A computing power arrangement and scheduling method and device based on a computing power service gateway and a medium

Abstract

The application discloses a kind of based on computing power service gateway's computing power arrangement scheduling method, equipment and medium, comprising: constructing task image;Constructing computing power node response image;Determine candidate computing power node set based on task-computing power adaptation matrix;Get the predicted result of resource state after scheduling;Calculate resource fragment risk value;Generate candidate arrangement scheme, obtain measured computing power response information;Get target arrangement scheme;According to execution feedback information updates task image, computing power node response image and resource fragment risk value.The application reduces the failure rate of task deployment and subsequent migration cost, improves the overall reliability and stability of computing power service.

Description

Technical Field

[0001] This invention relates to the field of computing power orchestration technology, and in particular to a computing power orchestration and scheduling method, device and medium based on a computing power service gateway. Background Technology

[0002] With the rapid development of cloud computing, edge computing, and multi-source heterogeneous computing resources, computing service gateways for multi-task concurrent processing are gradually becoming a crucial infrastructure connecting task requirements with underlying computing resources. In practical applications, a large number of computing tasks dynamically access the computing network with different resource requirements, quality of service constraints, and data security constraints. Simultaneously, computing nodes exhibit significant heterogeneity in hardware capabilities, load status, and network environment. Therefore, how to achieve efficient, stable, and sustainable computing orchestration and scheduling in complex and ever-changing computing environments has become a critical problem that urgently needs to be solved in the current technological field.

[0003] In existing technologies, computing power orchestration and scheduling methods mostly rely on scheduling strategies based on static resource views or short-term dynamic loads. They allocate tasks to computing nodes that meet current resource requirements through simple resource matching rules or greedy allocation algorithms. Typically, decisions are made only based on the current resource status, lacking the ability to model the evolution of resource occupancy trends after task deployment. This fails to reflect the impact of task execution on the future resource structure of computing nodes, easily leading to uneven distribution across different resource dimensions, generating a large number of fragmented resources that are difficult to reuse, resulting in computing power resource fragmentation and reduced overall resource utilization.

[0004] Most existing scheduling decisions rely on predictive models or rule-based reasoning results for direct task deployment, lacking online verification methods for actual operating environments. Since computing nodes experience unpredictable factors such as network jitter, container initialization delays, and transient system load fluctuations during actual operation, relying solely on model predictions for scheduling can easily lead to deployment failures or substandard performance, causing task migration and redundant scheduling issues, significantly increasing system operating overhead. Summary of the Invention

[0005] One objective of this invention is to propose a computing power orchestration and scheduling method, device, and medium based on a computing power service gateway. This invention reduces the task deployment failure rate and subsequent migration costs, and improves the overall reliability and stability of computing power services.

[0006] A computing power orchestration and scheduling method for a computing power service gateway according to an embodiment of the present invention includes: Receive computing power service requests, parse the computing power service intent information, and construct a task profile based on the computing power service intent information; Collect static capability information, dynamic load information, historical response information, data affinity information, and security domain information of multiple computing power nodes to construct a response profile of the computing power nodes; A task-computing power adaptation matrix is ​​constructed based on the task profile and the computing power node response profile, and a candidate computing power node set is determined based on the task-computing power adaptation matrix; For each candidate computing node in the candidate computing node set, sample the resource occupancy status after the LightGBM predicts the target task deployment, and obtain the resource status prediction result after scheduling. Calculate the resource fragmentation risk value based on the resource status prediction results after scheduling; The task-computing power adaptation matrix and resource fragmentation risk value are input into the improved two-layer Gale-Shapley game matching engine to generate candidate orchestration schemes. Lightweight trial deployments are then performed on the candidate orchestration schemes to obtain measured computing power response information. Based on the measured computing power response information, the candidate orchestration scheme is revised to obtain the target orchestration scheme; The target orchestration scheme is distributed to the target computing power nodes, the computing power service is deployed, the execution feedback information during the target task operation is collected, and the task profile, computing power node response profile and resource fragmentation risk value are updated based on the execution feedback information.

[0007] Optionally, receiving the computing power service request and parsing it to obtain the computing power service intent information includes: Receive computing power service requests, read the task type field, resource requirement field, service quality constraint field, and data security constraint field from the computing power service request, and generate a request field set; Validate the set of request fields to obtain a valid set of request fields; Parse the resource requirement fields to extract CPU requirements, memory requirements, video memory requirements, storage input / output requirements, and network bandwidth requirements, and generate a resource requirement vector according to the resource dimensions. Parse the service quality constraint fields to extract the maximum startup latency, maximum response latency, minimum service availability, and maximum number of migrations, and generate a service quality constraint vector according to the service quality dimensions. Parse the data security constraint fields to extract the data source location identifier, data affinity level, and security domain level, and generate a data security constraint vector according to the data security dimensions. Generate computing power service intent information based on the task type field, resource requirement vector, service quality constraint vector, and data security constraint vector; The computing power service intent information is normalized and encoded and the fields are concatenated according to the order of the profile fields to generate a task profile.

[0008] Optionally, the collection of static capability information, dynamic load information, historical response information, data affinity information, and security domain information of multiple computing nodes includes: Identify the multiple computing power nodes to be collected and generate a set of computing power nodes; Collect static capability information and dynamic load information of each computing node in the computing node set, and generate static capability vector and dynamic load vector; Calculate the residual hardware resource vector for each computing node based on the static capability vector and the dynamic load vector. Collect historical response information, data affinity information, and security domain information of each computing node in the computing node set, and generate historical response vector and node data security vector; The static capability vector, dynamic load vector, hardware residual resource vector, historical response vector, and node data security vector are normalized and encoded, and the fields are concatenated according to the order of the node profile fields to construct the computing power node response profile.

[0009] Optionally, the step of constructing a task-computing power adaptation matrix based on the task profile and the computing power node response profile includes: Based on the resource demand vector and the hardware residual resource vector, calculate the resource satisfaction degree between each computing power service task and each computing power node; Based on the service quality constraint vector and the historical response vector, calculate the service quality satisfaction between each computing power service task and each computing power node; Based on the data security constraint vector and the node data security vector, calculate the data security satisfaction degree between each computing power service task and each computing power node; Based on resource satisfaction, service quality satisfaction, and data security satisfaction, calculate the compatibility score between each computing power service task and each computing power node; The adaptation scores are arranged according to the computing power service task number and computing power node number, and a task-computing power adaptation matrix is ​​constructed. Based on the task-computing power adaptation matrix, determine the set of candidate computing power nodes.

[0010] Optionally, for each candidate computing power node in the candidate computing power node set, the steps include: For each candidate computing node in the candidate computing node set, read the resource demand vector in the task profile, and read the static capability vector, dynamic load vector, hardware residual resource vector and historical response vector in the computing node response profile to generate a set of predictive basic features. The dynamic load information of candidate computing power nodes before the current scheduling time is sampled within a time window to generate a dynamic load sampling sequence of candidate computing power nodes; Based on the predicted basic feature set and the dynamic load sampling sequence, LightGBM input feature samples are generated; Input the LightGBM input feature samples into the pre-trained LightGBM prediction model to predict the CPU usage, memory usage, GPU memory usage, storage input / output usage, and network bandwidth usage after the target task is deployed and scheduled. Based on the resource dimensions of CPU, memory, video memory, storage input / output, and network bandwidth, the predicted CPU usage, memory usage, video memory usage, storage input / output, and network bandwidth usage after scheduling are arranged in order to obtain the predicted resource status results after scheduling.

[0011] Optionally, calculating the resource fragmentation risk value based on the post-scheduling resource status prediction result includes: Read the static capability vector and the resource status prediction results after scheduling of the candidate computing power nodes, and calculate the hardware residual resource vector of the candidate computing power nodes after scheduling after the target task is deployed; Based on the remaining hardware resource vector after scheduling, calculate the proportion of remaining resources available for candidate computing power nodes; Based on the proportion of remaining available resources under each resource dimension, calculate the resource dimension imbalance risk value of the candidate computing power node; Based on the remaining resource availability ratio and the preset minimum reusable resource ratio under each resource dimension, calculate the non-reusable fragment risk value of the candidate computing power node; Based on the resource imbalance risk value and the non-reusable fragmentation risk value, the resource fragmentation risk value of the candidate computing power node is calculated.

[0012] Optionally, the step of inputting the task-computing power adaptation matrix and resource fragmentation risk value into the improved two-layer Gale-Shapley game matching engine includes: Extract the compatibility score between each computing power service task and each computing power node in the input task-computing power adaptation matrix, arrange the compatibility scores in descending order, and generate a task-side preference queue. Extract the input resource fragmentation risk value, combine it with the hardware residual resource vector of each computing node to calculate the receiving cost score, sort the receiving cost scores in ascending order, and generate a computing power side receiving willingness queue. The task-side preference queue and the computing power-side receiving willingness queue are simultaneously input into the improved two-layer Gael-Shapley game matching engine. Through multiple rounds of iterative hedging, in which tasks submit applications according to preferences and computing power nodes temporarily store or cancel according to receiving willingness, a stable matching state matrix is ​​output. Random perturbation and multi-version heterogeneous transformation of the topological boundary are performed on the stable matching state matrix to extract the set of non-conflict mappings that satisfy the equilibrium of the multi-party game and generate candidate orchestration schemes. The node routing paths in the candidate orchestration schemes are parsed and transformed into sandbox probing commands that only mount lightweight container images without loading core business datasets. These commands are then sent to the target physical nodes to execute empty runs without data that are not included in the formal billing process. The kernel-mode resource trajectories during the execution of the sandbox probing commands are captured, and transient feature data of the probing deployment is generated. Extract network I / O throughput jitter, container initialization latency, and instantaneous CPU burst heat from the transient characteristic data of the trial deployment, perform standardized weighted aggregation, and output measured computing power response information.

[0013] Optionally, the step of correcting the candidate orchestration scheme based on the measured computing power response information to obtain the target orchestration scheme includes: Read the candidate orchestration schemes and measured computing power response information to determine the target physical node, node routing path and measured response result for each computing power service task in the candidate orchestration schemes. Based on the measured computing power response information, a risk assessment is conducted on the trial deployment of target physical nodes in the candidate orchestration schemes. Based on the trial deployment of high-risk nodes, anomaly mapping identification is performed on candidate orchestration schemes to generate a set of mappings to be corrected. Based on the set of mappings to be corrected, alternative mappings are selected for abnormal mapping relationships in the candidate orchestration schemes to generate an alternative mapping set. The abnormal mapping relationships in the candidate orchestration schemes are replaced according to the alternative mapping set to generate the corrected orchestration scheme; Perform a pre-deployment consistency check on the revised orchestration scheme to obtain the target orchestration scheme.

[0014] An electronic device includes a processor and a memory, the memory storing programs or instructions executable on the processor, the programs or instructions being executed by the processor to implement the steps of the present invention.

[0015] A readable storage medium on which a program or instructions are stored, wherein the program or instructions, when executed by a processor, implement the steps of the present invention.

[0016] The beneficial effects of this invention are: (1) This invention unifies the modeling of predicted CPU usage, predicted memory usage, predicted GPU memory usage, predicted storage input / output usage, and predicted network bandwidth usage after scheduling, and further calculates the remaining resource availability ratio, resource dimension imbalance risk value, and non-reusable fragment risk value, enabling scheduling decisions to identify the generation trend of resource dimension imbalance and small-scale non-reusable fragments in advance. This invention can proactively constrain the resource structure before scheduling, suppress the accumulation problem of high-occupancy and low-reusability fragments from the source, and significantly improve the resource utilization rate of computing nodes and reduce the degree of resource fragmentation in continuous multi-task scheduling scenarios.

[0017] (2) This invention maps the fit score in the task-computing power fit matrix to the task-side preference queue, and merges the resource fragmentation risk value and the hardware residual resource vector into the receiving cost score and generates the computing power side receiving intention queue. During the matching process, through multiple rounds of application, temporary storage, rejection and contract termination, the matching result can simultaneously meet the performance requirements of the task side and the resource structure optimization constraints of the computing power side. This can avoid the problem of high-quality nodes being concentrated and low-quality nodes being idle for a long time. Under multiple constraints, it achieves collision-free mapping and stable matching, and significantly reduces multi-tenant competition conflicts and scheduling oscillations.

[0018] (3) This invention transforms candidate orchestration schemes into sandbox probing instructions that only mount lightweight container images and do not load core business datasets. These instructions are executed on target physical nodes without data, and network I / O throughput variation trajectories, container initialization latency trajectories, and CPU instantaneous burst heat trajectories are collected. Furthermore, network I / O throughput jitter values, container initialization latency, and CPU instantaneous burst heat are calculated and standardized and weighted to obtain measured computing power response information. High-risk nodes in the candidate orchestration schemes are identified and alternative mapping corrections are performed, eliminating potential failure paths before formal deployment. This invention can expose deployment risks caused by network jitter, abnormal container initialization, or transient load fluctuations in advance, reducing task deployment failure rates and subsequent migration costs, and improving the overall reliability and stability of computing power services. Attached Figure Description

[0019] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a computing power orchestration and scheduling method based on a computing power service gateway proposed in this invention; Figure 2 This is a structural diagram of a computing power orchestration and scheduling device based on a computing power service gateway proposed in this invention; Figure 3This is a diagram illustrating the improved two-layer Gale-Shapley game matching process in Embodiment 4 of the computing power orchestration and scheduling method, device, and medium based on a computing power service gateway proposed in this invention. Detailed Implementation

[0020] Example 1: Reference Figure 1 A computing power orchestration and scheduling method for a computing power service gateway, comprising: Receive computing power service requests, parse the computing power service intent information, and construct a task profile based on the computing power service intent information; In this embodiment, receiving a computing power service request and parsing it to obtain computing power service intent information includes: Receive computing power service requests, read the task type field, resource requirement field, service quality constraint field, and data security constraint field from the computing power service request, and generate a request field set; In Example 1, the task type field represents the task type corresponding to the target task, the resource requirement field represents the resource requirement corresponding to the target task, the service quality constraint field represents the service quality constraint corresponding to the target task, and the data security constraint field represents the data security constraint corresponding to the target task.

[0021] Validate the set of request fields to obtain a valid set of request fields; In Example 1, the validation of the request field set includes field integrity validation, field format validation, and field value range validation, filtering out computing power service requests with missing fields, abnormal field formats, or field values ​​exceeding the preset constraint range.

[0022] Parse the resource requirement fields to extract CPU requirements, memory requirements, video memory requirements, storage input / output requirements, and network bandwidth requirements, and generate a resource requirement vector according to the resource dimensions. In Example 1, CPU requirement represents the amount of CPU resources requested by the target task within a unit scheduling cycle, memory requirement represents the amount of memory resources requested by the target task within a unit scheduling cycle, video memory requirement represents the amount of video memory resources requested by the target task within a unit scheduling cycle, storage input / output requirement represents the amount of storage input / output resources requested by the target task within a unit scheduling cycle, and network bandwidth requirement represents the amount of network bandwidth resources requested by the target task within a unit scheduling cycle.

[0023] Parse the service quality constraint fields to extract the maximum startup latency, maximum response latency, minimum service availability, and maximum number of migrations, and generate a service quality constraint vector according to the service quality dimensions. In Example 1, maximum startup latency represents the maximum startup latency allowed for the target task, maximum response latency represents the maximum response latency allowed for the target task, minimum service availability represents the minimum service availability required for the target task, and maximum number of migrations represents the maximum number of migrations allowed for the target task.

[0024] Parse the data security constraint fields to extract the data source location identifier, data affinity level, and security domain level, and generate a data security constraint vector according to the data security dimensions. In Example 1, the data source location identifier represents the data source location identifier of the business data required by the target task, the data affinity level represents the data affinity level between the target task and the business data location, and the security domain level represents the security domain level corresponding to the target task.

[0025] Generate computing power service intent information based on the task type field, resource requirement vector, service quality constraint vector, and data security constraint vector; The computing power service intent information is normalized and encoded and the fields are concatenated according to the order of the profile fields to generate a task profile.

[0026] Collect static capability information, dynamic load information, historical response information, data affinity information, and security domain information of multiple computing power nodes to construct a response profile of the computing power nodes; In this embodiment, static capability information, dynamic load information, historical response information, data affinity information, and security domain information of multiple computing nodes are collected, including: Identify the multiple computing power nodes to be collected and generate a set of computing power nodes; In Example 1, the node registration information within the management scope of the computing power service gateway is read, and the node survival status and resource collection response status of each computing power node are received. The computing power nodes in the valid registration status are filtered according to the node registration information. The computing power nodes with heartbeat timeouts are filtered out according to the node survival status. The computing power nodes that cannot return resource collection data are filtered out according to the resource collection response status. The computing power nodes with the same node identifier are deduplicated to generate a set of computing power nodes.

[0027] The computing node set is used to record multiple computing nodes participating in computing power orchestration and scheduling. Each computing node in the computing node set is used as the object for subsequent collection of static capability information, dynamic load information, historical response information, data affinity information, and security domain information.

[0028] Collect static capability information and dynamic load information of each computing node in the computing node set, and generate static capability vector and dynamic load vector; In Example 1, the static capability information includes total CPU capacity, total memory capacity, total video memory capacity, total storage input / output capacity, and total network bandwidth capacity, which are arranged in order of static resource dimensions to generate a static capability vector.

[0029] Dynamic load information includes CPU usage, memory usage, video memory usage, storage I / O usage, and network bandwidth usage, arranged in order of dynamic load dimensions to generate a dynamic load vector.

[0030] The total CPU capacity and CPU usage are measured in the same unit; the total memory capacity and memory usage are measured in the same unit; the total video memory capacity and video memory usage are measured in the same unit; the total storage input / output capacity and storage input / output usage are measured in the same unit; and the total network bandwidth capacity and network bandwidth usage are measured in the same unit.

[0031] Calculate the residual hardware resource vector for each computing node based on the static capability vector and the dynamic load vector. In Example 1, the remaining CPU resources are obtained by subtracting the CPU usage from the total CPU capacity, the remaining memory resources are obtained by subtracting the memory usage from the total memory capacity, the remaining video memory resources are obtained by subtracting the video memory usage from the total video memory capacity, the remaining storage input / output resources are obtained by subtracting the storage input / output usage from the total storage input / output capacity, and the remaining network bandwidth resources are obtained by subtracting the network bandwidth usage from the total network bandwidth capacity.

[0032] The hardware residual resource vector is generated according to the resource dimension order of CPU residual resources, memory residual resources, video memory residual resources, storage input / output residual resources, and network bandwidth residual resources.

[0033] Collect historical response information, data affinity information, and security domain information of each computing node in the computing node set, and generate historical response vector and node data security vector; In Example 1, the historical response information includes historical deployment success rate, historical average startup latency, historical average response latency, and historical migration count. A historical response vector is generated according to the order of historical response dimensions.

[0034] Data affinity information includes data residence location identifier and data access distance level, and security domain information includes node security domain level and tenant isolation level. Node data security vectors are generated according to the data security dimensions of data residence location identifier, data access distance level, node security domain level, and tenant isolation level.

[0035] The static capability vector, dynamic load vector, hardware residual resource vector, historical response vector, and node data security vector are normalized and encoded, and the fields are concatenated according to the order of the node profile fields to construct the computing power node response profile.

[0036] In Example 1, the node profile fields are arranged in the following order: static capability vector, dynamic load vector, hardware residual resource vector, historical response vector, and node data security vector.

[0037] A task-computing power adaptation matrix is ​​constructed based on the task profile and the computing power node response profile, and a candidate computing power node set is determined based on the task-computing power adaptation matrix; In this embodiment, a task-computing power adaptation matrix is ​​constructed based on the task profile and the computing power node response profile, including: Based on the resource demand vector and the hardware residual resource vector, calculate the resource satisfaction degree between each computing power service task and each computing power node; In Example 1, for each computing power service task and each computing power node, the ratio of the remaining CPU resources to the CPU demand is calculated to obtain the single-dimensional resource satisfaction result in the CPU resource dimension. The ratio of the remaining memory resources to the memory demand is calculated to obtain the single-dimensional resource satisfaction result in the memory resource dimension. The ratio of the remaining video memory resources to the video memory demand is calculated to obtain the single-dimensional resource satisfaction result in the video memory resource dimension. The ratio of the remaining storage input / output resources to the storage input / output demand is calculated to obtain the single-dimensional resource satisfaction result in the storage input / output resource dimension. The ratio of the remaining network bandwidth resources to the network bandwidth demand is calculated to obtain the single-dimensional resource satisfaction result in the network bandwidth resource dimension.

[0038] When the single-dimensional resource satisfaction result under any resource dimension is greater than 1, the single-dimensional resource satisfaction result is counted as 1; the single-dimensional resource satisfaction results under the CPU resource dimension, memory resource dimension, video memory resource dimension, storage input / output resource dimension, and network bandwidth resource dimension are averaged to obtain the resource satisfaction degree.

[0039] Based on the service quality constraint vector and the historical response vector, calculate the service quality satisfaction between each computing power service task and each computing power node; In Example 1, for each computing power service task and each computing power node, the ratio of the historical deployment success rate to the minimum service availability rate is calculated to obtain the historical deployment success rate satisfaction result; the ratio of the maximum startup latency to the historical average startup latency is calculated to obtain the historical average startup latency satisfaction result; the ratio of the maximum response latency to the historical average response latency is calculated to obtain the historical average response latency satisfaction result; and the ratio of the maximum number of migrations to the historical number of migrations is calculated to obtain the historical number of migrations satisfaction result.

[0040] When the satisfaction result under any historical response dimension is greater than 1, the satisfaction result is counted as 1; according to the preset service quality weight, the satisfaction results of historical deployment success rate, historical average startup latency, historical average response latency, and historical migration times are weighted and summed to obtain the service quality satisfaction.

[0041] The preset service quality weights are obtained by offline statistical analysis and normalization of the impact of historical deployment success rate, historical average startup latency, historical average response latency, and historical migration count on the actual deployment success result from historical scheduling samples. The value range of each preset service quality weight is 0 to 1, and the sum of the weights is 1. The illustrated values ​​are: 0.4 for historical deployment success rate, 0.25 for historical average startup latency, 0.25 for historical average response latency, and 0.1 for historical migration count. The selection criteria are to prioritize deployment success rate, followed by startup latency and response latency, and to use the number of historical migrations as a stability constraint.

[0042] Based on the data security constraint vector and the node data security vector, calculate the data security satisfaction degree between each computing power service task and each computing power node; In Example 1, for each computing power service task and each computing power node, the data source location identifier is matched with the data residence location identifier to obtain the data source location matching result. If the data source location identifier and the data residence location identifier are consistent, the data source location matching result is 1; otherwise, it is 0.

[0043] The data affinity level is matched with the data access distance level to obtain the data affinity level satisfaction result. If the data affinity level and the data access distance level match perfectly, the data affinity level satisfaction result is 1. If they do not match perfectly, a step-down mapping is performed according to the difference between the two levels to obtain a scalar value between 0 and 1.

[0044] The security domain level is matched with the node security domain level and the tenant isolation level to obtain the security domain level satisfaction result. If both the node security domain level and the tenant isolation level are not lower than the security domain level requirement, the security domain level satisfaction result is 1; otherwise, it is 0.

[0045] Based on preset data security weights, the data security satisfaction rate is obtained by weighted summation of the data source location matching results, data affinity level satisfaction results, and security domain level satisfaction results.

[0046] The preset data security weights are obtained by offline statistical analysis and normalization based on the impact of data source location matching results, data affinity level satisfaction results, and security domain level satisfaction results on the actual deployment success results and security policy interception results in historical scheduling samples. The value range of each preset data security weight is 0 to 1, and the sum of the weights is 1. The illustrated values ​​are 0.3 for data source location matching results, 0.25 for data affinity level satisfaction results, and 0.45 for security domain level satisfaction results. The selection is based on prioritizing security domain level and tenant isolation requirements, and ensuring data source location matching and data access affinity.

[0047] Based on resource satisfaction, service quality satisfaction, and data security satisfaction, calculate the compatibility score between each computing power service task and each computing power node; In Example 1, for each computing power service task and each computing power node, the resource satisfaction, service quality satisfaction and data security satisfaction are weighted and summed according to the preset adaptation weights to obtain the adaptation score.

[0048] The preset adaptation weights are obtained by mapping the task type field, service quality constraint field, and data security constraint field in the computing power service intent information using rules. The weights corresponding to resource satisfaction, service quality satisfaction, and data security satisfaction all range from 0 to 1, and the sum of the three is 1. For example, the weight corresponding to resource satisfaction is 0.5, the weight corresponding to service quality satisfaction is 0.3, and the weight corresponding to data security satisfaction is 0.2. The weights are selected based on the relative priority of the target task's sensitivity to resource consumption, service quality, and data security.

[0049] The adaptation scores are arranged according to the computing power service task number and computing power node number, and a task-computing power adaptation matrix is ​​constructed. In Example 1, the computing power service task number is used as the row index of the task-computing power adaptation matrix, and the computing power node number is used as the column index of the task-computing power adaptation matrix. The adaptation score between each computing power service task and each computing power node is filled into the intersection of the corresponding row index and column index to construct the task-computing power adaptation matrix.

[0050] Based on the task-computing power adaptation matrix, determine the set of candidate computing power nodes.

[0051] In Example 1, for each computing power service task, all adaptation scores corresponding to the computing power service task in the task-computing power adaptation matrix are read, and computing power nodes with adaptation scores greater than or equal to the preset adaptation threshold are identified as candidate computing power nodes, generating a set of candidate computing power nodes.

[0052] The preset fit threshold is obtained by removing outliers from the fit scores between historically successfully deployed computing power service tasks and computing power nodes, and taking the lower quantile boundary. The value ranges from 0 to 1, with an example value of 0.75. The selection criterion is to ensure that the fit score of the candidate computing power node is not lower than the stable matching lower limit in the historically successfully deployed samples.

[0053] For each candidate computing node in the candidate computing node set, sample the resource occupancy status after the LightGBM predicts the target task deployment, and obtain the resource status prediction result after scheduling. In this embodiment, for each candidate computing power node in the candidate computing power node set, the following is included: For each candidate computing node in the candidate computing node set, read the resource demand vector in the task profile, and read the static capability vector, dynamic load vector, hardware residual resource vector and historical response vector in the computing node response profile to generate a set of predictive basic features. In Example 1, the resource demand vector, static capacity vector, dynamic load vector, and hardware residual resource vector are axially aligned according to the same five resource dimensions, and then normalized and encoded together with the historical response vector arranged according to the historical response dimensions. The encoded vectors are then concatenated horizontally to generate a set of predictive basic features.

[0054] The dynamic load information of candidate computing power nodes before the current scheduling time is sampled within a time window to generate a dynamic load sampling sequence of candidate computing power nodes; In Example 1, at each sampling time, the CPU usage, memory usage, video memory usage, storage input / output usage, and network bandwidth usage of the candidate computing power nodes are collected, and a dynamic load snapshot corresponding to the sampling time is generated. Multiple dynamic load snapshots are arranged in order from first to last according to the sampling time to generate a dynamic load sampling sequence of the candidate computing power nodes.

[0055] Based on the predicted basic feature set and the dynamic load sampling sequence, LightGBM input feature samples are generated; In Example 1, the predicted basic feature set and the dynamic load sampling sequence are normalized and encoded, and then concatenated according to the field order of the predicted basic feature set first and the dynamic load sampling sequence later to generate LightGBM input feature samples.

[0056] Input the LightGBM input feature samples into the pre-trained LightGBM prediction model to predict the CPU usage, memory usage, GPU memory usage, storage input / output usage, and network bandwidth usage after the target task is deployed and scheduled. In Example 1, the pre-trained LightGBM prediction model is trained using task profiles, computing node response profiles, and the actual resource usage status after the target task is deployed from historical scheduling samples as training data. The actual resource usage status after the target task is deployed includes the actual CPU usage, actual memory usage, actual GPU memory usage, actual storage input / output usage, and actual network bandwidth usage after scheduling.

[0057] The predicted CPU usage, memory usage, GPU memory usage, storage input / output usage, and network bandwidth usage after scheduling represent the predicted total usage of the candidate computing nodes in the corresponding resource dimensions after the target task is deployed to the candidate computing nodes.

[0058] Based on the resource dimensions of CPU, memory, video memory, storage input / output, and network bandwidth, the predicted CPU usage, memory usage, video memory usage, storage input / output, and network bandwidth usage after scheduling are arranged in order to obtain the predicted resource status results after scheduling.

[0059] Calculate the resource fragmentation risk value based on the resource status prediction results after scheduling; In this embodiment, the resource fragmentation risk value is calculated based on the resource status prediction results after scheduling, including: Read the static capability vector and the resource status prediction results after scheduling of the candidate computing power nodes, and calculate the hardware residual resource vector of the candidate computing power nodes after scheduling after the target task is deployed; In Example 1, the residual hardware resource vector after scheduling is obtained by subtracting the predicted occupancy of the same resource dimension in the resource state prediction result from the total capacity corresponding to each resource dimension in the static capability vector.

[0060] The resource dimensions include CPU resources, memory resources, video memory resources, storage input / output resources, and network bandwidth resources.

[0061] Based on the remaining hardware resource vector after scheduling, calculate the proportion of remaining resources available for candidate computing power nodes; In Example 1, for each candidate computing node, the ratio of the remaining CPU resources after scheduling to the total CPU capacity is calculated to obtain the remaining resource availability ratio in the CPU resource dimension. The ratio of the remaining memory resources after scheduling to the total memory capacity is calculated to obtain the remaining resource availability ratio in the memory resource dimension. The ratio of the remaining video memory resources after scheduling to the total video memory capacity is calculated to obtain the remaining resource availability ratio in the video memory resource dimension. The ratio of the remaining storage input / output resources after scheduling to the total storage input / output capacity is calculated to obtain the remaining resource availability ratio in the storage input / output resource dimension. The ratio of the remaining network bandwidth resources after scheduling to the total network bandwidth capacity is calculated to obtain the remaining resource availability ratio in the network bandwidth resource dimension.

[0062] Based on the proportion of remaining available resources under each resource dimension, calculate the resource dimension imbalance risk value of the candidate computing power node; In Example 1, the remaining resource availability ratios under the CPU resource dimension, memory resource dimension, video memory resource dimension, storage input / output resource dimension, and network bandwidth resource dimension are obtained. The maximum and minimum remaining resource availability ratios are determined from the remaining resource availability ratios under each resource dimension. The minimum remaining resource availability ratio is subtracted from the maximum remaining resource availability ratio to obtain the resource dimension imbalance risk value.

[0063] The resource dimension imbalance risk value represents the degree of residual resource difference between candidate computing nodes in different resource dimensions after the target task is deployed. The larger the resource dimension imbalance risk value, the higher the degree of resource dimension imbalance of the candidate computing nodes after the target task is deployed.

[0064] Based on the remaining resource availability ratio and the preset minimum reusable resource ratio under each resource dimension, calculate the non-reusable fragment risk value of the candidate computing power node; In Example 1, the remaining resource availability ratios under the CPU resource dimension, memory resource dimension, video memory resource dimension, storage input / output resource dimension, and network bandwidth resource dimension are compared with the preset minimum reusable resource ratio. If the remaining resource availability ratio under any resource dimension is greater than 0 and less than the preset minimum reusable resource ratio, the resource dimension is marked as an unreusable fragment dimension. The arithmetic mean of the remaining resource availability ratios corresponding to each unreusable fragment dimension is calculated to obtain the average fragment ratio. The ratio of the number of unreusable fragment dimensions to the total number of resource dimensions is calculated to obtain the quantity weight. The average fragment ratio and the quantity weight are multiplied to obtain the unreusable fragment risk value.

[0065] The preset minimum reusable resource ratio is obtained by removing outliers from the lowest resource occupancy ratio of historically successfully deployed computing power service tasks under each resource dimension, and taking the low quantile boundary. It is used to represent the minimum resource ratio lower limit that subsequent computing power service tasks can reuse residual resources.

[0066] Based on the resource imbalance risk value and the non-reusable fragmentation risk value, the resource fragmentation risk value of the candidate computing power node is calculated.

[0067] In Example 1, the resource fragmentation risk value is obtained by weighting and summing the resource dimension imbalance risk value and the non-reusable fragment risk value according to the preset fragmentation risk weight.

[0068] The preset fragmentation risk weights are obtained by offline statistical analysis and normalization of the impact of resource dimension imbalance risk values ​​and non-reusable fragmentation risk values ​​on the blocking results of subsequent tasks in historical scheduling samples. The value range of each preset fragmentation risk weight is 0 to 1 and the sum of the weights is 1. The illustrated values ​​are 0.45 for the resource dimension imbalance risk value and 0.55 for the non-reusable fragmentation risk value. The selection is based on the principle of prioritizing the suppression of non-reusable residual resource fragments and then suppressing the residual resource differences between different resource dimensions.

[0069] The task-computing power adaptation matrix and resource fragmentation risk value are input into the improved two-layer Gale-Shapley game matching engine to generate candidate orchestration schemes. Lightweight trial deployments are then performed on the candidate orchestration schemes to obtain measured computing power response information.

[0070] In this embodiment, the task-computing power adaptation matrix and resource fragmentation risk value are input into the improved two-layer Gale-Shapley game matching engine, including: Extract the compatibility score between each computing power service task and each computing power node in the input task-computing power adaptation matrix, arrange the compatibility scores in descending order, and generate a task-side preference queue. In Example 1, for each computing power service task, all the adaptation scores corresponding to the computing power service task in the task-computing power adaptation matrix are read, and the computing power nodes are sorted in descending order of adaptation scores to obtain the task-side preference queue corresponding to the computing power service task.

[0071] When there are computing power nodes with the same compatibility score, prioritize computing power nodes with lower resource fragmentation risk values; when resource fragmentation risk values ​​are the same, prioritize computing power nodes with higher historical deployment success rates.

[0072] Extract the input resource fragmentation risk value, combine it with the hardware residual resource vector of each computing node to calculate the receiving cost score, sort the receiving cost scores in ascending order, and generate a computing power side receiving willingness queue. In Example 1, for each computing node, the resource fragmentation risk value and hardware residual resource vector corresponding to the computing node are read. For each computing node, the residual CPU resources, residual memory resources, residual video memory resources, residual storage input / output resources, and residual network bandwidth resources are normalized according to the preset bearing normalization benchmark corresponding to each resource dimension to obtain the single-dimensional residual bearing result under each dimension.

[0073] When the single-dimensional residual carrying capacity result under any resource dimension is greater than 1, the single-dimensional residual carrying capacity result is counted as 1; the single-dimensional residual carrying capacity results under each resource dimension are averaged to obtain the residual resource carrying capacity margin.

[0074] The range transformation method is used to normalize the risk value of resource fragmentation to obtain the fragmentation risk standardization result. The residual resource carrying capacity is subtracted from the value 1 to obtain the residual resource consumption cost result. The fragmentation risk standardization result and the residual resource consumption cost result are weighted and summed according to the preset receiving cost weight to obtain the receiving cost score. The receiving cost score is arranged from low to high to generate the computing power side receiving intention queue.

[0075] The preset receiving cost weights are obtained by offline statistics and normalization based on the impact of fragmentation risk standardization results and residual resource consumption cost results on subsequent task blocking results and actual deployment failure results in historical scheduling samples. The values ​​range from 0 to 1 and the sum of the weights is 1. The illustrated values ​​are 0.6 for the fragmentation risk standardization results and 0.4 for the residual resource consumption cost results. The selection criteria are to prioritize suppressing resource fragmentation risk and then control the consumption of residual hardware resources.

[0076] The task-side preference queue and the computing power-side receiving willingness queue are simultaneously input into the improved two-layer Gael-Shapley game matching engine. Through multiple rounds of iterative hedging, in which tasks submit applications according to preferences and computing power nodes temporarily store or cancel according to receiving willingness, a stable matching state matrix is ​​output. In Example 1, an initial matching record table and a stable matching state matrix are established. The improved two-layer Gale-Shapley game matching engine reads the computing power service tasks that are in an unmatched state, selects the computing power node that has not yet applied for computing power service tasks and is ranked first according to the task-side preference queue as the current application computing power node, and submits a matching application. At the same time, the application computing power nodes of computing power service tasks are recorded in the initial matching record table.

[0077] After receiving a matching application, the node currently applying for computing power merges the computing power service tasks of the newly submitted matching application with the computing power service tasks that the node has already temporarily stored into a set of tasks to be sorted. Then, it sorts the computing power service tasks in the set of tasks to be sorted in ascending order according to the computing power side's willingness to receive, and obtains the node's receiving sorting result.

[0078] According to the node receiving sorting results, the computing power service tasks in the unsorted task set are read in sequence. Based on the hardware residual resource vector of the current requesting computing power node and the resource demand vector of the corresponding computing power service task, a resource capacity occupancy check is performed. The computing power service tasks that pass the resource capacity occupancy check and are ranked higher in the receiving intention are determined as temporary computing power service tasks, and the computing power service tasks that fail the resource capacity occupancy check or are ranked lower in the receiving intention are determined as rejected computing power service tasks.

[0079] For tasks requiring temporary computing power service, the current temporary computing power node is updated to the current requesting computing power node in the initial matching record table, and the current matching status is updated to temporary matching status. For tasks requiring rejected computing power service, the temporary relationship between the rejected task and the current requesting computing power node is cleared in the initial matching record table, and its current matching status is updated to unmatched status. If the rejected computing power service task was previously in a temporary matching state with the current requesting computing power node, a contract termination process is performed, and the matching status between the rejected computing power service task and the current requesting computing power node in the stable matching status matrix is ​​updated from temporary matching to unmatched.

[0080] After completing one round of application, temporary storage, rejection, and termination, the stable matching state matrix is ​​updated according to the initial matching record table. The matching status between the computing power service task in the temporary storage matching state and its current temporary storage computing power node is marked as temporary storage matching, and the matching status between the computing power service task not in the temporary storage matching state and its corresponding computing power node is marked as unmatched.

[0081] The process of repeatedly executing the following steps—submitting matching requests for unmatched computing power service tasks, generating nodes to receive sorting results, performing resource capacity occupancy checks on computing power nodes, temporarily storing or canceling computing power service tasks, and updating the stable matching state matrix—continues until there are no more unmatched computing power service tasks that have not yet applied for computing power nodes, or all computing power service tasks have entered the temporary matching state, thus obtaining the stable matching state matrix.

[0082] Random perturbation and multi-version heterogeneous transformation of the topological boundary are performed on the stable matching state matrix to extract the set of non-conflict mappings that satisfy the equilibrium of the multi-party game and generate candidate orchestration schemes. In Example 1, the temporary matching relationships in the stable matching state matrix are read, and node replacement perturbation is performed for computing power nodes that are in the same set of candidate computing power nodes and whose fitness score difference is less than the preset perturbation threshold. Routing path perturbation is performed for the temporary matching relationships with multiple node routing paths, and multi-version heterogeneous transformation is performed for different resource dimension priorities to generate multiple perturbation matching versions.

[0083] The preset perturbation threshold is obtained by removing outliers from the fit score difference between adjacent ranked computing power nodes in the same candidate computing power node set in the historical successful deployment samples and taking the high quantile boundary. The value range is from 0 to 1, and the illustrative value is 0.05.

[0084] Multiple heterogeneous transformations are performed on multiple perturbation matching versions, including replacing the same computing power service task among candidate computing power nodes that meet the task-side preference queue sorting requirements, replacing the same computing power node among computing power service tasks that meet the computing power side reception intention queue sorting requirements, and replacing node routing paths among reachable links, to obtain multiple heterogeneous matching versions.

[0085] Multiple heterogeneous matching versions are checked for node capacity conflict, tenant isolation conflict, node routing path conflict, and resource fragmentation risk conflict. Heterogeneous matching versions with node capacity conflict, tenant isolation conflict, node routing path conflict, or resource fragmentation risk conflict are filtered out to obtain feasible heterogeneous matching versions.

[0086] A multi-party game equilibrium verification is performed on feasible heterogeneous matching versions. If any computing power service task cannot switch to a computing power node ranked higher in the task-side preference queue without reducing the matching stability of other computing power service tasks, and any computing power node cannot replace the currently received computing power service task without increasing the risk value of resource fragmentation, then the feasible heterogeneous matching version is determined as a non-conflict mapping version that satisfies the multi-party game equilibrium.

[0087] All non-conflict mapping versions are aggregated to form a non-conflict mapping set. From the non-conflict mapping set, non-conflict mapping versions that meet the preset fit requirements in terms of fit score and resource fragmentation risk value are selected to generate candidate orchestration schemes.

[0088] The preset adaptability requirement is obtained by removing outliers from the adaptability scores between historically successfully deployed computing power service tasks and computing power nodes, and taking the low percentile boundary. The preset fragmentation risk requirement is obtained by removing outliers from the resource fragmentation risk values ​​of candidate computing power nodes that have not experienced subsequent task blocking in the past, and taking the high percentile boundary. Both values ​​range from 0 to 1.

[0089] The node routing paths in the candidate orchestration schemes are parsed and transformed into sandbox probing commands that only mount lightweight container images without loading core business datasets. These commands are then sent to the target physical nodes to execute empty runs without data that are not included in the formal billing process. The kernel-mode resource trajectories during the execution of the sandbox probing commands are captured, and transient feature data of the probing deployment is generated. In Example 1, the mapping relationship between computing power service tasks and computing power nodes in the candidate orchestration scheme and the corresponding node routing path are read. The target physical node, image repository address, image pull path, container startup entry and node-side resource namespace are determined according to the node routing path. Sandbox probing instructions for the target physical node are generated according to the target physical node, image repository address, image pull path, container startup entry and node-side resource namespace.

[0090] After receiving the sandbox probe command, the target physical node pulls the lightweight container image according to the image pull path, starts the lightweight container image in the node-side resource namespace, and masks the core business dataset loading process according to the empty data mount identifier, so that the lightweight container image can run without data without loading the core business dataset and will not be included in the formal billing.

[0091] During the data-free run, network I / O throughput observations, container initialization latency observations, and CPU instantaneous burst heat observations are continuously sampled according to the trajectory sampling period to obtain network I / O throughput sampling sequences, container initialization latency sampling sequences, and CPU instantaneous burst heat sampling sequences.

[0092] The network I / O throughput change trajectory is generated based on the network I / O throughput sampling sequence, the container initialization latency trajectory is generated based on the container initialization latency sampling sequence, and the CPU instantaneous burst heat trajectory is generated based on the CPU instantaneous burst heat sampling sequence. The network I / O throughput change trajectory, the container initialization latency trajectory, and the CPU instantaneous burst heat trajectory are then concatenated according to the order of the transient feature fields of the trial deployment to generate the transient feature data of the trial deployment.

[0093] Extract network I / O throughput jitter, container initialization latency, and instantaneous CPU burst heat from the transient characteristic data of the trial deployment, perform standardized weighted aggregation, and output measured computing power response information.

[0094] In Example 1, network I / O throughput jitter, container initialization latency, and CPU instantaneous burst heat are extracted from the transient characteristic data of the trial deployment. The variance of the network I / O throughput change trajectory is calculated as the network I / O throughput jitter value. The maximum latency value in the container initialization latency trajectory is extracted as the container initialization latency, and the highest temperature value in the CPU instantaneous burst heat trajectory is extracted as the CPU instantaneous burst heat. The network I / O throughput jitter value, container initialization latency, and CPU instantaneous burst heat are standardized respectively to obtain the standardized results of network I / O throughput jitter, container initialization latency, and CPU instantaneous burst heat.

[0095] Based on the preset measured response weights, the standardized results of network I / O throughput jitter, container initialization latency, and CPU instantaneous burst heat are weighted and aggregated to obtain the measured computing power response information.

[0096] The preset measured response weights are obtained by offline statistical analysis and normalization of the impact of standardized network I / O throughput jitter, standardized container initialization latency, and standardized CPU instantaneous burst heat on actual deployment failure results in historical trial deployment samples. The values ​​range from 0 to 1, and the sum of the weights is 1. The illustrated values ​​are: 0.4 for the standardized network I / O throughput jitter result, 0.35 for the standardized container initialization latency result, and 0.25 for the standardized CPU instantaneous burst heat result. The selection criteria are to prioritize the identification of deployment failure risks caused by network I / O throughput jitter, followed by the identification of deployment failure risks caused by container initialization latency, and to use CPU instantaneous burst heat as a transient load-bearing capacity constraint for the target physical node.

[0097] The improved two-layer Gael-Shapley game matching engine structure in this embodiment differs from traditional Gael-Shapley matching, which relies solely on unilateral or fixed bilateral preferences for stable matching. It transforms the task-side fit score into a task-side preference queue, allowing tasks to submit applications in descending order of fit score. It merges resource fragmentation risk values ​​and hardware residual resource vectors into a receiving cost score, enabling computing nodes to be temporarily stored or released in ascending order of receiving cost score. After iterative matching, it introduces topological boundary random perturbation, multi-version heterogeneous transformation, and multi-party game equilibrium verification, ensuring that the output not only satisfies stable matching relationships but also simultaneously meets constraints on node capacity, tenant isolation, node routing paths, and resource fragmentation risk, reducing resource contention conflicts, node overload, and subsequent task blocking risks under multi-task concurrency.

[0098] In this embodiment, the candidate orchestration scheme is corrected based on the measured computing power response information to obtain the target orchestration scheme, including: Determine the target physical node, node routing path, and measured response results for each computing power service task in the candidate orchestration scheme; In Example 1, the candidate orchestration scheme includes the mapping relationship between computing power service tasks and computing power nodes, as well as the corresponding node routing path. The computing power node corresponding to the computing power service task is determined according to the mapping relationship between the computing power service task and the computing power node. The target physical node corresponding to the computing power service task is determined according to the node routing path, and the node routing path is used as the transmission path for deploying the computing power service task to the target physical node.

[0099] The measured computing power response information is obtained by reading the network I / O throughput jitter standardization results, container initialization latency standardization results, and CPU instantaneous burst heat standardization results generated by the target physical node during the lightweight trial deployment process from the measured computing power response information. The measured computing power response information is obtained by weighted aggregation of the network I / O throughput jitter standardization results, container initialization latency standardization results, and CPU instantaneous burst heat standardization results. The network I / O throughput jitter standardization results, container initialization latency standardization results, CPU instantaneous burst heat standardization results, and measured computing power response information are collectively determined as the measured response results.

[0100] Based on the measured computing power response information, a risk assessment is conducted on the trial deployment of target physical nodes in the candidate orchestration schemes. In Example 1, the network I / O throughput jitter standardization result is compared with the preset network I / O throughput jitter threshold, the container initialization latency standardization result is compared with the preset container initialization latency threshold, the CPU instantaneous burst heat standardization result is compared with the preset CPU instantaneous burst heat threshold, and the measured computing power response information is compared with the preset measured response threshold.

[0101] If the network I / O throughput jitter standardization result is greater than the preset network I / O throughput jitter threshold, the container initialization latency standardization result is greater than the preset container initialization latency threshold, the CPU instantaneous burst heat standardization result is greater than the preset CPU instantaneous burst heat threshold, or the measured computing power response information is greater than the preset measured response threshold, then the corresponding target physical node will be marked as a high-risk node for trial deployment. If the network I / O throughput jitter standardization result is less than or equal to the preset network I / O throughput jitter threshold, the container initialization latency standardization result is less than or equal to the preset container initialization latency threshold, the CPU instantaneous burst heat standardization result is less than or equal to the preset CPU instantaneous burst heat threshold, and the measured computing power response information is less than or equal to the preset measured response threshold, then the corresponding target physical node will be marked as a usable node for trial deployment.

[0102] The risk assessment results of the trial deployment are used to record the high-risk node markers or available node markers for trial deployment of the target physical node.

[0103] The preset network I / O throughput jitter threshold, preset container initialization latency threshold, preset CPU instantaneous burst heat threshold, and preset measured response threshold are all obtained by removing outliers from the measured response results of historical trial deployment samples that have not experienced actual deployment failures and taking the high quantile boundary. The values ​​range from 0 to 1.

[0104] Based on the trial deployment of high-risk nodes, anomaly mapping identification is performed on candidate orchestration schemes to generate a set of mappings to be corrected. In Example 1, if the target physical node corresponding to any computing power service task in the candidate orchestration scheme is marked as a high-risk node for trial deployment, then the mapping relationship between the computing power service task and the target physical node is marked as an abnormal mapping relationship.

[0105] All abnormal mapping relationships are recorded according to the computing power service task number and the target physical node to generate a set of mappings to be corrected.

[0106] Based on the set of mappings to be corrected, alternative mappings are selected for abnormal mapping relationships in the candidate orchestration schemes to generate an alternative mapping set. In Example 1, for each abnormal mapping relationship in the mapping set to be corrected, the corresponding computing power service task, target physical node, node routing path, task-side preference queue, computing power-side receiving willingness queue, non-conflict mapping set, resource fragmentation risk value, and measured computing power response information are read.

[0107] From the non-conflict mapping set, select mapping relationships that are not marked as high-risk nodes for trial deployment, meet the task-side preference queue sorting requirements, meet the computing power-side receiving intention queue sorting requirements, have resource fragmentation risk values ​​that meet preset fragmentation risk requirements, and have measured computing power response information that meet preset measured response thresholds, and use them as candidate alternative mapping relationships.

[0108] When multiple candidate alternative mapping relationships exist, the candidate alternative mapping relationship with the higher fitness score is selected first; when the fitness scores are the same, the candidate alternative mapping relationship with the lower resource fragmentation risk value is selected first; when the resource fragmentation risk values ​​are the same, the candidate alternative mapping relationship with the lower measured computing power response information is selected first, and the finally selected candidate alternative mapping relationship is determined as the target alternative mapping relationship.

[0109] Establish a correspondence between the abnormal mapping relationship and the target substitution mapping relationship to generate a set of substitution mappings.

[0110] The abnormal mapping relationships in the candidate orchestration schemes are replaced according to the alternative mapping set to generate the corrected orchestration scheme; In Example 1, candidate orchestration schemes are copied to obtain orchestration schemes to be corrected. Abnormal mapping relationships are read one by one according to the set of mappings to be corrected, and the target replacement mapping relationship corresponding to the abnormal mapping relationship is read.

[0111] If the target substitution mapping relationship meets the requirements of resource fragmentation risk value, measured computing power response information, task-side preference queue sorting requirements, and computing power-side receiving willingness queue sorting requirements, then the target physical node corresponding to the abnormal mapping relationship in the orchestration scheme to be corrected will be replaced with the target physical node corresponding to the target substitution mapping relationship, the node routing path corresponding to the abnormal mapping relationship will be replaced with the node routing path corresponding to the target substitution mapping relationship, and the mapping relationship between computing power service tasks and target physical nodes will be updated synchronously.

[0112] For mapping relationships in the candidate orchestration schemes that are not marked as abnormal mapping relationships, the mapping relationships between the target physical node, the node routing path, and the computing power service task and the target physical node remain unchanged.

[0113] After replacing all abnormal mapping relationships in the set of mappings to be corrected, the corrected orchestration scheme is generated.

[0114] Perform a pre-deployment consistency check on the revised orchestration scheme to obtain the target orchestration scheme.

[0115] In Example 1, the modified orchestration scheme is subjected to node capacity consistency verification, tenant isolation consistency verification, node routing path consistency verification, resource fragmentation risk consistency verification, and measured response consistency verification.

[0116] If the revised orchestration scheme passes the node capacity consistency check, tenant isolation consistency check, node routing path consistency check, resource fragmentation risk consistency check, and measured response consistency check simultaneously, then the revised orchestration scheme will be determined as the target orchestration scheme.

[0117] If the revised orchestration scheme fails any consistency check, the process returns to selecting an alternative mapping until a target orchestration scheme that passes all consistency checks is generated. If the number of iterations of selecting an alternative mapping reaches the preset iteration limit threshold and still fails to pass all consistency checks, the global computing power orchestration rework retry mechanism is triggered.

[0118] The preset iteration upper limit threshold is obtained by offline statistics of the number of iteration rounds required to successfully pass all consistency checks in historical scheduling samples and taking the high quantile boundary. The value range is an integer from 5 to 50. The example value is 30. The selection is based on maximizing the convergence probability of the alternative mapping scheme while avoiding the risk of the computing power service gateway being suspended due to long-term unsolvable calculations.

[0119] The target orchestration scheme is distributed to the target computing power nodes, the computing power service is deployed, the execution feedback information during the target task operation is collected, and the task profile, computing power node response profile and resource fragmentation risk value are updated based on the execution feedback information.

[0120] In this embodiment, the target orchestration scheme is distributed to the target computing power nodes, computing power service deployment is executed, and execution feedback information during the target task execution process is collected, including: Read the target orchestration scheme and determine the target computing power node, node routing path, resource allocation results, and deployment execution order for each computing power service task in the target orchestration scheme. In Example 1, the target orchestration scheme includes the mapping relationship between computing power service tasks and target computing power nodes, as well as the corresponding node routing path. The target computing power node corresponding to the computing power service task is determined according to the mapping relationship between the computing power service task and the target computing power node. The transmission path for deploying the computing power service task to the target computing power node is determined according to the node routing path. The resource allocation result corresponding to the computing power service task is determined according to the resource occupancy arrangement in the target orchestration scheme. Finally, the deployment execution order is determined according to the orchestration order of each computing power service task in the target orchestration scheme.

[0121] Based on the target computing power node, node routing path, resource allocation results, and deployment execution order, a computing power service deployment instruction is generated and sent to the target computing power node. In Example 1, container image information and business data loading information are read from the computing power service request.

[0122] The target task to be deployed is determined based on the computing power service task identifier. The target computing power node to receive the deployment instruction is determined based on the target computing power node identifier. The transmission path of the computing power service deployment instruction is determined based on the node routing path. The amount of CPU resources, memory resources, video memory resources, storage input / output resources, and network bandwidth resources allocated by the target computing power node to the target task is determined based on the resource allocation results. The container image pull address corresponding to the target task is determined based on the container image information. The business data access path corresponding to the target task is determined based on the business data loading information. The order of distribution among multiple computing power service tasks is determined based on the deployment execution order.

[0123] The computing power service task identifier, target computing power node identifier, node routing path, resource allocation result, container image information, business data loading information, and deployment execution order are encapsulated according to the field order of the deployment instruction to generate a computing power service deployment instruction.

[0124] The computing power service gateway distributes the computing power service deployment instructions to the corresponding target computing power nodes via the node routing path according to the deployment execution order, so that the target computing power nodes can perform container image pulling, business data loading, runtime environment initialization and computing power service startup according to the resource allocation results.

[0125] During the deployment of computing services on the target computing nodes, feedback information during the deployment phase is collected. In Example 1, the deployment phase feedback information includes deployment success results, deployment failure results, actual startup latency, image retrieval latency, business data loading latency, and resource usage status during the deployment phase.

[0126] If the target computing node completes container image pulling, business data loading, runtime environment initialization, and computing service startup, a deployment success result will be generated; if the target computing node fails during container image pulling, business data loading, runtime environment initialization, or computing service startup, a deployment failure result will be generated.

[0127] The deployment success or failure result is identified as the deployment result identifier. The deployment result identifier, actual startup latency, image pull latency, business data loading latency, and resource usage status during the deployment stage are concatenated according to the order of the deployment stage feedback fields to generate deployment stage feedback information.

[0128] During the execution of the target task, feedback information during the execution phase is collected, and execution feedback information is generated based on the feedback information during the deployment phase and the feedback information during the execution phase. In Example 1, the feedback information during the operation phase includes actual response latency, actual service availability, actual migration count, CPU usage during the operation phase, memory usage during the operation phase, video memory usage during the operation phase, storage input / output usage during the operation phase, and network bandwidth usage during the operation phase.

[0129] Arrange the CPU usage, memory usage, video memory usage, storage input / output usage, and network bandwidth usage during the runtime phase in order of runtime resource dimensions to generate the runtime resource usage status. Concatenate the deployment success result, deployment failure result, actual startup latency, image pull latency, business data loading latency, deployment phase resource usage status, actual response latency, actual service availability, actual migration count, and runtime resource usage status in order of execution feedback fields to generate execution feedback information.

[0130] Based on the execution feedback information, update the task profile, computing node response profile, and resource fragmentation risk value.

[0131] In Example 1, the resource occupancy status during the deployment and operation phases of the execution feedback information are read, and compared with the resource demand vector in the task profile to obtain the resource demand deviation result. The actual startup latency, actual response latency, actual service availability, and actual migration counts in the execution feedback information are read, and compared with the service quality constraint vector in the task profile to obtain the service quality deviation result. Based on the resource demand deviation result and the service quality deviation result, the resource demand vector and service quality constraint vector in the task profile are updated to obtain the updated task profile.

[0132] Update the dynamic load vector of the target computing node based on the resource occupancy status during the deployment and operation phases; recalculate the residual hardware resource vector of the target computing node based on the updated dynamic load vector and the static capability vector of the target computing node.

[0133] Update the historical response vector of the target computing node based on deployment success results, deployment failure results, actual startup latency, actual response latency, and actual migration count.

[0134] The updated dynamic load vector, hardware residual resource vector, and historical response vector are normalized and encoded, and the fields are concatenated according to the order of the node profile fields to obtain the updated computing power node response profile.

[0135] The hardware residual resource vector of the target computing node is recalculated, and the residual resource availability ratio of the target computing node under each resource dimension is calculated based on the recalculated hardware residual resource vector. The resource dimension imbalance risk value is recalculated based on the residual resource availability ratio under each resource dimension, and the non-reusable fragmentation risk value is recalculated based on the residual resource availability ratio under each resource dimension and the preset minimum reusable resource ratio. The recalculated resource dimension imbalance risk value and the recalculated non-reusable fragmentation risk value are weighted and summed according to the preset fragmentation risk weight to obtain the updated resource fragmentation risk value.

[0136] The preset minimum reusable resource ratio is obtained by removing outliers and taking the low quantile boundary from the lowest resource occupancy ratio of historically successfully deployed computing power service tasks in each resource dimension.

[0137] Example 2: Reference Figure 2 This embodiment also provides a computer device, including a system memory 1005 and a processor 1001. The system memory 1005 stores a computer program, and the processor 1001 executes the computer program to implement the steps of any of the methods described above.

[0138] It should be noted that the processor 1001 is used to execute the steps in the above method embodiments according to the instructions in the program code. Alternatively, when the processor 1001 executes the computer program, it implements the functions of each module / unit in the above system / device embodiments.

[0139] Specifically, in this embodiment, the computer program can be divided into one or more modules / units, which are stored in the system memory 1005 and executed by the processor 1001 to complete this application. The one or more modules / units can be a series of computer program instruction segments capable of performing specific functions, which describe the execution process of the computer program in the terminal device.

[0140] The terminal device can be a desktop computer, laptop, handheld computer, or cloud server, etc. The terminal device may include, but is not limited to, a processor 1001 and a system memory 1005. Those skilled in the art will understand that this does not constitute a limitation on the terminal device; it may include more or fewer components than shown in the figures, or a combination of certain components, or different components. For example, the terminal device may also include an input / output device 1003, a network access device 1002, a bus 1006, etc.

[0141] The processor 1001 can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor.

[0142] System memory 1005 can be an internal storage unit of the terminal device, such as a hard drive or RAM. System memory 1005 can also be a storage device 1004 of the terminal device, such as an external hard drive, SmartMedia Card (SMC), Secure Digital (SD) card, or FlashCard. Furthermore, system memory 1005 can include both internal storage units and storage device 1004. System memory 1005 is used to store computer programs and other programs and data required by the terminal device. System memory 1005 can also be used to temporarily store data that has been output or will be output.

[0143] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0144] Example 3: This example provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of any of the methods described above.

[0145] The computer-readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device, or any combination thereof. More specific examples of computer-readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM), registers, hard disks, optical fibers, compact disc read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof, or any other form of computer-readable storage medium in the art.

[0146] An exemplary storage medium is coupled to a processor, enabling the processor to read information from and write information to the storage medium. Of course, the storage medium can also be a component of the processor. The processor and storage medium can reside within an application-specific integrated circuit (ASIC). In embodiments of the invention, the computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, system, or device.

[0147] Example 4: In this example, the computing power service gateway is applied to a heterogeneous computing power resource pool composed of CPU-type computing power nodes, GPU-type computing power nodes, and storage input / output enhanced computing power nodes. The resource pool has a total of 36 computing power nodes, including 12 GPU-type computing power nodes, 14 CPU-type computing power nodes, and 10 storage input / output enhanced computing power nodes. The resource pool receives 86 computing power service requests within a continuous scheduling window, including 21 model training tasks, 28 online inference tasks, 24 data cleaning tasks, and 13 temporary analysis tasks. During implementation, the implementers found that if the traditional current remaining resource priority scheduling method is adopted, multiple model training tasks will be concentrated in a few GPU-type computing power nodes with high video memory, resulting in rapid exhaustion of video memory, while CPU resources, memory resources, and network bandwidth still have a large surplus, forming resource fragments that are difficult to reuse by subsequent tasks; at the same time, although some online inference tasks are allocated to computing power nodes with static resource requirements, network I / O throughput jitter and abnormal container initialization latency occur during formal deployment, leading to deployment failure and task migration.

[0148] Within the continuous scheduling window, the computing power service gateway receives a model training task request and marks it as Task A. Task A's task type field is "model training task," and its resource requirements include 8 CPU cores, 48GB memory, 24GB GPU memory, 420MB / s storage I / O, and 180Mbps network bandwidth. Its quality of service constraints include a maximum startup latency of 45 seconds, a maximum response latency of 220 milliseconds, a minimum service availability of 0.985, and a maximum number of migrations of 1. Its data security constraints include a data source location identifier of internal data domain, a data affinity level of high affinity, and a security domain level of three. After reading the task type, resource requirement, quality of service constraint, and data security constraint fields, the computing power service gateway generates a request field set and performs field integrity, format, and value range checks on the request field set. Since there are no missing fields in Task A, the resource requirements do not exceed the preset resource limit, and the service quality constraints are within the preset constraint range, a valid request field set is generated. Furthermore, a resource requirement vector, a service quality constraint vector, and a data security constraint vector are generated to form the task profile corresponding to Task A.

[0149] The computing power service gateway reads the node registration information, node liveness status, and resource acquisition response status of the computing power nodes. Among the 36 registered computing power nodes, one node's heartbeat timed out, one node failed to return resource acquisition data, and two nodes had the same node identifier and were deduplicated and merged, ultimately resulting in 34 computing power nodes participating in this round of scheduling. Figure 3Node 1 in the dataset is a GPU-based computing node. The collected data shows a total CPU capacity of 48 cores, a total memory capacity of 256GB, a total GPU memory capacity of 80GB, a total storage I / O speed of 1800MB / s, and a total network bandwidth capacity of 2000Mbps. At the current scheduling moment, Node 1's CPU utilization is 24 cores, memory utilization is 128GB, GPU memory utilization is 44GB, storage I / O utilization is 760MB / s, and network bandwidth utilization is 690Mbps. By subtracting the utilization of the corresponding resource dimension from the total capacity of each resource dimension, the remaining resources of Node 1 are calculated as follows: CPU remaining resources: 24 cores; memory remaining resources: 128GB; GPU memory remaining resources: 36GB; storage I / O utilization remaining resources: 1040MB / s; and network bandwidth remaining resources: 1310Mbps. The historical deployment success rate of node 1 is 0.972, the historical average startup latency is 31 seconds, the historical average response latency is 168 milliseconds, and the historical migration count is 2. The data residence location identifier of node 1 is consistent with the data source location identifier of task A, the data access distance level meets the high affinity requirement, and the node security domain level and tenant isolation level are not lower than the security domain level requirement. Therefore, the computing power node response profile of node 1 is retained for constructing the task-computing power adaptation matrix.

[0150] The computing power service gateway calculates the compatibility score between Task A and each of the 34 computing power nodes. For Node 1, the ratios of residual CPU resources to CPU demand, residual memory resources to memory demand, residual GPU memory resources to GPU demand, residual storage I / O resources to storage I / O demand, and residual network bandwidth resources to network bandwidth demand are all greater than 1, and are all counted as 1, resulting in a resource satisfaction score of 1. Node 1's historical deployment success rate to minimum service availability ratio is 0.987, the ratio of maximum startup latency to historical average startup latency is greater than 1 and is counted as 1, the ratio of maximum response latency to historical average response latency is greater than 1 and is counted as 1, and the ratio of maximum migration count to historical migration count is 0.5. Based on the preset service quality weights, the service quality satisfaction score is 0.895. Since Node 1's data source location matching result, data affinity level satisfaction result, and security domain level satisfaction result are all 1, the data security satisfaction score is 1. Weighted according to resource satisfaction (0.5), service quality satisfaction (0.3), and data security satisfaction (0.2), the compatibility score between task A and node 1 is 0.969. Node 2 is another GPU-based computing node. Although node 2 has 42GB of residual GPU memory, its historical average startup latency is 36 seconds and it has migrated once. After comprehensive calculation, the compatibility score between task A and node 2 is 0.946. Node 3 is another GPU-based computing node, and after comprehensive calculation, the compatibility score between task A and node 3 is 0.931. Therefore, the task preference queues generated by the computing power service gateway for task A are, in order, node 1, node 2, and node 3. The compatibility score between task A and node 1 is 0.969, the compatibility score between task A and node 2 is 0.946, and the compatibility score between task A and node 3 is 0.931. Figure 3 The task-side preference of task A on the left corresponds to the queue.

[0151] Within the same continuous scheduling window, the computing power service gateway also received an online inference task request and marked it as Task B. Task B's resource requirements include 4 CPU cores, 16GB memory, 8GB GPU memory, 160MB / s storage I / O, and 260Mbps network bandwidth. Its quality of service constraints include a maximum startup latency of 20 seconds, a maximum response latency of 80 milliseconds, a minimum service availability of 0.99, and a maximum number of migrations of 1. The computing power service gateway generates a task profile for Task B using the same processing method as Task A, and calculates the compatibility score between Task B and each candidate computing power node. The calculation results show that the compatibility score between Task B and Node 2 is 0.955, the compatibility score between Task B and Node 4 is 0.921, and the compatibility score between Task B and Node 1 is 0.9. Therefore, the task-side preference queues generated by the computing power service gateway for task B are, in order, nodes 2, 4, and 1. The compatibility score between task B and node 2 is 0.955, the compatibility score between task B and node 4 is 0.921, and the compatibility score between task B and node 1 is 0.9. Figure 3 The task-side preference of task B on the left corresponds to the queue.

[0152] After the candidate computing power node set is determined, the computing power service gateway performs LightGBM prediction for each candidate computing power node. The implementers pre-trained the LightGBM prediction model using 11,970 valid historical scheduling samples, of which 8,379 were used as training samples, 1,795 as validation samples, and 1,796 as test samples. Each historical scheduling sample includes a task profile, a computing power node response profile, a dynamic load sampling sequence, and the actual resource usage status after the target task deployment. Model test results show that the average absolute error of actual CPU usage after scheduling is 5.8%, the average absolute error of actual memory usage after scheduling is 4.9%, the average absolute error of actual GPU memory usage after scheduling is 6.6%, the average absolute error of actual storage I / O usage after scheduling is 7.2%, and the average absolute error of actual network bandwidth usage after scheduling is 6.1%.

[0153] For the candidate mapping relationship between task A and node 1, the computing power service gateway collects 12 dynamic load snapshots of node 1 before the current scheduling time. The dynamic load sampling sequence shows that the memory usage of node 1 increased from 38GB to 44GB, the network bandwidth usage increased from 520Mbps to 690Mbps, and the storage I / O usage increased from 610MB / s to 760MB / s. The LightGBM prediction model outputs that after deploying task A, the predicted CPU usage of node 1 after scheduling is 35 cores, the predicted memory usage after scheduling is 184GB, the predicted memory usage after scheduling is 71GB, the predicted storage I / O usage after scheduling is 1295MB / s, and the predicted network bandwidth usage after scheduling is 1045Mbps. Based on the static capability vector and the predicted resource status after scheduling, the remaining CPU resources of node 1 after deploying task A are: 13 cores; remaining memory resources: 72GB; remaining video memory resources: 9GB; remaining storage I / O resources: 505MB / s; and remaining network bandwidth resources: 955Mbps. The remaining resource availability ratios for the five resource dimensions are 0.271, 0.281, 0.113, 0.281, and 0.478, respectively. The difference between the maximum and minimum remaining resource availability ratios is 0.365, resulting in a resource dimension imbalance risk value of 0.365. Since the remaining resource availability ratio of 0.113 for the video memory resource dimension is greater than 0 and less than the preset minimum reusable resource ratio of 0.18, the video memory resource dimension is marked as an unreusable fragmentation dimension, with an unreusable fragmentation risk value of 0.0226. After weighting according to the preset fragmentation risk weights, the resource fragmentation risk value of node 1 after deploying task A is 0.177. The resource fragmentation risk value of node 2 after deploying task A is 0.092, and the resource fragmentation risk value of node 3 after deploying task A is 0.118. Although the adaptability scores of nodes 2 and 3 are lower than those of node 1, they are more balanced in terms of residual resource structure after scheduling.

[0154] When 86 computing power service tasks participate in unified scheduling, the computing power service gateway generates a task-side preference queue based on the task-computing power adaptation matrix, and a computing power-side receiving willingness queue based on the resource fragmentation risk value and the hardware residual resource vector. For node 1, the computing power service gateway reads the resource fragmentation risk value and the hardware residual resource vector corresponding to node 1, and calculates the receiving cost score by combining the resource demand vectors of different computing power service tasks. Task C is a data cleaning task, and the receiving cost score for node 1 receiving task C is 0.314; Task A is a model training task, and the receiving cost score for node 1 receiving task A is 0.387; Task B is an online inference task, and the receiving cost score for node 1 receiving task B is 0.521. Since a lower reception cost score indicates a higher willingness of the computing power node to accept computing power service tasks, the reception willingness queue for node 1 is, in order, task C, task A, and task B. The reception cost score for node 1 to receive task C is 0.314, for task A it is 0.387, and for task B it is 0.521. The queue is... Figure 3 The computing power side of node 1 on the right corresponds to the willingness queue.

[0155] For node 2, the computing power service gateway also reads the resource fragmentation risk value and hardware residual resource vector corresponding to node 2, and calculates the reception cost score by combining it with the resource requirement vector of different computing power service tasks. The reception cost score for node 2 receiving task A is 0.296, the reception cost score for node 2 receiving task D is 0.401, and the reception cost score for node 2 receiving task B is 0.477, where task D is a temporary analysis task. Since a lower reception cost score indicates a higher willingness of the computing power node to receive computing power service tasks, the computing power side reception willingness queue for node 2 is, in order, task A, task D, and task B. The reception cost score for node 2 receiving task A is 0.296, the reception cost score for node 2 receiving task D is 0.401, and the reception cost score for node 2 receiving task B is 0.477. Figure 3 The computing power side of node 2 on the right corresponds to the willingness queue.

[0156] The improved two-layer Gael-Shapley game matching engine reads the task-side preference queue and the computing power-side acceptance willingness queue, and then begins multi-round iterative matching. In the first round of applications, task A submits a matching application to node 1 according to the task-side preference queue, and task B submits a matching application to node 2 according to the task-side preference queue. After receiving the matching application of task A, node 1 merges task A with other tasks it has already received into a set of tasks to be sorted, and sorts them according to node 1's computing power-side acceptance willingness queue. Since node 1's acceptance cost score for task C (0.314) is lower than that for task A (0.387), and the acceptance cost score for task A (0.387) is lower than that for task B (0.521), node 1 prioritizes temporarily storing tasks with higher acceptance willingness rankings and that have passed the resource capacity occupancy check, and rejects tasks with lower acceptance willingness rankings or that have not passed the resource capacity occupancy check. After receiving the matching request for task B, node 2 sorts task B together with tasks A and D that are waiting to be sorted. Since the receiving cost score of task A (0.296) is lower than the receiving cost score of task D (0.401), and the receiving cost score of task D (0.401) is lower than the receiving cost score of task B (0.477), node 2 will prioritize temporarily storing task A or task D and may reject task B.

[0157] In the second round of re-application, rejected computing power service tasks re-enter the unmatched state and submit matching applications to the next ranked computing power node according to their own task-side preference queue. If the newly submitted matching applications change the set of tasks to be ranked for a certain computing power node, the computing power node re-sorts according to the computing power side's receiving willingness queue and cancels tasks that were previously temporarily stored but ranked lower in receiving willingness. After the temporary storage and rejection in the first round of application, and the cancellation and temporary storage in the second round of re-application, the improved two-layer Gael-Shapley game matching engine continues to cyclically execute the process of unmatched computing power service tasks submitting matching applications, computing power nodes generating nodes to receive ranking results, computing power nodes performing resource capacity occupancy checks, computing power nodes temporarily storing or canceling computing power service tasks, and updating the stable matching state matrix. After 6 rounds of application, temporary storage, rejection, and cancellation, the system reaches a stable convergence state. Of the 86 computing power service tasks, 82 enter the temporary matching state, 4 tasks enter the delayed retry queue due to failure to pass the resource capacity occupancy check, and finally, a stable matching state matrix is ​​generated. Process and Figure 3 The process corresponds to "Round 1 application → temporary storage → rejection", "Round 2 reapplication → termination → temporary storage", and "Round 6 stable convergence → stable matching state matrix".

[0158] The computing power service gateway performs random perturbations and multi-version heterogeneous transformations on the topology boundaries of the stable matching state matrix, generating a total of 63 perturbation matching versions. After node capacity conflict verification, 17 versions were eliminated; after tenant isolation conflict verification, 5 versions were eliminated; after node routing path conflict verification, 9 versions were eliminated; and after resource fragmentation risk conflict verification, 11 versions were eliminated, leaving 21 feasible heterogeneous matching versions. After multi-party game equilibrium verification, 8 non-conflict mapping versions were retained, and the non-conflict mapping version with an average fit score of 0.914 and an average resource fragmentation risk value of 0.108 was selected as the candidate orchestration scheme. The traditional method has an average fit score of 0.887 and an average resource fragmentation risk value of 0.231 under the same task set.

[0159] After the candidate orchestration scheme is generated, the computing power service gateway resolves the node routing path and generates a sandbox probing instruction that only mounts a lightweight container image without loading the core business dataset. For task B, the candidate orchestration scheme originally mapped it to a target physical node with a historical deployment success rate of 0.981. During the probing deployment, the network IO throughput sampling sequence of the target physical node was 610Mbps, 890Mbps, 420Mbps, 970Mbps, 390Mbps, and 930Mbps, with a network IO throughput jitter normalization result of 0.78; the maximum latency value in the container initialization latency trajectory was 3.6 seconds, with a container initialization latency normalization result of 0.72; the highest temperature value in the CPU instantaneous burst heat trajectory was 82 degrees Celsius, with a CPU instantaneous burst heat normalization result of 0.69. After weighted aggregation according to the preset measured response weights, the measured computing power response information was 0.735, which is greater than the preset measured response threshold of 0.65. Therefore, the target physical node was marked as a high-risk node for probing deployment, and the mapping relationship between task B and the target physical node was marked as an abnormal mapping relationship.

[0160] When revising the candidate orchestration scheme, the computing power service gateway filters candidate alternative mapping relationships from the set of non-conflicting mappings. For task B, the system filters three candidate alternative mapping relationships: the first candidate alternative mapping relationship has a suitability score of 0.902, a resource fragmentation risk value of 0.104, and a measured computing power response value of 0.412; the second candidate alternative mapping relationship has a suitability score of 0.902, a resource fragmentation risk value of 0.097, and a measured computing power response value of 0.438; and the third candidate alternative mapping relationship has a suitability score of 0.891, a resource fragmentation risk value of 0.083, and a measured computing power response value of 0.392. Since the first and second candidate alternative mapping relationships have the same suitability score, the system prioritizes the second candidate alternative mapping relationship with the lower resource fragmentation risk value as the target alternative mapping relationship and replaces the original target physical node and node routing path. The system performs node capacity consistency verification, tenant isolation consistency verification, node routing path consistency verification, resource fragmentation risk consistency verification, and actual response consistency verification. The first verification found that a data cleaning task had a node routing path congestion conflict and returned an alternative mapping selection. The second verification passed all tests and the target orchestration scheme was obtained.

[0161] After the target orchestration scheme is issued, the target computing nodes execute container image pulling, business data loading, runtime environment initialization, and computing service startup according to the resource allocation results. For task A, the image pulling latency is 11.4 seconds, the business data loading latency is 8.7 seconds, the actual startup latency is 24.6 seconds, the actual response latency is 176 milliseconds, the actual service availability rate is 0.991, and the actual number of migrations is 0. During the runtime phase, the average CPU utilization is 7.4 cores, the average memory utilization is 46.1GB, the average video memory utilization is 23.2GB, the average storage I / O utilization is 398MB / s, and the average network bandwidth utilization is 169Mbps. The computing service gateway writes the data to the execution feedback information and updates the task profile, computing node response profile, and resource fragmentation risk value accordingly.

[0162] Five rounds of comparison were conducted under the same 86 computing power service requests and the same initial computing power nodes. The average deployment success rate of the traditional method was 88.6%, while that of the method of this invention was 97.4%. The average number of task migrations per task was 1.42 for the traditional method and 0.31 for the method of this invention. The average startup latency was 37.8 seconds for the traditional method and 25.9 seconds for the method of this invention. The average response latency for online inference tasks was 214 milliseconds for the traditional method and 163 milliseconds for the method of this invention. The average resource fragmentation risk value was 0.238 for the traditional method and 0.109 for the method of this invention. The proportion of unreusable fragments after scheduling was 23.7% for the traditional method and 9.4% for the method of this invention. The traditional method experienced 7 failures due to insufficient resources, 5 failures due to container initialization, 6 failures due to network I / O jitter, and 2 failures due to security policy interception, while the method of this invention experienced 1, 1, 2, and 0 failures respectively. The results show that the present invention can reduce the risk of resource fragmentation, reduce deployment failures and task migration, and improve the stability of computing service deployment in scenarios with multi-task concurrency and mixed deployment of heterogeneous computing power nodes.

[0163] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A computing power orchestration and scheduling method for a computing power service gateway, characterized in that, include: Receive computing power service requests, parse the computing power service intent information, and construct a task profile based on the computing power service intent information; Collect static capability information, dynamic load information, historical response information, data affinity information, and security domain information of multiple computing power nodes to construct a response profile of the computing power nodes; A task-computing power adaptation matrix is ​​constructed based on the task profile and the computing power node response profile, and a candidate computing power node set is determined based on the task-computing power adaptation matrix; For each candidate computing node in the candidate computing node set, sample the resource occupancy status after the LightGBM predicts the target task deployment, and obtain the resource status prediction result after scheduling. Calculate the resource fragmentation risk value based on the resource status prediction results after scheduling; The task-computing power adaptation matrix and resource fragmentation risk value are input into the improved two-layer Gale-Shapley game matching engine to generate candidate orchestration schemes. Lightweight trial deployments are then performed on the candidate orchestration schemes to obtain measured computing power response information. Based on the measured computing power response information, the candidate orchestration scheme is revised to obtain the target orchestration scheme; The target orchestration scheme is distributed to the target computing power nodes, the computing power service is deployed, the execution feedback information during the target task operation is collected, and the task profile, computing power node response profile and resource fragmentation risk value are updated based on the execution feedback information.

2. The computing power orchestration and scheduling method for a computing power service gateway according to claim 1, characterized in that, The process of receiving a computing power service request and parsing it to obtain computing power service intent information includes: Receive computing power service requests, read the task type field, resource requirement field, service quality constraint field, and data security constraint field from the computing power service request, and generate a request field set; Validate the set of request fields to obtain a valid set of request fields; Parse the resource requirement fields to extract CPU requirements, memory requirements, video memory requirements, storage input / output requirements, and network bandwidth requirements, and generate a resource requirement vector according to the resource dimensions. Parse the service quality constraint fields to extract the maximum startup latency, maximum response latency, minimum service availability, and maximum number of migrations, and generate a service quality constraint vector according to the service quality dimensions. Parse the data security constraint fields to extract the data source location identifier, data affinity level, and security domain level, and generate a data security constraint vector according to the data security dimensions. Generate computing power service intent information based on the task type field, resource requirement vector, service quality constraint vector, and data security constraint vector; The computing power service intent information is normalized, encoded, and concatenated according to the field order of the profile to generate a task profile.

3. The computing power orchestration and scheduling method for a computing power service gateway according to claim 1, characterized in that, The collection of static capability information, dynamic load information, historical response information, data affinity information, and security domain information of multiple computing nodes includes: Identify the multiple computing power nodes to be collected and generate a set of computing power nodes; Collect static capability information and dynamic load information of each computing node in the computing node set, and generate static capability vector and dynamic load vector; Calculate the residual hardware resource vector for each computing node based on the static capability vector and the dynamic load vector. Collect historical response information, data affinity information, and security domain information of each computing node in the computing node set, and generate historical response vector and node data security vector; The static capability vector, dynamic load vector, hardware residual resource vector, historical response vector, and node data security vector are normalized and encoded, and the fields are concatenated according to the order of the node profile fields to construct the computing power node response profile.

4. The computing power orchestration and scheduling method for a computing power service gateway according to claim 1, characterized in that, The construction of the task-computing power adaptation matrix based on the task profile and the computing power node response profile includes: Based on the resource demand vector and the hardware residual resource vector, calculate the resource satisfaction degree between each computing power service task and each computing power node; Based on the service quality constraint vector and the historical response vector, calculate the service quality satisfaction between each computing power service task and each computing power node; Based on the data security constraint vector and the node data security vector, calculate the data security satisfaction degree between each computing power service task and each computing power node; Based on resource satisfaction, service quality satisfaction, and data security satisfaction, calculate the compatibility score between each computing power service task and each computing power node; The adaptation scores are arranged according to the computing power service task number and computing power node number, and a task-computing power adaptation matrix is ​​constructed. Based on the task-computing power adaptation matrix, determine the set of candidate computing power nodes.

5. The computing power orchestration and scheduling method for a computing power service gateway according to claim 1, characterized in that, The step of targeting each candidate computing power node in the candidate computing power node set includes: For each candidate computing node in the candidate computing node set, read the resource demand vector in the task profile, and read the static capability vector, dynamic load vector, hardware residual resource vector and historical response vector in the computing node response profile to generate a set of predictive basic features. The dynamic load information of candidate computing power nodes before the current scheduling time is sampled within a time window to generate a dynamic load sampling sequence of candidate computing power nodes; Based on the predicted basic feature set and the dynamic load sampling sequence, LightGBM input feature samples are generated; Input the LightGBM input feature samples into the pre-trained LightGBM prediction model to predict the CPU usage, memory usage, GPU memory usage, storage input / output usage, and network bandwidth usage after the target task is deployed and scheduled. Based on the resource dimensions of CPU, memory, video memory, storage input / output, and network bandwidth, the predicted CPU usage, memory usage, video memory usage, storage input / output, and network bandwidth usage after scheduling are arranged in order to obtain the predicted resource status results after scheduling.

6. The computing power orchestration and scheduling method for a computing power service gateway according to claim 1, characterized in that, The step of calculating the resource fragmentation risk value based on the resource status prediction results after scheduling includes: Read the static capability vector and the resource status prediction results after scheduling of the candidate computing power nodes, and calculate the hardware residual resource vector of the candidate computing power nodes after scheduling after the target task is deployed; Based on the remaining hardware resource vector after scheduling, calculate the proportion of remaining resources available for candidate computing power nodes; Based on the proportion of remaining available resources under each resource dimension, calculate the resource dimension imbalance risk value of the candidate computing power node; Based on the remaining resource availability ratio and the preset minimum reusable resource ratio under each resource dimension, calculate the non-reusable fragment risk value of the candidate computing power node; Based on the resource imbalance risk value and the non-reusable fragmentation risk value, the resource fragmentation risk value of the candidate computing power node is calculated.

7. The computing power orchestration and scheduling method for a computing power service gateway according to claim 1, characterized in that, The improved two-layer Gale-Shapley game matching engine, which inputs the task-computing power adaptation matrix and resource fragmentation risk value, includes: Extract the compatibility score between each computing power service task and each computing power node in the input task-computing power adaptation matrix, arrange the compatibility scores in descending order, and generate a task-side preference queue. Extract the input resource fragmentation risk value, combine it with the hardware residual resource vector of each computing node to calculate the receiving cost score, sort the receiving cost scores in ascending order, and generate a computing power side receiving willingness queue. The task-side preference queue and the computing power-side receiving willingness queue are simultaneously input into the improved two-layer Gael-Shapley game matching engine. Through multiple rounds of iterative hedging, in which tasks submit applications according to preferences and computing power nodes temporarily store or cancel according to receiving willingness, a stable matching state matrix is ​​output. Random perturbation and multi-version heterogeneous transformation of the topological boundary are performed on the stable matching state matrix to extract the set of non-conflict mappings that satisfy the equilibrium of the multi-party game and generate candidate orchestration schemes. The node routing paths in the candidate orchestration schemes are parsed and transformed into sandbox probing commands that only mount lightweight container images without loading core business datasets. These commands are then sent to the target physical nodes to execute empty runs without data that are not included in the formal billing process. The kernel-mode resource trajectories during the execution of the sandbox probing commands are captured, and transient feature data of the probing deployment is generated. Extract network I / O throughput jitter, container initialization latency, and instantaneous CPU burst heat from the transient characteristic data of the trial deployment, perform standardized weighted aggregation, and output measured computing power response information.

8. The computing power orchestration and scheduling method for a computing power service gateway according to claim 1, characterized in that, The step of correcting the candidate orchestration scheme based on the measured computing power response information to obtain the target orchestration scheme includes: Read the candidate orchestration schemes and measured computing power response information to determine the target physical node, node routing path and measured response result for each computing power service task in the candidate orchestration schemes. Based on the measured computing power response information, a risk assessment is conducted on the trial deployment of target physical nodes in the candidate orchestration schemes. Based on the trial deployment of high-risk nodes, anomaly mapping identification is performed on candidate orchestration schemes to generate a set of mappings to be corrected. Based on the set of mappings to be corrected, alternative mappings are selected for abnormal mapping relationships in the candidate orchestration schemes to generate an alternative mapping set. The abnormal mapping relationships in the candidate orchestration schemes are replaced according to the alternative mapping set to generate the corrected orchestration scheme; Perform a pre-deployment consistency check on the revised orchestration scheme to obtain the target orchestration scheme.

9. An electronic device, characterized in that, The electronic device includes a processor and a memory, the memory storing programs or instructions that can run on the processor, the programs or instructions being executed by the processor to implement the steps of the method as described in any one of claims 1-8.

10. A readable storage medium, characterized in that, The readable storage medium stores a program or instructions that, when executed by a processor, implement the steps of the method as described in any one of claims 1-8.

Images