Algorithm warehouse management method and system based on servitization tool chain

Abstract

The embodiment of the invention provides an algorithm warehouse management method and system based on a servitization tool chain, and belongs to the technical field of algorithm warehouse management. According to the method, resource requirements are matched through task category identifiers, dynamic resource occupation prediction of parallel tasks of an execution terminal is combined, a hardware resource available parameter sequence is generated, and the matching degree is calculated based on the moment-by-moment resource fitting degree through a first screening action, so that the resource adaptation precision of an algorithm model and the execution terminal is greatly improved, and resource waste is reduced; the second screening action is combined with historical execution information of the target service terminal and the non-target service terminal, it is guaranteed that the confidence coefficient of a single task reaches the standard, execution terminal selection is adjusted through service terminal group confidence coefficient distribution, and the overall execution stability of the system is guaranteed. And the selection range of the execution terminal is dynamically adjusted based on the confidence ratio of the service terminal group, intelligent management of system confidence and computing power reserve is realized, extreme imbalance is avoided, and adjustability of long-term operation is guaranteed.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of algorithm warehouse management, in particular to an algorithm warehouse management method based on a service tool chain and an algorithm warehouse management system based on a service tool chain. BACKGROUND

[0002] In a distributed computing scenario, the periodic execution of tasks by a service terminal group needs to be implemented through a scheduling terminal to achieve efficient matching of algorithm models and execution terminals. The management efficiency of the algorithm warehouse as a core resource pool directly affects the stability of task execution and resource utilization.

[0003] In the prior art, the algorithm warehouse management method has the following deficiencies: first, the resource matching lacks precision. Most methods only match based on task type and static hardware configuration of the execution terminal, without considering the dynamic resource occupation rules of the execution terminal in parallel tasks, resulting in a deviation in the estimation of available hardware resources, a low degree of adaptation of algorithm models and execution terminals in terms of resource requirements, and a high risk of resource waste or task execution failure. Second, the overall task execution stability is not well controlled. Existing solutions focus on the execution results of a single task and do not take into account the historical execution information of the service terminal group, ignoring the impact of the confidence level distribution of target service terminals and non-target service terminals on the overall system stability, which may lead to a long-term substandard confidence level of some service terminals and a decline in the overall execution quality of the system. Third, there is a lack of dynamic adjustment mechanism. The existing screening logic is mostly fixed rules, and the execution terminal selection strategy is not dynamically adjusted according to the system confidence level, which may lead to extreme cases such as waste of computing power when the confidence level is high or insufficient computing power when the confidence level is low, and the dynamic balance between system confidence level and computing power reserve cannot be achieved, which may lead to an unadjustable deadlock in the long run.

[0004] Therefore, the existing algorithm warehouse management method cannot meet the three core requirements of resource precision matching, overall execution stability, and dynamic balance adjustment, which limits the task processing capacity and long-term running stability of the distributed system. SUMMARY

[0005] The purpose of the embodiments of the present application is to provide an algorithm warehouse management method and system based on a service tool chain to at least solve the problems of inaccurate resource matching, poor overall execution stability, and lack of dynamic adjustment mechanism in the prior art.

[0006] To achieve the above-mentioned purpose, the first aspect of the present application provides an algorithm warehouse management method based on a service tool chain, which is used for a scheduling terminal, and the method comprises: When receiving an algorithm calling request uploaded by a target service terminal in a service terminal group for a periodic execution task, extracting a task category identifier of the periodic execution task; According to the task category identifier, a task execution resource requirement of the periodic execution task is matched, and a hardware resource available parameter sequence of each execution terminal in a task execution window is generated; query first historical task execution information of a target service terminal and second historical task execution information of a non-target service terminal in the service terminal group; wherein, the first historical task execution information and the second historical task execution information include algorithm analysis results of corresponding service terminals in historical groups of periodic execution tasks, and the algorithm analysis results include algorithm execution time and algorithm analysis confidence; Based on the hardware resource available parameter sequence of each execution terminal in the task execution window, a first screening action associated with hardware resource matching and a second screening action associated with overall task execution stability determined by algorithm analysis results of the target service terminal and the non-target service terminal in the historical groups of periodic execution tasks are respectively performed in the algorithm warehouse of the service tool chain and the execution terminal group, to generate an optimal combination scheme of the current algorithm model and the execution terminal for the periodic execution task; According to the optimal combination scheme, algorithm model scheduling instructions from the algorithm warehouse to the target execution terminal and task data scheduling instructions from the service terminal to the execution terminal are generated to complete task distribution and execution.

[0007] Optionally, when receiving an algorithm calling request uploaded by a target service terminal in a service terminal group for a current periodic execution task, a task category identifier of the periodic execution task is extracted, including: receiving an algorithm calling request uploaded by a target service terminal in a service terminal group for a current periodic execution task, parsing and extracting a task category field as an initial category identifier; querying a pre-stored standard task category library in the algorithm warehouse of the service tool chain; wherein, the standard task category library contains all supported task category identifiers and corresponding coding rules; matching the extracted initial category identifier with the standard task category library, and if the matching is successful, the corresponding standard code is taken as the task category identifier of the periodic execution task.

[0008] Optionally, according to the task category identifier, a task execution resource requirement of the periodic execution task is matched, and a hardware resource available parameter sequence of each execution terminal in a task execution window is generated, including: According to the task category identifier, a task execution resource requirement of the periodic execution task is matched; wherein, the task execution resource requirement includes task execution window requirement and hardware resource occupation parameter requirement; querying a parallel task execution list of a plurality of execution terminals in the execution terminal group, and according to execution period and resource occupation information of each parallel task in the parallel task execution list, generating a hardware resource available parameter sequence of each execution terminal in a task execution window.

[0009] Optionally, the task execution resource requirement of the matching period is executed by using the task category identifier, including: calling a pre-stored task category and resource requirement mapping table in an algorithm warehouse of a service tool chain; wherein each record is associated with a task category identifier, a corresponding standard task execution window range and a hardware resource occupation parameter interval; querying the standard task execution window range in the task category and resource requirement mapping table based on the task category identifier, combining the real-time index corresponding to the task category identifier, taking the intersection of the standard task execution window range and the real-time index to determine the task execution window requirement; based on the task category identifier, matching the corresponding hardware resource occupation parameter interval in the task category and resource requirement mapping table, taking the interval value of each hardware resource to determine the hardware resource occupation parameter requirement.

[0010] Optionally, the parallel task execution list of a plurality of execution terminals in the execution terminal group is queried, and based on the execution period and resource occupation information of each parallel task in the parallel task execution list, a hardware resource available parameter sequence of each execution terminal in the task execution window is generated, including: sending a state query instruction to all online execution terminals in the execution terminal group to obtain the parallel task execution list returned by each execution terminal; wherein the parallel task execution list contains the execution period and resource occupation information of each parallel task; for each execution terminal, extracting the parallel tasks in the parallel task execution list that have time overlap with the task execution window of the current period execution task to form an overlapping task set; for each task in the overlapping task set, based on the historical resource occupation trend, the resource occupation sequence in the task execution window is predicted by fitting the resource occupation curve of the last three similar tasks; obtaining the hardware configuration parameters of each execution terminal, for each resource type, calculating the available resource parameter at each time in the task execution window; wherein the available resource parameter is the maximum available value of the corresponding resource type minus the total predicted occupation value of the overlapping task set to the resource at that time; arranging all available resource parameters of each resource type in the task execution window in time sequence to form the hardware resource available parameter sequence of the execution terminal.

[0011] Optionally, the first historical task execution information of the target service terminal and the second historical task execution information of the non-target service terminal in the service terminal group are queried, including: access a historical task database; wherein the historical task database stores execution records of periodic execution tasks of all service terminals in the service terminal group in the past several periods, and each execution record contains an algorithm execution time and an algorithm analysis confidence level; query all periodic execution task execution records of the target service terminal in the past several periods to form first historical task execution information, and query all periodic execution task execution records of all non-target service terminals in the service terminal group except the target service terminal in the past several periods, and group them according to the service terminals to form second historical task execution information.

[0012] Optionally, the first screening action includes: query all algorithm models and corresponding confidence levels matching the task category identifier of the current periodic execution task from the algorithm repository of the service tool chain; wherein each confidence level is associated with a unique resource requirement sequence; extract the resource requirement sequence of each confidence level of each algorithm model; for each execution terminal, call the hardware resource available parameter sequence to calculate the resource matching degree; wherein the resource matching degree is obtained by multiplying the inverse of the task execution window length by the sum of the fitting degrees of the resource requirement and the available resource parameter at each time point in the window, and the fitting degree at each time point is the smaller value divided by the larger value of the resource requirement and the available resource parameter at the time point; based on a preset matching degree threshold, filter out algorithm model confidence levels and execution terminal combinations with a resource matching degree greater than or equal to the threshold to form a first screening result set.

[0013] Optionally, the second screening action includes: based on a preset confidence level threshold, query the algorithm analysis confidence levels of all periodic execution tasks of the target service terminal in the first historical task execution information, calculate the confidence level average of the algorithm analysis confidence levels of all periodic execution tasks and the confidence level determined by each combination in the first screening result set, filter out combinations with a confidence level average greater than or equal to the confidence level threshold to form an intermediate screening set; based on a preset confidence level reference value, query the periodic execution tasks of each non-target service terminal in the second historical task execution information, calculate the historical confidence level average of each non-target service terminal, count the number of positive terminals with a historical confidence level average greater than zero and the number of negative terminals with a historical confidence level average less than zero, calculate the ratio of the number of positive terminals to the number of negative terminals plus 1, collect the current total hardware resource occupation rate of all execution terminals, sort them in ascending order to form an execution terminal sorting table, determine the selection position according to the ratio by a preset rule, select execution terminals from the execution terminal sorting table according to the selection position to form an execution terminal candidate set; The preset rule is that when the ratio of the number of positive terminals to the number of negative terminals is greater than 2, the execution terminals in the last 30% of the execution terminal sorting table are selected, when the ratio of the number of positive terminals to the number of negative terminals is less than 1, the execution terminals in the first 30% of the execution terminal sorting table are selected, and when the ratio of the number of positive terminals to the number of negative terminals is greater than or equal to 1 and less than or equal to 2, the execution terminals in the middle 40% of the execution terminal sorting table are selected. The combination of the execution terminal not in the execution terminal candidate set is removed from the middle screening set, and any one combination is selected as the optimal combination scheme of the periodic execution task and the execution terminal of the current algorithm model.

[0014] Optionally, according to the optimal combination scheme, the algorithm model scheduling instruction from the algorithm warehouse to the target execution terminal and the task data scheduling instruction from the service terminal to the execution terminal are generated, and the task distribution and execution are completed, including: According to the target execution terminal and the algorithm model confidence version in the optimal combination scheme, the algorithm model scheduling instruction from the algorithm warehouse to the target execution terminal and the task data scheduling instruction from the service terminal to the execution terminal are generated; Based on the algorithm model scheduling instruction, the algorithm model confidence version corresponding to the algorithm warehouse is sent to the corresponding target execution terminal, based on the task data scheduling instruction, the corresponding task data in the algorithm calling request is sent to the target execution terminal, so that the target execution terminal executes the task data processing of the periodic execution task according to the algorithm model confidence version.

[0015] The second aspect of the application provides an algorithm warehouse management system based on a service tool chain, which comprises: The extraction module is configured to extract the task category identifier of the periodic execution task when receiving the algorithm calling request uploaded by the target service terminal in the service terminal group for the periodic execution task; The generation module is configured to match the task execution resource demand of the periodic execution task by using the task category identifier, and generate a sequence of hardware resource availability parameters of each execution terminal in a task execution window; The query module is configured to query the first historical task execution information of the target service terminal and the second historical task execution information of the non-target service terminal in the service terminal group; wherein the first historical task execution information and the second historical task execution information comprise algorithm analysis results of corresponding service terminals in historical groups of periodic execution tasks, and the algorithm analysis results comprise algorithm execution time and algorithm analysis confidence; The screening module is used for performing a first screening action associated with hardware resource matching and a second screening action associated with overall task execution stability determined by algorithm analysis results of the target service terminal and the non-target service terminal in historical groups of periodic execution tasks based on a hardware resource available parameter sequence of each execution terminal in a task execution window, and generating an optimal combination scheme of the current algorithm model and the execution terminal for the periodic execution task. The execution module is used for generating algorithm model scheduling instructions of the algorithm warehouse to the target execution terminal and task data scheduling instructions of the service terminal to the execution terminal according to the optimal combination scheme, and completing task distribution and execution.

[0016] Through the above technical solution, the present application generates a hardware resource available parameter sequence by matching resource requirements through task category identification and combining dynamic resource occupation prediction of parallel tasks of the execution terminal, the first screening action calculates the matching degree based on resource fitting degree at each time point, greatly improves the resource adaptation accuracy of the algorithm model and the execution terminal, and reduces resource waste. On this basis, the second screening action combines historical execution information of the target service terminal and the non-target service terminal, not only ensures that the confidence of a single task meets the standard, but also adjusts the selection of the execution terminal through the confidence distribution of the service terminal group, and ensures the overall execution stability of the system. Based on the service terminal group confidence ratio, the selection range of the execution terminal is dynamically adjusted, the intelligent management of the system confidence and the computing power reserve is realized, extreme imbalance is avoided, and the adjustability of long-term operation is ensured. Finally, the problems of inaccurate resource matching, poor overall execution stability and lack of dynamic adjustment mechanism in the prior art are solved.

[0017] Other features and advantages of the present application will be described in detail in the following detailed description. BRIEF DESCRIPTION OF DRAWINGS

[0018] The accompanying drawings are included to provide a further understanding of the present application and constitute a part of the specification, and are used together with the following detailed description to explain the present application, but do not constitute a limitation on the present application. In the drawings: Figure 1 is a step flowchart of an algorithm warehouse management method based on a service-oriented tool chain provided by an embodiment of the present application; Figure 2 is a system structure diagram of an algorithm warehouse management system based on a service-oriented tool chain provided by an embodiment of the present application. DETAILED DESCRIPTION

[0019] The specific embodiments of the present application will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are only used to illustrate and explain the present application, and are not used to limit the present application.

[0020] As Figure 1 shown, the embodiment of the application provides a service-based tool chain algorithm warehouse management method, which comprises the following steps: Step S1: When receiving the algorithm call request of the periodic execution task uploaded by the target service terminal in the service terminal group, the task category identifier of the periodic execution task is extracted.

[0021] Specifically, the target service terminal generates an algorithm call request according to the business demand of the periodic execution task and uploads it to the scheduling terminal. The algorithm call request contains task basic information, including task ID, task category field, real-time index, data volume and other core parameters, wherein the task category field is used to clearly define the business type of the task (such as data detection, signal analysis, etc.).

[0022] After the scheduling terminal receives the algorithm call request, it performs structured analysis on the request data through a preset analysis rule, and extracts the task category field as the initial category identifier. The initial category identifier adopts a standardized string format to ensure compatibility with the subsequent standard task category library.

[0023] The scheduling terminal queries the pre-stored standard task category library in the algorithm warehouse of the service-based tool chain. The standard task category library is a structured database constructed offline, which contains all task category identifiers supported by the system, corresponding 6-digit coding rules (the first 2 digits are task category codes, and the last 4 digits are task subcategory codes) and adaptive algorithm model types, and is updated regularly according to business expansion.

[0024] The scheduling terminal accurately matches the extracted initial category identifier with the records in the standard task category library. If the matching is successful, the corresponding 6-digit code is directly used as the task category identifier of the periodic execution task; if the matching fails, the scheduling terminal returns a matching failure prompt to the target service terminal, requiring the re-upload of a compliant task category field, while starting a fuzzy matching algorithm to determine the closest standard task category identifier based on string similarity (such as edit distance), and labeling the to-be-verified label to the scheduling terminal background for manual review by the management personnel to ensure the accuracy of the task category identifier. Step S2: Using the task category identifier, matching the task execution resource demand of the periodic execution task to generate a sequence of hardware resource availability parameters of each execution terminal within the task execution window.

[0025] Specifically, step S2 comprises the following steps: Step S21: Using the task category identifier, matching the task execution resource demand of the periodic execution task; wherein the task execution resource demand includes task execution window demand and hardware resource occupation parameter demand.

[0026] Specifically, the scheduling terminal calls the pre-stored task category and resource demand mapping table in the algorithm warehouse of the service tool chain. The mapping table is structured data generated based on historical task execution data offline training, and each record is associated with a task category identifier, a corresponding standard task execution window range, a CPU occupancy rate interval, a memory occupancy interval, a GPU memory occupancy interval, and a bandwidth occupancy rate interval, ensuring the pertinence of resource demand matching.

[0027] Based on the task category identifier, the mapping table is queried to obtain the corresponding standard task execution window range (such as 10-30 seconds). At the same time, the real-time indicator (i.e. the maximum allowed delay of the task) carried in the algorithm call request is extracted, and the intersection of the standard task execution window range and the real-time indicator is determined to determine the task execution window demand , wherein is the current timestamp plus 2 seconds of scheduling preparation time, is the window length, ensuring that the task execution window meets both the regular execution demand of the task type and the real-time constraint.

[0028] Based on the task category identifier matching the corresponding hardware resource occupancy parameter interval in the mapping table, the middle value of each hardware resource (CPU, memory, GPU, and bandwidth) interval is taken as the initial resource occupancy parameter. For example, when the CPU occupancy rate interval is 30%-50%, the initial value is 40%; when the memory occupancy interval is 2-4 GB, the initial value is 3 GB, and finally the hardware resource occupancy parameter demand of the periodic execution task is formed, providing a benchmark for subsequent hardware resource availability calculation.

[0029] Step S22: Query the parallel task execution list of a plurality of execution terminals in the execution terminal group, and generate a hardware resource available parameter sequence of each execution terminal in the task execution window according to the execution period and resource occupancy information of each parallel task in the parallel task execution list. Specifically, the scheduling terminal sends a state query instruction to all online execution terminals in the execution terminal group. The instruction uses a standardized communication protocol (such as HTTP / 2) and includes parameters such as query timestamp and data format requirements, requiring the execution terminal to return a response within 1 second.

[0030] After each execution terminal receives the state query instruction, it collects the current parallel task execution list and returns it to the scheduling terminal. The parallel task execution list includes the task ID, execution start timestamp, estimated end timestamp, task category identifier, and real-time resource occupancy information (instantaneous occupancy rate of CPU, memory, GPU, and bandwidth, with a sampling frequency of 1 time / second) of each parallel task, fully reflecting the current load state of the execution terminal.

[0031] The scheduling terminal analyzes the parallel task execution list of each execution terminal and extracts the task execution window Parallel tasks with overlapping execution times form an overlapping task set. The criterion for determining time overlap is: the execution start timestamp of the parallel tasks ≤ And the expected end timestamp is ≥ This ensures that no parallel tasks that might affect the execution of the target task are overlooked.

[0032] For each task in the overlapping task set, the scheduling terminal queries the historical task database for the three most recent execution records of the task's category, extracting the corresponding resource usage curves (time-resource utilization relationship). A multinomial fitting algorithm is used to fit the three curves, generating a resource usage prediction model for the task in the current execution scenario. Based on this model, the resource usage within the task execution window is predicted. Resource usage sequence within ( , (Memory, GPU, bandwidth), prediction accuracy error is controlled within ±5%.

[0033] The scheduling terminal obtains the hardware configuration parameters of each execution terminal, including the number of CPU cores, total memory, total GPU memory, and maximum bandwidth, and calculates the maximum available value of each resource. (For example, maximum available CPU value = number of cores × 100%, maximum available memory value = total memory).

[0034] For each resource type In the task execution window Every moment within Calculate the available resource parameters at that moment: in, For overlapping task sets at time Resources The total predicted occupancy rate. If the calculated result is less than zero, then... Set it to 0 to ensure the reasonableness of available resource parameters.

[0035] All available resource parameters for each resource type within the task execution window are arranged in chronological order to form a sequence of available hardware resource parameters for the execution terminal. This sequence is stored as an array, with the array length being the window duration multiplied by the sampling frequency (sampling frequency is 1 time / second). For example, when the window duration is 20 seconds, the array length is 20. Each element corresponds to an available resource parameter at a given moment, providing accurate data support for subsequent resource matching.

[0036] Step S3: querying the first historical task execution information of the target service terminal and the second historical task execution information of the non-target service terminal in the service terminal group; wherein the first historical task execution information and the second historical task execution information include algorithm analysis results of corresponding service terminals in historical several groups of period execution tasks, and the algorithm analysis results include algorithm execution time and algorithm analysis confidence.

[0037] Specifically, the scheduling terminal accesses the historical task database associated with itself. The database adopts a distributed storage architecture and stores past period execution task execution records of all service terminals in the service terminal group, each execution record containing a service terminal ID, a task ID, a task category identifier, an execution timestamp, an algorithm model ID, an algorithm version (confidence version), an algorithm execution time, an algorithm analysis confidence (accurate to three decimal places), and an execution state (success / failure), and the data retention period is consistent with the preset 5, which can be configured according to business needs.

[0038] The scheduling terminal queries all period execution task execution records of the target service terminal in the past periods based on the target service terminal ID and the task category identifier determined in step S1, and forms the first historical task execution information. The information is arranged in descending order of execution timestamp, facilitating quick extraction of recent execution data; if the number of records is less than , all valid records (records with execution state as success) are taken.

[0039] The scheduling terminal queries all period execution task execution records of all non-target service terminals in the service terminal group in the past periods, groups the records by service terminal ID, and arranges the records in each group in descending order of execution timestamp to form the second historical task execution information. The information is used to reflect the overall execution quality distribution of the service terminal group. The scheduling terminal performs data cleaning on the first historical task execution information and the second historical task execution information. The records with execution state as failure, the abnormal records with algorithm analysis confidence as 0, and the invalid records with algorithm execution time exceeding twice the corresponding task execution window are removed to obtain the cleaned valid historical task execution information, ensuring the reliability of the subsequent analysis data.

[0040] ​Step S4: Based on the hardware resource available parameter sequence of each execution terminal within the task execution window, the first screening action associated with hardware resource matching and the second screening action associated with the overall task execution stability determined by the algorithm analysis results of the target service terminal and the non-target service terminal in the historical several groups of periodic execution tasks are respectively performed in the algorithm warehouse of the service tool chain and the execution terminal group, and the optimal combination scheme of the current algorithm model and the execution terminal for the periodic execution task is generated.

[0041] Specifically, step S4 includes: S41: query all algorithm models and corresponding confidence versions matching the task category identifier of the current periodic execution task from the algorithm warehouse of the service tool chain; wherein each confidence version is associated with a unique resource requirement sequence; S42: for each confidence version of each algorithm model, extract its resource requirement sequence; S43: for each execution terminal, call the hardware resource available parameter sequence to calculate the resource matching degree; wherein the resource matching degree is obtained based on the inverse of the task execution window length multiplied by the sum of the fitting degrees of the resource requirement and the available resource parameter at each time within the window, and the fitting degree at each time is the smaller value divided by the larger value of the resource requirement and the available resource parameter at that time; S44: based on the preset matching degree threshold, filter out the algorithm model confidence version and the execution terminal combination with the resource matching degree greater than or equal to the threshold to form a first screening result set.

[0042] Specifically, the scheduling terminal queries all algorithm models and corresponding confidence versions matching the task category identifier of the current periodic execution task from the algorithm warehouse of the service tool chain. Each algorithm model in the algorithm warehouse is divided into multiple versions according to confidence (such as high confidence version, balanced version, and lightweight version), different versions are associated with a unique resource requirement sequence (CPU, memory, GPU, bandwidth requirement rate sequence within the task execution window, which is consistent with the time dimension of the hardware resource available parameter sequence, with a sampling frequency of 1 time / second), and the higher the confidence version, the higher the resource requirement sequence value.

[0043] For each confidence version of each algorithm model, the scheduling terminal extracts its resource requirement sequence , which reflects the real-time demand of the algorithm version on each hardware resource during execution.

[0044] For each execution terminal, the scheduling terminal calls its hardware resource available parameter sequence generated in step 2 to calculate the resource matching degree. The calculation rule of the resource matching degree is: ​​ in, for from arrive The summation operation, this formula quantifies the degree of fit between resource demand and available resources by weighting the average of their time-by-time matching. The range of values ​​is The closer the value is to 1, the higher the compatibility.

[0045] The scheduling terminal has a pre-set allocation threshold (default is 0.7, which can be dynamically adjusted according to the system resource shortage) to filter out... The algorithm model confidence level, version, and execution terminal are combined to form the first screening result set. This result set includes the combination ID, algorithm model ID, version confidence level, execution terminal ID, and matching score, providing a basic combination pool for subsequent stability screening.

[0046] S45: Based on a preset confidence threshold, query the algorithm analysis confidence of all periodically executed tasks of the target service terminal in the first historical task execution information, calculate the average confidence of the algorithm analysis confidence of all periodically executed tasks and the confidence of each combination in the first screening result set, and screen out the combinations whose average confidence is greater than or equal to the confidence threshold to form an intermediate screening set; S46: Based on the preset confidence reference value, query the periodic execution tasks of each non-target service terminal in the second historical task execution information, calculate the historical average confidence value of each non-target service terminal, count the number of positive terminals with a difference greater than zero between the historical average confidence value and the confidence reference value and the number of negative terminals with a difference less than zero, calculate the ratio of the number of positive terminals to the number of negative terminals plus 1, collect the current total hardware resource utilization rate of all execution terminals, sort them in ascending order to form an execution terminal sorting table, determine the selection position according to the ratio and preset rules, select execution terminals from the execution terminal sorting table according to the selection position, and form an execution terminal candidate set; The preset rule is as follows: when the ratio of the number of positive terminals to the number of negative terminals is greater than 2, the last 30% of the execution terminals in the execution terminal sorting table are selected; when the ratio of the number of positive terminals to the number of negative terminals is less than 1, the first 30% of the execution terminals in the execution terminal sorting table are selected; when the ratio of the number of positive terminals to the number of negative terminals is greater than or equal to 1 and less than or equal to 2, the middle 40% of the execution terminals in the execution terminal sorting table are selected. S47: Eliminate combinations from the intermediate filtering set that are not in the candidate set of execution terminals, and select any combination as the optimal combination of the periodic execution task with the current algorithm model and execution terminal.

[0047] Specifically, the second screening action is performed in three steps, and the core goal is to screen out the combination scheme that is most conducive to the stability of the overall task execution: First, generate the intermediate screening set: The scheduling terminal pre-sets the confidence threshold (preset to 0.9, which can be adjusted according to the requirements of the business on the execution quality). Query the algorithm analysis confidence of all valid records in the first historical task execution information, and calculate the historical confidence mean: ; Wherein, is the number of valid records in the first historical task execution information, is the algorithm analysis confidence of the th record.

[0048] For each algorithm model confidence version-execution terminal combination in the first screening result set, calculate the combination confidence mean: Wherein, is the confidence of the algorithm model version in the combination. Screen out combinations to form the intermediate screening set, ensuring that the task execution result confidence of the target service terminal meets the standard.

[0049] Second, determine the execution terminal candidate set: The scheduling terminal pre-sets the confidence reference value (preset to 0.85, used to measure the historical execution quality of the service terminal). Query the valid records of each non-target service terminal in the second historical task execution information, and calculate the historical confidence mean of each non-target service terminal: Wherein, is the number of valid records of the service terminal, is the algorithm analysis confidence of the th record of the service terminal.

[0050] Statistical non-target service terminal that meets positive terminal number , and the negative terminal number that meets , calculate the ratio: The denominator plus 1 avoids the case where the divisor is 0 when . The ratio reflects the overall confidence distribution of the service terminal group, the larger the better the overall execution quality.

[0051] Collect all the current hardware resource total occupancy of the execution terminal (the average of the total real-time resource occupancy of all parallel tasks), and sort them from small to large to form an execution terminal sorting table. In the sorting table, the front sequence terminal indicates that the hardware resource occupancy is low, the remaining available resources are sufficient, and the adaptability to high confidence algorithm version is strong; the later sequence terminal indicates that the hardware resource occupancy is high, the remaining available resources are few, and the adaptability to high confidence algorithm version is weak.

[0052] According to the ratio Determine the execution terminal selection range according to the preset rule: When , the overall confidence of the service terminal group is high, the execution terminal sorting table is selected for the last 30% of the execution terminal, and the execution terminal candidate set is formed. At this time, the terminal with high resource occupancy is selected, which can activate the idle computing power and increase the system computing power reserve. At the same time, because the overall confidence of the system has redundancy, it will not affect the task execution quality, and the high confidence → algorithm reserve adjustment is realized.

[0053] When , the overall confidence of the service terminal group is low, the execution terminal sorting table is selected for the first 30% of the execution terminal, and the execution terminal candidate set is formed. At this time, the high-quality terminal with sufficient resources is selected to preferentially guarantee the task execution confidence, quickly improve the overall confidence of the system, and realize the low confidence → confidence adjustment.

[0054] When , the overall confidence of the service terminal group is in a balanced state, the middle 40% of the execution terminal (31%-70%) in the execution terminal sorting table is selected to form the execution terminal candidate set, and the dynamic balance of confidence and algorithm reserve is maintained to avoid the system deviating to the extreme direction.

[0055] Third, generate the optimal combination scheme: The scheduling terminal eliminates the combinations of the execution terminal not in the execution terminal candidate set from the middle selection set to obtain a second selection result set. The result set retains the combination ID, algorithm model ID, version confidence, execution terminal ID, matching degree score and Value.

[0056] Select the optimal combination scheme from the second selection result set. The selection rule can be: the matching degree score in descending order is the first priority, The value in descending order is the second priority; if there is a tie, select the combination with the highest historical task success rate (success execution times / total allocation times) of the execution terminal, finally determine the optimal algorithm model ID, optimal confidence version and target execution terminal ID, and form the optimal combination scheme for the current period execution task.

[0057] Step S5: Based on the optimal combination scheme, generate algorithm model scheduling instructions from the algorithm repository to the target execution terminal and task data scheduling instructions from the service terminal to the execution terminal, and complete task distribution and execution.

[0058] Specifically, based on the optimal combination scheme, the scheduling terminal extracts core parameters such as the target execution terminal ID, the optimal algorithm model ID, and the optimal confidence version, and generates two types of scheduling instructions: Firstly, the algorithm model scheduling instructions: These use a JSON structure and include the instruction type (algorithm scheduling), target execution terminal ID, optimal algorithm model ID, optimal confidence version, algorithm storage path (a unique access address within the algorithm repository), transmission priority (set to high / medium / low based on task real-time metrics), checksum (generated based on the hash value of the algorithm model file, used for integrity verification), and instruction validity period. (The first 30 seconds) ensure the accuracy and timeliness of the algorithm model transmission.

[0059] Secondly, the task data scheduling instructions adopt a JSON structure, including the instruction type (data scheduling), target execution terminal ID, target service terminal ID, task ID, task data storage path (task data access address of the target service terminal), and task execution window. Hardware resource usage parameters and result feedback address (the address where the scheduling terminal receives the execution results) are specified to ensure the matching between task data and algorithm model.

[0060] The scheduling terminal sends corresponding scheduling instructions to the algorithm repository and the target service terminal through the communication module of the service-oriented toolchain. Simultaneously, an instruction transmission confirmation mechanism is activated, with a preset timeout of 5 seconds. If no confirmation response is received from the receiver within the timeout period, the instruction is resent, with a maximum of 3 retries to ensure reliable instruction transmission.

[0061] After receiving the algorithm model scheduling instruction, the algorithm repository first verifies the validity of the checksum and the instruction's expiration date. If the verification passes, it transmits the optimal algorithm model and its corresponding version (including model weights, inference configuration, etc.) to the target execution terminal via a high-speed transmission channel (such as FTP protocol). Upon completion of the transmission, the algorithm repository sends an algorithm-ready notification to the scheduling terminal, informing it that the algorithm model has been successfully deployed to the target execution terminal.

[0062] After receiving the task data scheduling instruction, the target service terminal packages the task data according to the format specified in the instruction, ensuring that the data format is consistent with the input requirements of the optimal algorithm model. After packaging, the task data is transmitted to the target execution terminal through a secure transmission channel, and a data ready notification is sent to the scheduling terminal after the transmission is complete.

[0063] The target execution terminal starts the optimal algorithm model, loads the task data and performs the calculation after receiving both the algorithm ready and data ready notifications During the execution, the target execution terminal reports the resource occupation status (real-time occupation rate of CPU, memory, GPU and bandwidth) and the execution progress (completed data volume / total data volume) to the scheduling terminal at a preset frequency (1 time / 5 seconds); after the execution is completed, the algorithm analysis result (including algorithm execution time, algorithm analysis confidence and task processing result) is fed back to the scheduling terminal, the scheduling terminal synchronizes the result to the historical task database, and the distribution and execution of the current period execution task are completed.

[0064] As shown in Figure 2 The embodiment of the present application provides an algorithm warehouse management system based on a service tool chain, which comprises: An extraction module is configured to extract a task category identifier of a period execution task when receiving an algorithm calling request of the period execution task uploaded by a target service terminal in a service terminal group; A generation module is configured to match a task execution resource requirement of the period execution task by using the task category identifier, and generate a hardware resource available parameter sequence of each execution terminal in a task execution window; A query module is configured to query first historical task execution information of the target service terminal and second historical task execution information of non-target service terminals in the service terminal group; wherein the first historical task execution information and the second historical task execution information comprise algorithm analysis results of corresponding service terminals in historical period execution tasks, and the algorithm analysis result comprises algorithm execution time and algorithm analysis confidence; A screening module is configured to perform a first screening action associated with hardware resource matching and a second screening action associated with overall task execution stability determined by algorithm analysis results of the target service terminal and the non-target service terminals in the historical period execution tasks in the algorithm warehouse of the service tool chain and the execution terminal group based on the hardware resource available parameter sequence of each execution terminal in the task execution window, and generate an optimal combination scheme of the algorithm model and the execution terminal for the current period execution task; An execution module is configured to generate algorithm model scheduling instructions from the algorithm warehouse to the target execution terminal and task data scheduling instructions from the service terminal to the execution terminal according to the optimal combination scheme, and complete the task distribution and execution.

[0065] Those skilled in the art can understand that all or part of the steps of the method for implementing the above-mentioned embodiments can be completed by programs instructing relevant hardware, the programs are stored in a storage medium, and the programs include a plurality of instructions for enabling a single-chip microcomputer, a chip or a processor to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and various storage media capable of storing program codes.

[0066] The optional embodiments of the present application are described in detail above in combination with the drawings, but the embodiments of the present application are not limited to the specific details in the above-described embodiments. Within the technical concept range of the embodiments of the present application, various simple modifications can be made to the technical solutions of the embodiments of the present application, and these simple modifications all belong to the protection scope of the embodiments of the present application. In addition, it should be noted that each specific technical feature described in the above-described specific embodiments can be combined in any appropriate manner without contradiction. In order to avoid unnecessary repetition, the embodiments of the present application will not be described again for various possible combinations.

[0067] In addition, various different embodiments of the present application can also be combined in any manner, as long as it does not deviate from the idea of the embodiments of the present application, and it should also be considered as disclosed by the embodiments of the present application.

Claims

1. A service-based tool chain based algorithm warehouse management method, characterized in that, A method for scheduling terminals, the method comprising: Upon receiving an algorithm invocation request for a periodic execution task at a current algorithm call request from a target service terminal in a service terminal group, extracting a task category identifier of the periodic execution task; Using the task category identifier, matching a task execution resource requirement of the periodic execution task to generate a hardware resource availability parameter sequence of each execution terminal within a task execution window; Querying first historical task execution information of the target service terminal and second historical task execution information of non-target service terminals in the service terminal group; wherein the first historical task execution information and the second historical task execution information include algorithm analysis results of corresponding service terminals in historical groups of periodic execution tasks, and the algorithm analysis results include algorithm execution time and algorithm analysis confidence; Based on the hardware resource availability parameter sequence of each execution terminal within the task execution window, performing a first screening action associated with hardware resource matching and a second screening action associated with overall task execution stability determined by algorithm analysis results of the target service terminal and the non-target service terminals in the historical groups of periodic execution tasks in the algorithm repository of the service tool chain and the execution terminal group respectively, to generate an optimal combination scheme of the algorithm model and the execution terminal for the periodic execution task at the current algorithm call request; According to the optimal combination scheme, generating an algorithm model scheduling instruction from the algorithm repository to the target execution terminal and a task data scheduling instruction from the service terminal to the execution terminal to complete task distribution and execution.

2. The service-based tool chain based algorithm warehouse management method according to claim 1, characterized in that, Upon receiving an algorithm invocation request for a periodic execution task at a current algorithm call request from a target service terminal in a service terminal group, extracting a task category identifier of the periodic execution task, comprising: Receiving an algorithm invocation request for a periodic execution task at a current algorithm call request from a target service terminal in a service terminal group, parsing and extracting a task category field as an initial category identifier; Querying a pre-stored standard task category library in an algorithm repository of a service tool chain; wherein the standard task category library contains all supported task category identifiers and corresponding coding rules; Matching the extracted initial category identifier with the standard task category library, and if the matching is successful, using the corresponding standard code as the task category identifier of the periodic execution task.

3. The service-based tool chain based algorithm warehouse management method according to claim 1, characterized in that, Using the task category identifier, matching a task execution resource requirement of the periodic execution task to generate a hardware resource availability parameter sequence of each execution terminal within a task execution window, comprising: Using the task category identifier, matching a task execution resource requirement of the periodic execution task; wherein the task execution resource requirement includes task execution window requirement and hardware resource occupation parameter requirement; Querying a parallel task execution list of a plurality of execution terminals in the execution terminal group, and generating a hardware resource availability parameter sequence of each execution terminal within a task execution window according to execution period and resource occupation information of each parallel task in the parallel task execution list.

4. The service-based tool chain based algorithm warehouse management method according to claim 3, characterized in that, Using the task category identifier, matching a task execution resource requirement of the periodic execution task, comprising: Call the pre-stored task category and resource demand mapping table in the algorithm warehouse of the service tool chain; wherein each record is associated with a task category identifier, a corresponding standard task execution window range and a hardware resource occupation parameter interval; Based on the task category identifier, the standard task execution window range in the task category and resource demand mapping table is queried, and the intersection of the standard task execution window range and the real-time index corresponding to the task category identifier is taken to determine the task execution window demand; Based on the task category identifier, the corresponding hardware resource occupation parameter interval in the task category and resource demand mapping table is matched, and the middle value of each hardware resource interval is taken to determine the hardware resource occupation parameter demand.

5. The service-based tool chain based algorithm warehouse management method according to claim 3, characterized in that, Query the parallel task execution list of a plurality of execution terminals in the execution terminal group, and generate the hardware resource available parameter sequence of each execution terminal in the task execution window according to the execution period and resource occupation information of each parallel task in the parallel task execution list, including: Send a state query instruction to all online execution terminals in the execution terminal group, and obtain the parallel task execution list returned by each execution terminal; wherein the parallel task execution list contains the execution period and resource occupation information of each parallel task; For each execution terminal, extract the parallel tasks in the parallel task execution list that have time overlap with the task execution window of the current period execution task to form an overlapping task set; For each task in the overlapping task set, based on the historical resource occupation trend, the resource occupation sequence in the task execution window is predicted through the fitting of the resource occupation curves of the last three similar tasks; Obtain the hardware configuration parameters of each execution terminal, and for each resource type, calculate the available resource parameter at each time in the task execution window; wherein the available resource parameter is the maximum available value of the corresponding resource type minus the total predicted occupation value of the overlapping task set at that time for that resource; Arrange all available resource parameters of each resource type in the task execution window in chronological order to form the hardware resource available parameter sequence of the execution terminal.

6. The service-based tool chain based algorithm warehouse management method according to claim 1, characterized in that, Query the first historical task execution information of the target service terminal and the second historical task execution information of the non-target service terminal in the service terminal group, including: Access the historical task database; wherein the historical task database stores the periodic execution task execution records of all service terminals in the service terminal group in the past several periods, and each execution record contains algorithm execution time and algorithm analysis confidence; Query all periodic execution task execution records of the target service terminal in the past several periods to form the first historical task execution information, and query all periodic execution task execution records of all non-target service terminals other than the target service terminal in the service terminal group in the past several periods, and group them by service terminal to form the second historical task execution information.

7. The service-based tool chain based algorithm warehouse management method according to claim 1, characterized in that, The first screening action includes: Query all algorithm models and corresponding confidence versions matching the task category identifier of the current period execution task from the algorithm warehouse of the service tool chain; wherein each confidence version is associated with a unique resource demand sequence; For each confidence version of each algorithm model, extract its resource demand sequence; For each execution terminal, the hardware resource available parameter sequence is called to calculate the resource matching degree; wherein, the resource matching degree is obtained based on the reciprocal of the task execution window length multiplied by the sum of the fitting degrees of the resource demand and the available resource parameters at each time in the window, and the fitting degree at each time is the smaller value divided by the larger value of the resource demand and the available resource parameter at the time; Based on the preset matching degree threshold, the algorithm model confidence version and the execution terminal combination with the resource matching degree greater than or equal to the threshold are screened out to form a first screening result set.

8. The service-based tool chain based algorithm warehouse management method according to claim 7, characterized in that, The second screening action includes: Based on the preset confidence threshold, the algorithm analysis confidence of all periodic execution tasks of the target service terminal in the first historical task execution information is queried, the confidence average of the algorithm analysis confidence of all periodic execution tasks and the confidence corresponding to each combination in the first screening result set is calculated, the combination with the confidence average greater than or equal to the confidence threshold is screened out to form an intermediate screening set; Based on the preset confidence reference value, the periodic execution tasks of each non-target service terminal in the second historical task execution information are queried, the historical confidence average of each non-target service terminal is calculated, the number of positive terminals with the difference between the historical confidence average and the confidence reference value greater than zero and the number of negative terminals with the difference less than zero are counted, the ratio of the number of positive terminals and the number of negative terminals plus 1 is calculated, the current total hardware resource occupation rate of all execution terminals is collected, the execution terminal sorting table is formed by sorting from small to large, the selection position is determined according to the ratio by a preset rule, the execution terminal is selected from the execution terminal sorting table according to the selection position, and the execution terminal candidate set is formed; The preset rule is: when the ratio of the number of positive terminals and the number of negative terminals is greater than 2, the last 30% of the execution terminal sorting table is selected, when the ratio of the number of positive terminals and the number of negative terminals is less than 1, the first 30% of the execution terminal sorting table is selected, and when the ratio of the number of positive terminals and the number of negative terminals is greater than or equal to 1 and less than or equal to 2, the middle 40% of the execution terminal sorting table is selected. The combination in which the execution terminal is not in the execution terminal candidate set is removed from the intermediate screening set, and any one combination is selected as the optimal combination scheme of the periodic execution task in the current algorithm model and the execution terminal.

9. The service-based tool chain based algorithm warehouse management method according to claim 1, characterized in that, According to the optimal combination scheme, the algorithm model scheduling instruction from the algorithm warehouse to the target execution terminal and the task data scheduling instruction from the service terminal to the execution terminal are generated to complete task distribution and execution, including: According to the target execution terminal and the algorithm model confidence version in the optimal combination scheme, the algorithm model scheduling instruction from the algorithm warehouse to the target execution terminal and the task data scheduling instruction from the service terminal to the execution terminal are generated; Based on the algorithm model scheduling instruction, the algorithm model confidence version corresponding to the algorithm warehouse is sent to the corresponding target execution terminal, based on the task data scheduling instruction, the corresponding task data in the algorithm calling request is sent to the target execution terminal, so that the target execution terminal executes the task data processing of the periodic execution task according to the algorithm model confidence version.

10. A service-based tool chain based algorithm warehouse management system, characterized in that, The system includes: An extraction module is configured to extract a task category identifier of a periodic execution task when a current algorithm calling request of the periodic execution task uploaded by a target service terminal in a service terminal group is received; A generation module is configured to generate a sequence of hardware resource availability parameters of each execution terminal within a task execution window by matching a task execution resource requirement of the periodic execution task with the task category identifier; A query module is configured to query first historical task execution information of the target service terminal and second historical task execution information of non-target service terminals in the service terminal group; wherein the first historical task execution information and the second historical task execution information include algorithm analysis results of corresponding service terminals in historical groups of periodic execution tasks, and the algorithm analysis results include algorithm execution time and algorithm analysis confidence; A screening module is configured to perform a first screening action associated with hardware resource matching and a second screening action associated with overall task execution stability determined by algorithm analysis results of the target service terminal and the non-target service terminals in the historical groups of periodic execution tasks in the algorithm repository and the execution terminal group based on the sequence of hardware resource availability parameters of each execution terminal within the task execution window, and generate an optimal combination scheme of the algorithm model and the execution terminal for the periodic execution task at present; An execution module is configured to generate algorithm model scheduling instructions from the algorithm repository to the target execution terminal and task data scheduling instructions from the service terminal to the execution terminal according to the optimal combination scheme, and complete task distribution and execution.

Images