight matching, prediction driving dynamic resource view construction and multi-dimensional weighted node screening and path quality optimization and resource specification adaptation collaborative arrangement execution, for the scene adaptive intelligent arrangement and full-link high-quality scheduling execution of computing power task in heterogeneous computing power network provide technical support.

Description

Technical Field

[0001] This application relates to the field of data processing, specifically to a scenario-based computing task orchestration method and apparatus. Background Technology

[0002] Existing computing task orchestration methods have significant shortcomings. Traditional systems perform poorly in parsing computing task requests and acquiring scenario-adaptive resource weights. They typically rely on manually preset static scheduling rules or simple resource utilization thresholds to allocate computing tasks in a coarse-grained manner. They lack the ability to dynamically match scenario-aware adaptive resource weights by inputting task feature data into a scenario classification model to perform classification inference to obtain scenario type identifiers, and then retrieving the corresponding multi-dimensional resource weight vectors from the scenario weight mapping table based on the scenario type identifiers. This results in the inability to accurately identify the differences in resource requirements of computing tasks under different business scenarios, making it difficult to provide a scenario-adaptive weight basis for subsequent differentiated node scoring and resource orchestration decisions.

[0003] Furthermore, existing technologies face bottlenecks in real-time status acquisition of computing nodes, construction of dynamic resource views driven by time-series prediction, and weighted node scoring and selection. Most systems lack the ability to collect computing node status information and network link transmission status information and input them into a time-series prediction model to obtain predicted resource changes; to fuse real-time status data with predicted resource changes to generate a dynamic resource view; and to perform weighted scoring calculations on candidate nodes based on a multi-dimensional resource weight vector and the dynamic resource view to obtain a node score sequence, which is then included in the candidate node set according to selection threshold conditions. This results in node selection relying solely on static snapshot data at the time of acquisition, leading to insufficient timeliness and accuracy in orchestration decisions under scenarios of dynamic fluctuations in node resource status, thus affecting the overall service quality assurance capability of computing task scheduling.

[0004] Existing systems have technical shortcomings in transmission path quality optimization, resource combination matching, and the encapsulation and issuance of complete orchestration decision instructions. They lack a path quality awareness and resource specification adaptation collaborative orchestration execution mechanism. This mechanism is insufficient for calculating the transmission path from the data source to each node in a candidate node set, selecting the optimal path based on path quality indicators, performing resource combination matching on the candidate node set based on task specification identifiers to obtain the target node and resource quota, and encapsulating the target node, resource quota, and optimal path into orchestration decision instructions for issuance and execution. These shortcomings affect the comprehensive orchestration execution capability of computing tasks under the dual constraints of end-to-end transmission link quality assurance and accurate node resource quota matching. Solving these problems is crucial for improving the scenario adaptability and end-to-end collaborative scheduling capability of task orchestration in heterogeneous computing power networks. Summary of the Invention

[0005] To address the problems in existing technologies, this application provides a scenario-based computing task orchestration method and apparatus, which can effectively solve the shortcomings of traditional technologies in scenario-aware adaptive resource weight matching, prediction-driven dynamic resource view construction and multi-dimensional weighted node selection, path quality optimization and resource specification adaptation collaborative orchestration execution, etc., and provide technical support for scenario-adaptive intelligent orchestration and high-quality scheduling execution of computing tasks in heterogeneous computing networks.

[0006] To solve at least one of the above problems, this application provides the following technical solution: Firstly, this application provides a scenario-based computing task orchestration method, including: Receive computing power task requests and parse the request content to obtain task feature data. Input the task feature data into the scene classification model to perform classification inference to obtain scene type identifier. Based on the scene type identifier, read the corresponding multi-dimensional resource weight vector from the scene weight mapping table. Real-time status data is obtained by collecting computing power status information of computing power nodes and transmission status information of network links. The real-time status data is input into a time series prediction model to obtain resource change prediction values. The real-time status data and the resource change prediction values ​​are fused to generate a dynamic resource view. Based on the multi-dimensional resource weight vector and the dynamic resource view, a weighted scoring calculation is performed on the candidate nodes to obtain a node scoring sequence. Nodes in the node scoring sequence that are higher than the screening threshold are included in the candidate node set. For each node in the candidate node set, the transmission path from the data source to the node is calculated, and the optimal path is selected based on the path quality index. Based on the task specification identifier, resource combination matching is performed on the candidate node set to obtain the target node and resource quota. The target node, the resource quota and the optimal path are encapsulated into an orchestration decision instruction and issued for execution.

[0007] Furthermore, it also includes: receiving computing power task requests and parsing request messages to obtain raw request data; performing field integrity verification and format normalization processing on the raw request data to obtain normalized request data; and extracting data type field, computational load estimation field, response time requirement field, data source location field, and task type identifier field from the normalized request data and assembling them into task feature data. The task feature data is numerically normalized to obtain a normalized feature vector. The normalized feature vector is then input into the input layer of the scene classification model and forward propagated through a multi-layer neural network to obtain a scene probability distribution. The scene category with the highest probability value is selected from the scene probability distribution as the scene type identifier.

[0008] Furthermore, it also includes: loading a scene weight mapping table from the memory and establishing an index relationship between scene type identifiers and weight records to obtain a weight index structure, and using the scene type identifier as a query key to perform a matching retrieval in the weight index structure to obtain the target weight record; The computing power weight value, latency weight value, bandwidth weight value, storage weight value, price weight value and energy consumption weight value are read from the target weight record and assembled into a multi-dimensional resource weight vector in a preset dimension order. The multi-dimensional resource weight vector is written into the task context cache for subsequent orchestration calculation and calling.

[0009] Furthermore, it also includes: collecting computing power utilization, available memory capacity and computing power idle rate by probes deployed on each computing power node to obtain computing power status information; collecting end-to-end latency, remaining bandwidth and packet loss rate by network measurement protocol to obtain transmission status information; aligning the computing power status information and the transmission status information by timestamp and aggregating them into real-time status data; organizing the real-time status data into a historical time series according to a preset time window and inputting it into a time series prediction model to perform forward inference to obtain resource change prediction values; The fused status record is obtained by concatenating the corresponding prediction index values ​​in the real-time status data and the corresponding prediction index values ​​in the resource change prediction values ​​at the field level according to the node identifier and the link identifier. The fused status record is then indexed according to the computing power node dimension and the network link dimension and written into the memory data structure to generate a dynamic resource view.

[0010] Furthermore, it also includes: traversing each computing node from the dynamic resource view and reading the corresponding computing power index value, latency index value, bandwidth index value, storage index value and energy consumption index value; performing normalization processing on each index value to obtain a normalized index vector; performing dimension-by-dimensional multiplication on the normalized index vector and summing it to obtain the comprehensive score value of the node; and associating the comprehensive score values ​​of each node with the node identifier to assemble a node score sequence. The comprehensive score of each node in the node scoring sequence is compared with the screening threshold. The identifiers of nodes with comprehensive scores higher than the screening threshold are extracted and written into the candidate node set. The candidate node set is arranged in descending order of comprehensive score for subsequent path optimization and resource combination.

[0011] Furthermore, it also includes: reading the data source location identifier from the task feature data, retrieving the reachable link sequence from the data source location to each node in the candidate node set from the network topology data, performing path traversal calculation based on the reachable link sequence to obtain multiple candidate transmission paths, and reading the latency value, remaining bandwidth value and packet loss rate value of each link in the candidate transmission path from the dynamic resource view and aggregating them into a path quality index. The latency, remaining bandwidth, and packet loss rate values ​​in the path quality indicators are weighted and calculated with the corresponding weight values ​​in the multi-dimensional resource weight vector to obtain a comprehensive path score. The candidate transmission paths are sorted in descending order according to the comprehensive path score, and the path with the highest score is selected as the optimal path for that node. A mapping relationship is established between each node and the corresponding optimal path and written into the path decision cache.

[0012] Furthermore, it also includes: parsing the computational load estimation field and the task execution mode field from the task feature data to obtain the task specification identifier; comparing the task specification identifier with the preset specification threshold condition to determine the task specification category; for large-specification category tasks, selecting nodes with continuous computing power resources from the candidate node set and allocating centralized resource quotas; for small-specification category tasks, selecting nodes with preheating containers from the candidate node set and allocating lightweight resource quotas; determining the selection result as the target node and the allocation result as the resource quota. The optimal path corresponding to the target node is read from the path decision cache. The node identifier of the target node, the quota parameter of the resource quota, and the path identifier of the optimal path are encapsulated into an orchestration decision instruction according to a preset message format. The orchestration decision instruction is sent to the target node through the communication interface to execute the task deployment.

[0013] Secondly, this application provides a scenario-based computing task orchestration device, comprising: The task parsing module is used to receive computing power task requests and parse the request content to obtain task feature data. The task feature data is then input into the scene classification model to perform classification inference to obtain a scene type identifier. Based on the scene type identifier, the corresponding multi-dimensional resource weight vector is read from the scene weight mapping table. The node selection module is used to collect computing power status information and network link transmission status information of computing power nodes to obtain real-time status data, input the real-time status data into the time series prediction model to obtain resource change prediction values, fuse the real-time status data and the resource change prediction values ​​to generate a dynamic resource view, perform weighted scoring calculation on candidate nodes based on the multi-dimensional resource weight vector and the dynamic resource view to obtain a node scoring sequence, and include nodes in the node scoring sequence that are higher than the screening threshold into the candidate node set. The computing power orchestration module is used to calculate the transmission path from the data source to each node in the candidate node set and select the optimal path based on the path quality index. Based on the task specification identifier, it performs resource combination matching on the candidate node set to obtain the target node and resource quota. It encapsulates the target node, the resource quota and the optimal path into an orchestration decision instruction and issues it for execution.

[0014] Thirdly, this application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the scenario-based computing task orchestration method.

[0015] Fourthly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the scenario-based computing task orchestration method described above.

[0016] Fifthly, this application provides a computer program product, including a computer program / instructions, which, when executed by a processor, implement the steps of the scenario-based computing task orchestration method.

[0017] As can be seen from the above technical solution, this application provides a scenario-based computing task orchestration method and apparatus. It infers the scenario type identifier through a scenario classification model and reads the multi-dimensional resource weight vector from the scenario weight mapping table. It combines a time-series prediction model to drive the construction of a dynamic resource view and a multi-dimensional weighted node scoring and filtering to obtain a candidate node set. It selects the optimal path through transmission path quality indicators and drives resource combination matching with task specifications to obtain the target node and resource quota. Then, it encapsulates the orchestration decision instruction and issues it for execution. It effectively solves the shortcomings of traditional technologies in scenario-aware adaptive resource weight matching, prediction-driven dynamic resource view construction and multi-dimensional weighted node filtering, and path quality optimization and resource specification adaptation collaborative orchestration execution. It provides technical support for scenario-adaptive intelligent orchestration and high-quality end-to-end scheduling execution of computing tasks in heterogeneous computing networks. Attached Figure Description

[0018] 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 some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart illustrating the scenario-based computing task orchestration method in the embodiments of this application; Figure 2 This is a structural diagram of the scenario-based computing task orchestration device in the embodiments of this application. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, 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, 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.

[0021] The acquisition, storage, use, and processing of data in this application comply with relevant laws and regulations.

[0022] In view of the problems existing in the prior art, this application provides a scenario-based computing task orchestration method and apparatus. It infers the scenario type identifier through a scenario classification model and reads the multi-dimensional resource weight vector from the scenario weight mapping table. Combined with a time-series prediction model to drive the construction of a dynamic resource view and the selection of candidate nodes by multi-dimensional weighted node scoring, the optimal path is selected by transmission path quality index and the resource combination matching driven by task specifications is obtained to obtain the target node and resource quota. Then, the orchestration decision instruction is encapsulated and issued for execution. It effectively solves the shortcomings of traditional technologies in scenario-aware adaptive resource weight matching, prediction-driven dynamic resource view construction and multi-dimensional weighted node selection, path quality optimization and resource specification adaptation collaborative orchestration execution, etc. It provides technical guarantee for scenario-adaptive intelligent orchestration and high-quality scheduling execution of computing tasks in heterogeneous computing power networks.

[0023] To effectively address the shortcomings of traditional technologies in areas such as scene-aware adaptive resource weight matching, prediction-driven dynamic resource view construction and multi-dimensional weighted node selection, path quality optimization, and resource specification adaptation for collaborative orchestration and execution, and to provide technical support for scene-adaptive intelligent orchestration and high-quality end-to-end scheduling and execution of computing tasks in heterogeneous computing networks, this application provides an embodiment of a scene-based computing task orchestration method, see [link to embodiment]. Figure 1 The scenario-based computing task orchestration method specifically includes the following: Step S101: Receive computing power task request and parse the request content to obtain task feature data. Input the task feature data into the scene classification model to perform classification inference to obtain scene type identifier. Read the corresponding multi-dimensional resource weight vector from the scene weight mapping table according to the scene type identifier. In this embodiment, the communication interface unit receives externally submitted computing task requests and obtains the original request messages. The communication interface unit is deployed in the access layer of the computing task orchestration device, acquiring all inbound requests via a high-speed network interface. Upon receipt, an arrival time identifier is appended to each original request message, and the original request message is temporarily stored as a byte sequence in a request buffer queue, awaiting field extraction operations. If the message length of the original request message exceeds a preset capacity threshold, pruning processing is performed and a pruning bit flag is recorded.

[0024] After the original request message is temporarily stored, the data type field, computational load estimate field, response time requirement field, data source location field, and task type identifier field are extracted and encapsulated in a fixed order to form task feature data. Each field is located and read based on the offset in the message header. The data type field identifies the data format processed by the task; the computational load estimate field identifies the computing power required by the task; the response time requirement field identifies the latency constraints of the task; the data source location field identifies the network location of the task's input data; and the task type identifier field identifies the service category explicitly declared by the user. If a field is missing, it is registered as a placeholder and a missing bitmap is marked. This task feature data serves as a structured description of the computing power task request for subsequent normalization processing.

[0025] Accordingly, the task feature data is numerically normalized to generate a normalized feature vector. The normalization process performs a range mapping on each field value, converting the original numerical values ​​to a uniform dimension range. The normalized feature vector, in a compact numerical form, replaces the original task feature data in classification inference. The normalized feature vector is written to a feature cache for forward propagation by the scene classification model.

[0026] After the normalized feature vector is generated, it is input into the scene classification model and forward propagated to obtain the scene probability distribution. The scene classification model adopts a multi-layer neural network structure. The number of nodes in the input layer is consistent with the dimension of the normalized feature vector. The hidden layer performs nonlinear transformation and feature aggregation on the input vector, and the number of nodes in the output layer is consistent with the number of preset scene types. The output layer uses a normalized exponential function to convert the output values ​​of each node into probability values. The scene category with the highest probability value is selected from the scene probability distribution as the scene type identifier. The scene type identifier is written into the classification result cache for weight retrieval operations.

[0027] After the scene type identifier is generated, it is used as the query key to retrieve the corresponding target weight record from the scene weight mapping table. The scene weight mapping table is pre-stored in memory, with the scene type identifier as the primary key to establish an index relationship. If a match is found, the target weight record is read; otherwise, the default weight record is returned and a miss flag is registered. Computing power weight value, latency weight value, bandwidth weight value, storage weight value, price weight value, and energy consumption weight value are extracted from the target weight record and assembled into a multi-dimensional resource weight vector according to a preset dimension order. The multi-dimensional resource weight vector defines the order of importance of various resource indicators under the current scene type.

[0028] Based on the aforementioned assembly results, the multidimensional resource weight vector is associated with the scene type identifier, the task feature data, and the arrival time identifier, and written into the task context cache to be aggregated into task context records. The task context cache organizes all context records of tasks to be orchestrated in ascending order of time, using the arrival time identifier as the sorting key. Each task context record carries four attribute fields: scene type identifier, task feature data, multidimensional resource weight vector, and arrival time identifier. The task context record serves as the output of step S101, and is used in step S102 to read the task feature data at the status acquisition entry point for data source location. Step S102: Collect computing power status information of computing power nodes and transmission status information of network links to obtain real-time status data. Input the real-time status data into a time series prediction model to obtain resource change prediction values. Fuse the real-time status data and the resource change prediction values ​​to generate a dynamic resource view. Perform weighted scoring calculation on candidate nodes based on the multi-dimensional resource weight vector and the dynamic resource view to obtain a node scoring sequence. Include nodes in the node scoring sequence that are higher than the screening threshold into the candidate node set. This embodiment collects computing power status information by deploying probes on each computing power node to monitor node operation. The probes read three indicators at a preset collection period: computing power utilization, available memory capacity, and computing power idle rate. During collection, each piece of computing power status information is appended with a collection time identifier and a node identifier, and the information is temporarily stored in a structured record in a status buffer queue, awaiting retrieval during status aggregation. If a node does not respond within the collection period, it is registered with a missing placeholder and marked as offline.

[0029] After the computing power status information is temporarily stored, network links are probed using a network measurement protocol to obtain transmission status information. The probe process reads three metrics for each link: end-to-end latency, remaining bandwidth, and packet loss rate. A probe time identifier and a link identifier are appended to each transmission status information entry. The transmission status information and the computing power status information are aligned according to the time identifier, and records within the same acquisition period are aggregated into real-time status data. If there is a time identifier discrepancy, a backfilling strategy is used to correct it to the most recent acquisition period. The real-time status data is written to a status aggregation cache for use by the time-series prediction model.

[0030] Accordingly, the real-time status data is organized into a historical time series according to a preset time window, and input into a time series prediction model to perform forward inference to obtain resource change prediction values. The time series prediction model adopts a long short-term memory network structure, with the input sequence length covering a preset number of historical acquisition periods. The hidden layer performs state memory and forgetting gating on the time series features, and the output layer generates prediction index values ​​for a preset number of future acquisition periods. The resource change prediction values ​​include the computing power load prediction curves of each node and the congestion probability prediction values ​​of each link. The resource change prediction values ​​are written into the prediction result cache for dynamic resource view generation operations to read.

[0031] After the resource change prediction value is generated, it is concatenated with the real-time status data at the field level according to node identifier and link identifier to obtain the fused status record. The fused status record contains two types of fields: real-time indicator value and predicted indicator value. The fused status record is indexed according to the computing power node dimension and the network link dimension respectively, and written into an in-memory data structure to generate a dynamic resource view. The dynamic resource view uses the node identifier and link identifier as the primary key to carry all fused status records, which are read by the weighted scoring calculation operation.

[0032] After the dynamic resource view is generated, a multi-dimensional resource weight vector is read from the weight vector entry in the task context cache, and a weighted score calculation is performed on the candidate nodes. Each computing node in the dynamic resource view is traversed, and its corresponding computing power, latency, bandwidth, storage, and energy consumption metrics are read. Each metric value is normalized to obtain a normalized metric vector. The normalized metric vector is multiplied dimension-by-dimensionally by the multi-dimensional resource weight vector and summed to obtain the node's comprehensive score. The comprehensive scores of each node are associated and assembled into a node score sequence according to the node identifier. The node score sequence is written to the score result cache for use in the filtering operation.

[0033] Based on the aforementioned scoring results, the comprehensive score value of each node in the node scoring sequence is compared and determined with the screening threshold. Nodes with comprehensive scores higher than the screening threshold are extracted and written into a candidate node set. The candidate node set is arranged in descending order of comprehensive score value. The candidate node set carries three attribute fields: node identifier, comprehensive score value, and normalized index vector. The candidate node set serves as the output of step S102, and is read by step S103 at the candidate node entry point for transmission path calculation and resource combination matching.

[0034] Step S103: Calculate the transmission path from the data source to each node in the candidate node set and select the optimal path based on the path quality index. Perform resource combination matching on the candidate node set according to the task specification identifier to obtain the target node and resource quota. Encapsulate the target node, the resource quota and the optimal path into an orchestration decision instruction and issue it for execution.

[0035] This embodiment reads task feature data from the status acquisition entry point of the task context cache and extracts the data source location field as the starting point identifier for path calculation. For each node in the candidate node set, a sequence of reachable links from the data source location to that node is retrieved from the network topology data. The network topology data is pre-stored in memory, and all connection relationships are organized using node identifiers and link identifiers as index keys. Path traversal calculation is performed based on the reachable link sequence to obtain multiple candidate transmission paths. If a node has no reachable links, the node is marked as unreachable and removed from the candidate node set.

[0036] After the candidate transmission paths are generated, the latency, remaining bandwidth, and packet loss rate of each link in the path are read from the dynamic resource view and aggregated into a path quality index. The latency value is the sum of the latencies of all links in the path, the remaining bandwidth value is the minimum of the remaining bandwidth of all links in the path, and the packet loss rate value is the product of the packet loss rates of all links in the path. The path quality index is written to the path evaluation cache for use in the calculation of the overall path score.

[0037] Accordingly, the latency, remaining bandwidth, and packet loss rate values ​​in the path quality indicators are weighted and calculated with their corresponding weight values ​​in the multi-dimensional resource weight vector to obtain a comprehensive path score. The reciprocal of the latency value is multiplied by the latency weight value, the remaining bandwidth value is multiplied by the bandwidth weight value, and the reciprocal of the packet loss rate value is multiplied by the reliability weight value. The sum of these three products forms the comprehensive path score. Candidate transmission paths are sorted in descending order according to their comprehensive path scores, and the path with the highest score is selected as the optimal path for that node. A mapping relationship is established between each node and its corresponding optimal path, and this mapping is written to the path decision cache. The path decision cache is used for resource combination and matching operations.

[0038] After the optimal path is determined, the estimated computational load and task execution mode fields are parsed from the task feature data to obtain the task specification identifier. The task specification identifier is then compared with a preset specification threshold to determine the task specification category. If the estimated computational load exceeds the specification threshold or the task execution mode field indicates a long-running mode, the task is classified as a large-specification task. If the estimated computational load is below the specification threshold and the task execution mode field indicates a short-running batch mode, the task is classified as a small-specification task. The task specification category is written to a specification determination cache for read by resource quota allocation operations.

[0039] After determining the task specification category, nodes with continuous computing power resources are selected from the candidate node set for large-specification tasks. The computing power idle rate and resource continuity flag of each candidate node are read from the dynamic resource view, and nodes with a computing power idle rate higher than a preset capacity threshold and a true resource continuity flag are selected as target nodes. A centralized resource quota is allocated to the target node, which includes three parameters: the number of continuous computing power units, memory capacity, and resource lockout time window.

[0040] For small-scale tasks, nodes with preheated containers are selected from the candidate node set. The container preheating status flags of each candidate node are read from the dynamic resource view, and the node with a true container preheating status flag and the highest overall score is selected as the target node. A lightweight resource quota is allocated to the target node, which includes three parameters: container instance identifier, number of computing units, and memory capacity. If no node has a preheated container, the node with the highest overall score is selected, and the container creation process is triggered.

[0041] Based on the aforementioned matching results, the optimal path corresponding to the target node is read from the path decision cache. The node identifier of the target node, the quota parameters of the resource quota, and the path identifier of the optimal path are encapsulated into an orchestration decision instruction according to a preset message format. The orchestration decision instruction carries four attribute fields: target node identifier, resource quota parameters, path identifier, and task characteristic data. The orchestration decision instruction is sent to the target node through the communication interface unit to execute task deployment. The orchestration decision instruction, as the output of step S103, is read by the target node's task execution module at the instruction receiving entry point for resource allocation and task initiation.

[0042] As can be seen from the above description, the scenario-based computing task orchestration method provided in this application can infer the scenario type identifier through a scenario classification model and read the multi-dimensional resource weight vector from the scenario weight mapping table. It combines a time-series prediction model to drive the construction of a dynamic resource view and a multi-dimensional weighted node scoring to obtain a candidate node set. After selecting the optimal path through transmission path quality indicators and matching the resource combination driven by task specifications to obtain the target node and resource quota, it encapsulates the orchestration decision instruction and issues it for execution. This effectively solves the shortcomings of traditional technologies in scenario-aware adaptive resource weight matching, prediction-driven dynamic resource view construction and multi-dimensional weighted node selection, as well as path quality optimization and resource specification adaptation collaborative orchestration execution. It provides technical support for scenario-adaptive intelligent orchestration and high-quality end-to-end scheduling execution of computing tasks in heterogeneous computing power networks.

[0043] In one embodiment of the scenario-based computing task orchestration method of this application, it may further include the following: Step S201: Receive computing power task request and parse the request message to obtain original request data. Perform field integrity verification and format normalization processing on the original request data to obtain normalized request data. Extract data type field, computational load estimation field, response time requirement field, data source location field and task type identifier field from the normalized request data and assemble them into task feature data. Step S202: Perform numerical normalization on the task feature data to obtain a normalized feature vector. Input the normalized feature vector into the input layer of the scene classification model and perform forward propagation through a multi-layer neural network to obtain the scene probability distribution. Select the scene category with the highest probability value from the scene probability distribution as the scene type identifier.

[0044] In this embodiment, the communication interface unit receives externally submitted computing task requests and obtains the original request messages. The communication interface unit listens to the inbound data stream via a high-speed network interface and writes the received request messages into the receive buffer queue in arrival order. An arrival time identifier and message sequence number are appended to each original request message. The original request messages are temporarily stored in the receive buffer queue as a byte sequence, awaiting retrieval during message parsing. If the length of the original request message exceeds a preset capacity threshold, truncation is performed and a truncation bit is registered.

[0045] After the original request message is temporarily stored, it is parsed to obtain the original request data. The parsing process identifies the header and body boundaries according to a preset message protocol format, and extracts key-value pairs of fields from the body. The original request data carries all parsed fields in a structured record format. The original request data is written to a parsing result cache for use in field integrity verification operations.

[0046] Accordingly, field integrity checks and format normalization are performed on the original request data to obtain normalized request data. Field integrity checks iterate through each field in the original request data, determining the existence of required fields based on a preset field list. If a required field is missing, it is filled with a default placeholder and a missing bitmap is added. Format normalization performs type conversion and format unification on each field value, converting string-type numeric values ​​to numeric types and standardizing the time format to a standard timestamp format. The normalized request data is written to a normalization cache for use in field extraction operations.

[0047] After the normalized request data is generated, the following fields are extracted: data type field, computational load estimate field, response time requirement field, data source location field, and task type identifier field. The data type field identifies the data format processed by the task, with values ​​including image, video, text, and numerical types. The computational load estimate field identifies the computing power required by the task, measured in floating-point operations. The response time requirement field identifies the latency constraints of the task, measured in milliseconds. The data source location field identifies the network address of the task input data. The task type identifier field identifies the service category explicitly declared by the user. The extracted five fields are assembled into task feature data in a fixed order. This task feature data serves as the output of step S201 and is read by step S202 at the feature input entry for normalization processing.

[0048] After the task feature data is assembled, it undergoes numerical normalization to obtain a normalized feature vector. The normalization process maps the values ​​of each field to a range, linearly scaling the original values ​​to the zero-to-one interval according to a preset field value range. Data type fields are converted to multi-dimensional binary vectors using one-hot encoding and then concatenated into the normalized feature vector. The normalized feature vector carries a standardized representation of the task features in a compact numerical form. This normalized feature vector is written to a feature vector cache for use by the scene classification model.

[0049] Accordingly, the normalized feature vector is input into the input layer of the scene classification model, and the scene probability distribution is obtained by forward propagation through a multi-layer neural network. The scene classification model adopts a multi-layer perceptron structure. The number of nodes in the input layer is consistent with the dimension of the normalized feature vector. The hidden layer contains two fully connected layers and uses a non-linear activation function to perform feature transformation. The number of nodes in the output layer is consistent with the number of preset scene types. The output layer uses a normalized exponential function to convert the output values ​​of each node into probability values, and the sum of all probability values ​​is one. The scene probability distribution carries the confidence of each scene category in vector form.

[0050] After the scenario probability distribution is generated, the probability values ​​of each scenario category are traversed and a maximum value search is performed. The scenario category with the highest probability value is selected as the scenario type identifier. If multiple scenario categories have equal and all of the highest probability values, the category with the highest probability value is selected according to the preset scenario priority order. The scenario type identifier identifies the business scenario to which the current task belongs in the form of an enumeration value. The scenario type identifier is the output of step S202 and is used by subsequent steps to read from the scenario identifier entry point for weight vector retrieval and orchestration strategy generation.

[0051] In one embodiment of the scenario-based computing task orchestration method of this application, it may further include the following: Step S301: Load the scene weight mapping table from the memory and establish an index relationship between scene type identifier and weight record to obtain a weight index structure. Use the scene type identifier as the query key to perform a matching retrieval in the weight index structure to obtain the target weight record. Step S302: Read the computing power weight value, latency weight value, bandwidth weight value, storage weight value, price weight value and energy consumption weight value from the target weight record and assemble them into a multi-dimensional resource weight vector in a preset dimension order. Write the multi-dimensional resource weight vector into the task context cache for subsequent orchestration calculation and calling.

[0052] This embodiment loads a scene weight mapping table from the non-volatile storage medium of the memory to obtain a weight data set. The scene weight mapping table is pre-stored in the memory as a persistent file, containing mapping records of all preset scene types and their corresponding weight parameters. During loading, the weight data set is read into the working area of ​​the volatile storage medium, and a loading time identifier is attached for version management. If file corruption or verification failure is detected during loading, the data is reloaded from the backup storage path, and an anomaly recovery flag is registered.

[0053] After the weighted data set is loaded, each weight record is traversed, and an index relationship between scene type identifiers and weight records is established to obtain a weighted index structure. The weighted index structure is organized in hash table form, using the scene type identifier as the index key and the corresponding weight record as the index value. During index creation, a hash calculation is performed on each scene type identifier to obtain a storage slot, and the weight record is written to the corresponding slot. If a hash collision exists, the colliding record is appended to the tail of the linked list in the same slot using a chaining method. The weighted index structure is written to the index cache for reading during matching and retrieval operations.

[0054] Accordingly, the scene type identifier is read from the scene identifier entry in step S202, and used as the query key to perform a matching retrieval in the weight index structure to obtain the target weight record. The matching retrieval performs a hash calculation on the query key to locate the storage slot, traverses the record list of that slot, and compares it with the scene type identifier field. If a match is successful, the corresponding weight record is returned as the target weight record; if no match is found after traversal, a default weight record is returned and a miss flag is registered. The target weight record is written to the retrieval result cache for use in weight field extraction operations.

[0055] After the target weight record retrieval is completed, six weight fields are read: computing power weight value, latency weight value, bandwidth weight value, storage weight value, price weight value, and energy consumption weight value. The computing power weight value defines the importance of computing power resources in orchestration decisions; the latency weight value defines the importance of network latency; the bandwidth weight value defines the importance of remaining bandwidth; the storage weight value defines the importance of storage read / write performance; the price weight value defines the importance of computing power price; and the energy consumption weight value defines the importance of node energy consumption. The sum of the six weight values ​​is one, and each weight value ranges from zero to one.

[0056] Accordingly, the extracted five weight values ​​are assembled into a multi-dimensional resource weight vector according to a preset dimensional order. The preset dimensional order is as follows: computing power, latency, bandwidth, storage, and energy consumption. During assembly, each weight value is filled into the corresponding position of the vector according to the dimensional order, forming a fixed-length numerical vector. The multi-dimensional resource weight vector carries the importance ranking of each resource indicator under the current scenario type in a compact numerical form. The multi-dimensional resource weight vector is written to a weight vector cache for reading by task context association operations.

[0057] After the multidimensional resource weight vector is assembled, it is associated with the scene type identifier, task feature data, and arrival time identifier, and written into the task context cache to be aggregated into task context records. The task context cache organizes all context records of tasks to be orchestrated in ascending order of time, using the arrival time identifier as the sorting key. During concurrent writing, a stable sort is performed by the arrival time identifier, and records with the same time identifier are arranged in order of message sequence number. The task context record carries four attribute fields: scene type identifier, task feature data, multidimensional resource weight vector, and arrival time identifier. The task context record serves as the output of step S302, and is used by subsequent orchestration calculation steps to read the multidimensional resource weight vector at the weight vector entry point for weighted scoring calculation of candidate nodes.

[0058] In one embodiment of the scenario-based computing task orchestration method of this application, it may further include the following: Step S401: Obtain computing power status information by collecting computing power utilization, available memory capacity and computing power idle rate through probes deployed on each computing power node; obtain transmission status information by collecting end-to-end latency, remaining bandwidth and packet loss rate through network measurement protocol; align the computing power status information and the transmission status information by timestamp and aggregate them into real-time status data; organize the real-time status data into a historical time series according to a preset time window and input it into the time series prediction model to perform forward inference to obtain resource change prediction values; Step S402: Perform field-level concatenation of each indicator value in the real-time status data and the corresponding predicted indicator value in the resource change prediction value according to the node identifier and link identifier to obtain the fused status record. Build indexes on the fused status record according to the computing power node dimension and the network link dimension respectively, and write them into the memory data structure to generate a dynamic resource view.

[0059] This embodiment collects node operating status data using probes deployed on each computing node to obtain computing power status information. The probes poll the node's system interface at a preset collection period, reading three indicators: computing power utilization, available memory capacity, and computing power idle rate. Computing power utilization indicates the current proportion of computing power resources being used; available memory capacity indicates the remaining allocable memory space; and computing power idle rate indicates the proportion of idle capacity for each type of computing power unit. During collection, a collection timestamp and node identifier are appended to each piece of computing power status information, and this information is temporarily stored in a structured record in a computing power status buffer queue, awaiting read during status alignment operations. If a node does not respond within the collection period, a missing placeholder is used to register the node and mark it as offline.

[0060] After the computing power status information is temporarily stored, the network links are probed using a network measurement protocol to obtain transmission status information. The probe process involves sending measurement messages to each link endpoint and receiving responses, calculating three metrics: end-to-end latency, remaining bandwidth, and packet loss rate. End-to-end latency indicates the transmission time of a message from the source to the destination; remaining bandwidth indicates the maximum transmission rate the link can currently support; and packet loss rate indicates the proportion of messages lost during transmission. A probe timestamp and link identifier are appended to each transmission status message. This transmission status information is written to a transmission status buffer queue for read by status alignment operations.

[0061] Accordingly, the computing power status information and the transmission status information are aligned according to timestamps to obtain a time-aligned status set. The alignment operation uses a preset collection period as the time window, grouping computing power status records and transmission status records within the same window into the same time slot. If the timestamp of a record deviates from the current window boundary, it is backfilled into an adjacent window according to the nearest neighbor principle, and a backfilling mark is registered. The aligned computing power status information and transmission status information are aggregated into real-time status data according to time slots. The real-time status data carries all node indicators and link indicators at the same moment in the form of composite records. The real-time status data is written to a status aggregation cache for reading by time-series organization operations.

[0062] After the real-time status data is aggregated, it is organized into a historical time series according to a preset time window. The time window covers a preset number of continuous acquisition periods, and the real-time status data of each period is arranged in chronological order to form a sequence structure. The historical time series is updated in a sliding window manner, with newly arriving real-time status data appended to the end of the sequence and historical data exceeding the window range removed from the beginning of the sequence. The historical time series is written into a time series buffer for the time series prediction model to read.

[0063] Accordingly, the historical time series is input into the time series prediction model for forward inference to obtain predicted resource changes. The time series prediction model employs a Long Short-Term Memory (LSTM) network structure. The input layer receives a vectorized representation of the historical time series, the hidden layer performs state memory and selective forgetting of the time series features through input gates, forget gates, and output gates, and the output layer generates predicted index values ​​for a predetermined number of future data collection periods. The predicted resource changes include the predicted computing load curves for each node and the predicted congestion probability for each link. These predicted resource changes are written to a prediction result cache for retrieval by field-level concatenation operations.

[0064] After the resource change prediction values ​​are generated, the indicator values ​​in the real-time status data are concatenated with the corresponding predicted indicator values ​​in the resource change prediction values ​​at the field level, based on node identifiers and link identifiers, to obtain a fused status record. The concatenation operation uses the node identifier as the association key to merge the real-time computing power indicator values ​​and predicted computing power indicator values ​​of the same node into a node fusion field. Similarly, it uses the link identifier as the association key to merge the real-time transmission indicator values ​​and predicted transmission indicator values ​​of the same link into a link fusion field. The fused status record carries the parallel fields of real-time and predicted values ​​in the form of an extended record.

[0065] Based on the aforementioned concatenation results, the fusion status records are indexed separately according to the computing power node dimension and the network link dimension, and written into an in-memory data structure to generate a dynamic resource view. The node-dimensional index uses the node identifier as the primary key, organizing the fusion status records of each node into a node status table. The link-dimensional index uses the link identifier as the primary key, organizing the fusion status records of each link into a link status table. The dynamic resource view uses a dual-index structure to carry the global resource status, supporting fast retrieval by node identifier or link identifier. This dynamic resource view, as the output of step S402, is available for subsequent steps to read at the resource view entry point for candidate node screening and path quality evaluation.

[0066] In one embodiment of the scenario-based computing task orchestration method of this application, it may further include the following: Step S501: Traverse each computing node from the dynamic resource view and read the corresponding computing power index value, latency index value, bandwidth index value, storage index value and energy consumption index value. Perform normalization processing on each index value to obtain a normalized index vector. Perform dimension-wise multiplication of the normalized index vector and multidimensional resource weight vector and sum them to obtain the comprehensive score value of the node. Associate the comprehensive score values ​​of each node with the node identifier and assemble them into a node score sequence. Step S502: Compare and determine the comprehensive score of each node in the node scoring sequence with the screening threshold condition, extract the node identifiers of the nodes with comprehensive scores higher than the screening threshold condition and write them into the candidate node set, and arrange the candidate node set in descending order of comprehensive score for subsequent path optimization and resource combination.

[0067] This embodiment iterates through the node status table of the dynamic resource view to read the corresponding fusion status record of each computing node. For each computing node, five resource indicators are extracted from the fusion status record: computing power indicator, latency indicator, bandwidth indicator, storage indicator, and energy consumption indicator. The computing power indicator is the weighted average of the real-time value and the predicted value of the computing power idle rate; the latency indicator is the average end-to-end latency of the links associated with the node; the bandwidth indicator is the minimum remaining bandwidth of the links associated with the node; the storage indicator is the normalized representation of the available memory capacity; and the energy consumption indicator is the ratio of the node's current power consumption to its rated power consumption. These five resource indicators are written to the node indicator cache for normalization processing operations to read.

[0068] After the resource indicators are extracted, normalization is performed on each indicator value to obtain a normalized indicator vector. The normalization process linearly scales the original values ​​according to a preset range of indicator values. Computing power and bandwidth indicators are forward normalized; the higher the value, the closer the normalization result is to one. Latency, energy consumption, and packet loss-related indicators are inversely normalized; the lower the value, the closer the normalization result is to one. Storage indicators are proportionally normalized based on the maximum memory capacity. The five normalized indicator values ​​are then assembled into a normalized indicator vector in a preset dimensional order. This normalized indicator vector is written to a normalization cache for use in weighted scoring calculations.

[0069] Accordingly, a multi-dimensional resource weight vector is read from the weight vector entry point of the task context cache. The normalized index vector is multiplied dimension-by-dimensionally by the multi-dimensional resource weight vector and then summed to obtain the comprehensive score value of the node. The dimension-by-dimensional multiplication involves multiplying the first dimension of the normalized index vector by the computing power weight value of the multi-dimensional resource weight vector, the second dimension by the latency weight value, the third dimension by the bandwidth weight value, the fourth dimension by the storage weight value, and the fifth dimension by the energy consumption weight value. The summation operation adds up these five multiplications to form the comprehensive score value. The comprehensive score value, expressed as a floating-point number, indicates the degree of adaptability of the node to the current task.

[0070] After the comprehensive score is calculated, the comprehensive score values ​​of each node are associated and assembled into a node score sequence according to the node identifier. The assembly operation traverses all calculated nodes, encapsulates the node identifier and the corresponding comprehensive score value into a score record, and appends it to the end of the node score sequence in traversal order. The node score sequence carries the score records of all nodes in the form of an ordered list. The node score sequence is written to the score result cache for reading by the filtering and judgment operation.

[0071] After the node scoring sequence is assembled, the comprehensive score value of each node is traversed and compared with the screening threshold. The screening threshold is read from the system configuration parameters and identifies the minimum score requirement for candidate node admission. The comparison is performed by numerically comparing the comprehensive score value of each scoring record with the screening threshold. If the comprehensive score value is greater than or equal to the screening threshold, the node is determined to pass the screening; if the comprehensive score value is less than the screening threshold, the node is determined to fail the screening and is filtered out.

[0072] Based on the aforementioned judgment results, the identifiers of nodes with comprehensive scores higher than the filtering threshold are extracted and written into the candidate node set. The extraction operation iterates through the filtered score records, reads the node identifier field, and appends it to the candidate node set. The candidate node set is sorted in descending order of comprehensive score, with the node with the highest score at the beginning of the set. The candidate node set carries three attribute fields: node identifier, comprehensive score, and normalized index vector. The candidate node set serves as the output of step S502, and is used by subsequent steps to read at the candidate node entry point for transmission path optimization and resource combination matching.

[0073] In one embodiment of the scenario-based computing task orchestration method of this application, it may further include the following: Step S601: Read the data source location identifier from the task feature data, retrieve the reachable link sequence from the network topology data for each node in the candidate node set, perform path traversal calculation based on the reachable link sequence to obtain multiple candidate transmission paths, read the latency value, remaining bandwidth value and packet loss rate value of each link in the candidate transmission path from the dynamic resource view and aggregate them into path quality indicators. Step S602: Perform weighted calculations on the latency value, remaining bandwidth value, and packet loss rate value in the path quality indicators with the corresponding weight values ​​in the multi-dimensional resource weight vector to obtain a comprehensive path score. Sort each candidate transmission path in descending order according to the comprehensive path score and select the path with the highest score as the optimal path for that node. Establish a mapping relationship between each node and the corresponding optimal path and write it into the path decision cache.

[0074] This embodiment reads task feature data from the task context cache and extracts the data source location identifier as the starting point for path calculation. The data source location identifier identifies the storage node or access gateway of the task input data in the form of a network address. The data source location identifier is written into the path calculation context for read by reachable link retrieval operations. If the data source location identifier field is missing, an alternative address is read from the default data source configuration and a missing backfill flag is registered.

[0075] After the data source location identifier is extracted, for each node in the candidate node set, a sequence of reachable links from the data source location to that node is retrieved from the network topology data. The network topology data is pre-stored in memory, organizing the connections and link attributes between all nodes in the form of an adjacency list. The retrieval operation starts with the data source location identifier and ends with the current candidate node identifier, traversing the link records in the adjacency list and filtering for links with connectivity. The sequence of reachable links carries all available link segments from the start point to the end point in an ordered list. If a candidate node has no reachable links, the node is marked as unreachable and removed from the candidate node set.

[0076] Accordingly, path traversal calculations are performed based on the reachable link sequence to obtain multiple candidate transmission paths. The path traversal employs a breadth-first search strategy, expanding reachable nodes layer by layer from the data source location, recording the path prefixes to each intermediate node. When the search expands to the target candidate node, the path prefixes are backtracked to form a complete transmission path. During the traversal, visited nodes are marked with access markers to avoid loops. Since multiple different paths may exist for the same candidate node, all valid paths are collected to form a candidate transmission path set. This candidate transmission path set is written to a path traversal cache for reading during path quality metric aggregation operations.

[0077] After the candidate transmission path set is generated, the latency, remaining bandwidth, and packet loss rate of each link in the dynamic resource view are read from the link status table and aggregated into a path quality index. For each candidate transmission path, all its link segments are traversed and the corresponding fusion status record is read. The latency value is the sum of the latency of all links in the path, representing the total time it takes for a packet to be transmitted through that path. The remaining bandwidth value is the minimum remaining bandwidth of all links in the path, representing the bottleneck transmission capacity of that path. The packet loss rate value is the product of the transmission success rates of all links in the path and the complement, representing the probability of a packet being lost through that path. The path quality index is encapsulated in the form of triples and written to the path evaluation cache for weighted scoring calculation.

[0078] After the path quality indicators are aggregated, a multi-dimensional resource weight vector is read from the weight vector entry point in the task context cache. The path quality indicator values ​​are then weighted and calculated with their corresponding weight values ​​to obtain a comprehensive path score. The latency value is multiplied by its reciprocal and then by the latency weight value; a smaller value results in a higher weighted score. The remaining bandwidth value is directly multiplied by the bandwidth weight value; a larger value results in a higher weighted score. The packet loss rate value is multiplied by its complement and then by the reliability coefficient; a lower packet loss rate results in a higher weighted score. These three weighted results are summed to form the comprehensive path score. The comprehensive path score, expressed as a floating-point number, indicates the degree of adaptation of the path to the current task's transmission.

[0079] Accordingly, the candidate transmission paths are sorted in descending order according to their comprehensive path scores, and the path with the highest score is selected as the optimal path for that node. The sorting operation sorts all candidate transmission paths associated with the same candidate node in descending order based on their comprehensive path scores, with the highest-scoring path at the beginning of the sequence. The selection operation reads the first path of the sorted sequence as the optimal path. If multiple paths have the same score and are all the highest, the path with the fewest hops is selected in ascending order. The optimal path is represented by a path identifier and a link sequence.

[0080] Based on the aforementioned selection results, a mapping relationship is established between each node and its corresponding optimal path, and written into the path decision cache. The mapping relationship uses the candidate node identifier as the primary key, and the optimal path identifier, link sequence, and path comprehensive score as attribute fields. The path decision cache organizes all mapping records in the form of a hash table, supporting fast retrieval by node identifier. The path decision cache, as the output of step S602, is used by subsequent steps at the path decision entry point for orchestrating decision instruction encapsulation and target node determination.

[0081] In one embodiment of the scenario-based computing task orchestration method of this application, it may further include the following: Step S701: Parse the computational load estimation field and the task execution mode field from the task feature data to obtain the task specification identifier. Compare the task specification identifier with the preset specification threshold condition to determine the task specification category. For large-specification category tasks, select nodes with continuous computing power resources from the candidate node set and allocate centralized resource quotas. For small-specification category tasks, select nodes with preheating containers from the candidate node set and allocate lightweight resource quotas. Determine the selection results as target nodes and the allocation results as resource quotas. Step S702: Read the optimal path corresponding to the target node from the path decision cache, encapsulate the node identifier of the target node, the quota parameters of the resource quota, and the path identifier of the optimal path into an orchestration decision instruction according to a preset message format, and send the orchestration decision instruction to the target node through the communication interface to execute the task deployment.

[0082] This embodiment reads task feature data from the task context cache, and parses the computational load estimation field and the task execution mode field to obtain the task specification identifier. The computational load estimation field indicates the computing power required by the task in terms of floating-point operations, and the task execution mode field indicates the task's execution mode in the form of an enumeration of values. The value range of the task execution mode field includes two categories: long-term execution mode and short-term batch mode. The task specification identifier encapsulates the computational load estimation value and the execution mode value in the form of a tuple, and is written to the specification parsing cache for the specification category determination operation to read.

[0083] After the task specification identifier is parsed, it is compared with a preset specification threshold to determine the task specification category. The preset specification threshold is read from the system configuration parameters and includes two items: a computational load threshold and a mode determination rule. The comparison first compares the estimated computational load with the computational load threshold, and then matches the execution mode value with the mode determination rule. If the estimated computational load is greater than or equal to the computational load threshold, or the execution mode value matches a long-running mode, the task is determined to be a large-specification category task. If the estimated computational load is less than the computational load threshold and the execution mode value matches a short-running batch mode, the task is determined to be a small-specification category task. The task specification category is written to the specification determination cache for resource combination matching operations to read.

[0084] Accordingly, for large-scale tasks, nodes with continuous computing power resources are selected from the candidate node set. The selection operation traverses each node in the candidate node set, reading the node's computing power idle rate and resource continuity flag from the dynamic resource view. The resource continuity flag indicates whether the node has continuously allocable computing power unit blocks. The selection criteria require that the computing power idle rate is higher than a preset capacity threshold and the resource continuity flag is true. The node that meets the selection criteria and has the highest comprehensive score is determined as the target node. A centralized resource quota is allocated to the target node, which includes three parameters: the number of continuous computing power units, memory capacity, and resource lockout time window.

[0085] For small-scale tasks, nodes with preheated containers are selected from the candidate node set. The selection operation iterates through each node in the candidate node set, reading the container preheating status flag of the node from the dynamic resource view. The container preheating status flag indicates whether the node has a container instance that is started and in an idle state. The selection criterion requires the container preheating status flag to be true. The node that meets the selection criteria and has the highest comprehensive score is determined as the target node. Lightweight resource quotas are allocated to the target node, and the lightweight resource quotas include three parameters: container instance identifier, number of computing units, and memory capacity. If no node has a preheated container, the node with the highest comprehensive score is selected and the container creation process is triggered, and the identifier of the newly created container instance is written into the resource quota.

[0086] After the target node and resource quota are determined, a retrieval is performed from the path decision cache using the target node's node identifier as the query key to read the optimal path corresponding to that node. The retrieval operation locates the mapping record corresponding to the target node identifier in the hash table of the path decision cache and extracts the path identifier and link sequence fields of the optimal path. The optimal path identifier represents the combination of transmission links from the origin to the target node. If the retrieval fails, a path recalculation process is triggered, and a path missing marker is registered.

[0087] Based on the aforementioned determinations, the node identifier of the target node, the quota parameters of the resource quota, and the path identifier of the optimal path are encapsulated into an orchestration decision instruction according to a preset message format. The preset message format defines the field layout of the instruction header and body. The instruction header includes three fields: instruction type identifier, instruction sequence number, and timestamp. The instruction body includes four fields: target node identifier, resource quota parameters, path identifier, and task characteristic data. The encapsulation operation fills each field into the corresponding position in the instruction message according to a preset offset, forming a complete orchestration decision instruction.

[0088] After the orchestration decision instruction is encapsulated, it is sent to the target node for task deployment via the communication interface unit. The communication interface unit determines the message forwarding route based on the optimal path link sequence and transmits the orchestration decision instruction to the target node along the specified path. The task execution module of the target node reads the orchestration decision instruction at the instruction receiving entry, parses the resource quota parameters and performs resource allocation, and starts the task instance based on the task characteristic data. The orchestration decision instruction, as the output of step S702, completes the orchestration decision and deployment process for computing power tasks.

[0089] To effectively address the shortcomings of traditional technologies in areas such as scene-aware adaptive resource weight matching, prediction-driven dynamic resource view construction and multi-dimensional weighted node selection, path quality optimization, and resource specification adaptation for collaborative orchestration and execution, and to provide technical support for scene-adaptive intelligent orchestration and high-quality end-to-end scheduling and execution of computing tasks in heterogeneous computing networks, this application provides an embodiment of a scene-based computing task orchestration device for implementing all or part of the aforementioned scene-based computing task orchestration method. See [link to embodiment]. Figure 2 The scenario-based computing task orchestration device specifically includes the following components: Task parsing module 10 is used to receive computing power task requests and parse the request content to obtain task feature data, input the task feature data into the scene classification model to perform classification inference to obtain scene type identifier, and read the corresponding multi-dimensional resource weight vector from the scene weight mapping table according to the scene type identifier; The node selection module 20 is used to collect computing power status information and network link transmission status information of computing power nodes to obtain real-time status data, input the real-time status data into the time series prediction model to obtain resource change prediction values, fuse the real-time status data and the resource change prediction values ​​to generate a dynamic resource view, perform weighted scoring calculation on candidate nodes based on the multi-dimensional resource weight vector and the dynamic resource view to obtain a node scoring sequence, and include nodes in the node scoring sequence that are higher than the screening threshold into the candidate node set. The computing power orchestration module 30 is used to calculate the transmission path from the data source to each node in the candidate node set and select the optimal path based on the path quality index. Based on the task specification identifier, it performs resource combination matching on the candidate node set to obtain the target node and resource quota. It encapsulates the target node, the resource quota and the optimal path into an orchestration decision instruction and issues it for execution.

[0090] As can be seen from the above description, the scenario-based computing task orchestration device provided in this application can infer the scenario type identifier through a scenario classification model and read the multi-dimensional resource weight vector from the scenario weight mapping table. It combines a time-series prediction model to drive the construction of a dynamic resource view and a multi-dimensional weighted node scoring to obtain a candidate node set. After selecting the optimal path through transmission path quality indicators and matching the resource combination driven by task specifications to obtain the target node and resource quota, it encapsulates the orchestration decision instruction and issues it for execution. This effectively solves the shortcomings of traditional technologies in scenario-aware adaptive resource weight matching, prediction-driven dynamic resource view construction and multi-dimensional weighted node selection, and path quality optimization and resource specification adaptation collaborative orchestration execution. It provides technical support for scenario-adaptive intelligent orchestration and high-quality scheduling execution of computing tasks in heterogeneous computing networks.

[0091] This invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the scenario-based computing task orchestration method.

[0092] This invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described scenario-based computing task orchestration method.

[0093] This invention also provides a computer program product, which includes a computer program that, when executed by a processor, implements the above-described scenario-based computing task orchestration method.

[0094] In this embodiment of the invention, a scene type identifier is inferred through a scene classification model and a multi-dimensional resource weight vector is read from a scene weight mapping table. A candidate node set is obtained by combining a time-series prediction model to drive the construction of a dynamic resource view and multi-dimensional weighted node scoring and filtering. The optimal path is selected through transmission path quality indicators and the resource combination is matched with task specifications to obtain the target node and resource quota. Then, the orchestration decision instruction is encapsulated and issued for execution. This effectively solves the shortcomings of traditional technologies in scene-aware adaptive resource weight matching, prediction-driven dynamic resource view construction and multi-dimensional weighted node filtering, path quality optimization and resource specification adaptation collaborative orchestration execution, etc. It provides technical support for scene-adaptive intelligent orchestration and high-quality end-to-end scheduling execution of computing tasks in heterogeneous computing power networks.

[0095] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0096] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0097] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0098] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0099] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A scenario-based computing task orchestration method, characterized in that, The method includes: Receive computing power task requests and parse the request content to obtain task feature data. Input the task feature data into the scene classification model to perform classification inference to obtain scene type identifier. Based on the scene type identifier, read the corresponding multi-dimensional resource weight vector from the scene weight mapping table. Real-time status data is obtained by collecting computing power status information of computing power nodes and transmission status information of network links. The real-time status data is input into a time series prediction model to obtain resource change prediction values. The real-time status data and the resource change prediction values ​​are fused to generate a dynamic resource view. Based on the multi-dimensional resource weight vector and the dynamic resource view, a weighted scoring calculation is performed on the candidate nodes to obtain a node scoring sequence. Nodes in the node scoring sequence that are higher than the screening threshold are included in the candidate node set. For each node in the candidate node set, the transmission path from the data source to the node is calculated, and the optimal path is selected based on the path quality index. Based on the task specification identifier, resource combination matching is performed on the candidate node set to obtain the target node and resource quota. The target node, the resource quota and the optimal path are encapsulated into an orchestration decision instruction and issued for execution.

2. The scenario-based computing task orchestration method according to claim 1, characterized in that, The process of receiving computing power task requests and parsing the request content to obtain task feature data, and then inputting the task feature data into a scene classification model to perform classification inference to obtain a scene type identifier, includes: The system receives computing power task requests and parses the request messages to obtain raw request data. It then performs field integrity verification and format normalization on the raw request data to obtain normalized request data. Finally, it extracts data type field, computational load estimation field, response time requirement field, data source location field, and task type identifier field from the normalized request data and assembles them into task feature data. The task feature data is numerically normalized to obtain a normalized feature vector. The normalized feature vector is then input into the input layer of the scene classification model and forward propagated through a multi-layer neural network to obtain a scene probability distribution. The scene category with the highest probability value is selected from the scene probability distribution as the scene type identifier.

3. The scenario-based computing task orchestration method according to claim 1, characterized in that, The step of reading the corresponding multi-dimensional resource weight vector from the scene weight mapping table based on the scene type identifier includes: Load the scene weight mapping table from the memory and establish an index relationship between scene type identifiers and weight records to obtain a weight index structure. Use the scene type identifier as the query key to perform a matching retrieval in the weight index structure to obtain the target weight record. The computing power weight value, latency weight value, bandwidth weight value, storage weight value, price weight value and energy consumption weight value are read from the target weight record and assembled into a multi-dimensional resource weight vector in a preset dimension order. The multi-dimensional resource weight vector is written into the task context cache for subsequent orchestration calculation and calling.

4. The scenario-based computing task orchestration method according to claim 1, characterized in that, The process involves collecting computing power status information from computing nodes and transmission status information from network links to obtain real-time status data. This real-time status data is then input into a time-series prediction model to obtain resource change prediction values. Finally, the real-time status data and the resource change prediction values ​​are fused to generate a dynamic resource view, including: Computing power status information is obtained by collecting computing power utilization, available memory capacity and computing power idle rate by probes deployed on each computing power node. Transmission status information is obtained by collecting end-to-end latency, remaining bandwidth and packet loss rate by network measurement protocol. The computing power status information and the transmission status information are aligned by timestamp and aggregated into real-time status data. The real-time status data is organized into a historical time series according to a preset time window and input into a time series prediction model to perform forward inference to obtain resource change prediction values. The fused status record is obtained by concatenating the corresponding prediction index values ​​in the real-time status data and the corresponding prediction index values ​​in the resource change prediction values ​​at the field level according to the node identifier and the link identifier. The fused status record is then indexed according to the computing power node dimension and the network link dimension and written into the memory data structure to generate a dynamic resource view.

5. The scenario-based computing task orchestration method according to claim 1, characterized in that, The step of performing a weighted scoring calculation on candidate nodes based on the multidimensional resource weight vector and the dynamic resource view to obtain a node score sequence, and including nodes in the node score sequence that are higher than the screening threshold into the candidate node set, includes: The system iterates through each computing node in the dynamic resource view and reads the corresponding computing power index, latency index, bandwidth index, storage index, and energy consumption index. It then performs normalization processing on each index value to obtain a normalized index vector. The normalized index vector is multiplied dimension by dimension by the multidimensional resource weight vector and summed to obtain the comprehensive score value of the node. Finally, the comprehensive score values ​​of each node are associated and assembled into a node score sequence according to the node identifier. The comprehensive score of each node in the node scoring sequence is compared with the screening threshold. The identifiers of nodes with comprehensive scores higher than the screening threshold are extracted and written into the candidate node set. The candidate node set is arranged in descending order of comprehensive score for subsequent path optimization and resource combination.

6. The scenario-based computing task orchestration method according to claim 1, characterized in that, The step of calculating the transmission path from the data source to each node in the candidate node set and selecting the optimal path based on path quality indicators includes: Read the data source location identifier from the task feature data, retrieve the reachable link sequence from the data source location to each node in the candidate node set from the network topology data, perform path traversal calculation based on the reachable link sequence to obtain multiple candidate transmission paths, read the latency value, remaining bandwidth value and packet loss rate value of each link in the candidate transmission path from the dynamic resource view and aggregate them into path quality indicators; The latency, remaining bandwidth, and packet loss rate values ​​in the path quality indicators are weighted and calculated with the corresponding weight values ​​in the multi-dimensional resource weight vector to obtain a comprehensive path score. The candidate transmission paths are sorted in descending order according to the comprehensive path score, and the path with the highest score is selected as the optimal path for that node. A mapping relationship is established between each node and the corresponding optimal path and written into the path decision cache.

7. The scenario-based computing task orchestration method according to claim 1, characterized in that, The step of performing resource combination matching on the candidate node set based on the task specification identifier to obtain the target node and resource quota, and encapsulating the target node, the resource quota, and the optimal path into an orchestration decision instruction and issuing it for execution includes: The task specification identifier is obtained by parsing the computational load estimation field and the task execution mode field from the task feature data. The task specification identifier is compared with the preset specification threshold condition to determine the task specification category. For large-specification category tasks, nodes with continuous computing power resources are selected from the candidate node set and a centralized resource quota is allocated. For small-specification category tasks, nodes with preheating containers are selected from the candidate node set and a lightweight resource quota is allocated. The selection result is determined as the target node and the allocation result is determined as the resource quota. The optimal path corresponding to the target node is read from the path decision cache. The node identifier of the target node, the quota parameter of the resource quota, and the path identifier of the optimal path are encapsulated into an orchestration decision instruction according to a preset message format. The orchestration decision instruction is sent to the target node through the communication interface to execute the task deployment.

8. A scenario-based computing task orchestration device, characterized in that, The device includes: The task parsing module is used to receive computing power task requests and parse the request content to obtain task feature data. The task feature data is then input into the scene classification model to perform classification inference to obtain a scene type identifier. Based on the scene type identifier, the corresponding multi-dimensional resource weight vector is read from the scene weight mapping table. The node selection module is used to collect computing power status information and network link transmission status information of computing power nodes to obtain real-time status data, input the real-time status data into the time series prediction model to obtain resource change prediction values, fuse the real-time status data and the resource change prediction values ​​to generate a dynamic resource view, perform weighted scoring calculation on candidate nodes based on the multi-dimensional resource weight vector and the dynamic resource view to obtain a node scoring sequence, and include nodes in the node scoring sequence that are higher than the screening threshold into the candidate node set. The computing power orchestration module is used to calculate the transmission path from the data source to each node in the candidate node set and select the optimal path based on the path quality index. Based on the task specification identifier, it performs resource combination matching on the candidate node set to obtain the target node and resource quota. It encapsulates the target node, the resource quota and the optimal path into an orchestration decision instruction and issues it for execution.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the scenario-based computing task orchestration method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the steps of the scenario-based computing task orchestration method according to any one of claims 1 to 7.

Images