A service node selection method based on computing power state

Abstract

The application discloses a service node selection method based on computing power state, relates to the technical field of computing power scheduling, and specifically comprises the following steps: receiving a service request and determining a candidate node to obtain the resource state, task occupation state and state generation time of the candidate node, converting heterogeneous computing resources into equivalent computing power based on the performance mapping relationship corresponding to the service type, forming the equivalent computing power time sequence within the response time limit according to the computing power change event, performing trusted correction on the equivalent computing power corresponding to the resource release event, forming the trusted computing power time sequence, determining the demand computing power and the predicted processing time length according to the processing demand, identifying the serviceable time period from the trusted computing power time sequence and forming the time domain matching result, determining the node scheduling priority and the target node according to the time domain matching result, generating a resource reservation record and sending a scheduling instruction, and the application can reduce the overestimation of computing power, improve the reliability of service node selection and resource reservation.

Description

Technical Field

[0001] This invention belongs to the field of computing power scheduling technology, specifically relating to a service node selection method based on computing power status. Background Technology

[0002] With the development of cloud computing, edge computing, and distributed service systems, the same service request can typically be handled by multiple service nodes. Different nodes have varying configurations of processors, graphics processors, storage access resources, and data transfer resources, and nodes may also be running or reserving other tasks simultaneously. Current service node selection methods often rely on resource availability, load rate, or historical average processing speed at the time of scheduling to sort nodes and select those with lower current load or shorter estimated completion times.

[0003] Service requests often involve multiple processing stages with sequential, parallel, or data-dependent relationships, each relying on different resource types and consuming varying amounts of resources. Simply comparing individual resource reserves or summarizing resource metrics fails to reflect the resource constraints limiting the processing flow's execution speed, and also makes it difficult to compare nodes with different resource components on a uniform scale. Furthermore, when data migration occurs between processing stages, the stage transition latency caused by the transmission process further impacts the actual completion time.

[0004] Node resource status and task occupancy status are typically collected, generated, and reported by different components, which may lead to inconsistencies in generation time, communication delays, fluctuations in reporting intervals, or missing status data. Existing methods directly treat received status data as the current status without fully considering the lag between status generation time and the current scheduling time, making it easy to make selections based on distorted resource status. Tasks that are currently executing or have been reserved may also cause changes in available resources at their expected start time, expected completion time, or the end time of resource reservation; relying solely on current availability cannot reflect the changes in computing power over time within the response timeframe. Especially when the actual completion time of a task is later than the expected completion time, the resources expected to be released may not have been released, leading to an overestimation of computing power in subsequent periods.

[0005] Some scheduling methods use a fixed weight to comprehensively consider completion time, load, and resource availability, or simply select the node with the earliest expected completion time. However, they fail to evaluate the stability of the matching results in response to changes in start time, resource release delays, and decreased computing power during bottleneck processing. After resource reservation is completed, if a critical resource release event deviates, the fixed resource reservation record is difficult to adjust in a timely manner, which may lead to response time limits being exceeded, duplicate scheduling, resource conflicts, or resource fragmentation.

[0006] Therefore, there is an urgent need for a service node selection method based on computing power status to solve the problems existing in the current technology. Summary of the Invention

[0007] The purpose of this invention is to provide a service node selection method based on computing power status, which solves the technical problems in the prior art such as difficulty in uniformly evaluating heterogeneous computing resources, lag and fluctuation in node status data, difficulty in accurately reflecting subsequent computing power changes caused by pending tasks, resulting in overestimation of available computing power of nodes, inaccurate identification of serviceable time periods, and low reliability of service node selection and resource reservation.

[0008] To achieve the above objectives, the present invention adopts the following technical solution: A service node selection method based on computing power status includes: S1. Receive service requests, parse the service type, processing requirements and response time limit, and determine candidate nodes from the nodes with corresponding service capabilities; S2. Obtain the resource status, task occupancy status and status generation time of candidate nodes at multiple sampling times. Based on the performance mapping relationship corresponding to the service type, convert heterogeneous computing resources into equivalent computing power. Determine the computing power change event according to the expected start time, expected completion time and resource reservation end time of the task to be processed, and form the equivalent computing power time sequence within the response time limit. S3. Determine the baseline update interval based on historical state data. Determine the state reliability based on the state generation time, baseline update interval, and fluctuation dispersion of equivalent computing power. Divide resource release events into confirmed and unconfirmed events. Fully include the marginal equivalent computing power increment of confirmed events and reduce the marginal equivalent computing power increment of unconfirmed events according to state reliability to obtain the reliable computing power time series. Determine the required computing power and expected processing time based on processing needs. Identify serviceable time periods from the reliable computing power time series and generate time-domain matching results. S4. Determine the node scheduling priority and target node based on the time-domain matching results, generate resource reservation records, and send scheduling instructions.

[0009] Furthermore, initial candidate nodes are selected based on the necessary processing components and necessary resource types corresponding to the service type; the scheduling status of the initial candidate nodes is obtained, and nodes with communication abnormalities, unavailable resource status, unavailable necessary processing components, or prohibited from receiving new tasks are excluded to obtain candidate nodes.

[0010] Furthermore, the performance mapping relationship includes processing stages, resource dependencies, and resource consumption per unit task; the initial resource consumption per unit task is obtained through benchmark task calibration, and is corrected according to the input data scale, data complexity, and processing component adaptation, with the upper bound of resource consumption that meets the preset coverage ratio being taken as the resource consumption per unit task.

[0011] Specifically, the minimum value of the ratio between the available resources of each processing stage and the resource consumption of the corresponding unit task is determined as the stage processing rate; the stage processing time is determined based on the stage baseline workload and the stage processing rate; the baseline processing time is determined by the maximum value of the cumulative stage processing time of each execution path; and the ratio of the baseline workload to the baseline processing time is determined as the equivalent computing power.

[0012] Furthermore, the expected start time is defined as the resource occupancy increase event, and the expected completion time or the end time of resource reservation is defined as the resource release event. The time period is divided by the current scheduling time, the expected occurrence time of each computing power change event and the response time limit. The available resources in each time period are cumulatively updated and converted into equivalent computing power to form the equivalent computing power time sequence.

[0013] Furthermore, outliers are removed from the generation time intervals of historical adjacent state data, and the median of the remaining generation time intervals is determined as the baseline update interval. The earlier of the resource state generation time and the task occupation state generation time is taken as the effective state generation time. When the actual duration exceeds the baseline update interval, the excess is progressively amplified according to the number of complete baseline update intervals crossed to obtain the state lag.

[0014] Specifically, the fluctuation dispersion is determined by the weighted absolute deviation of the equivalent computing power relative to the median level at multiple sampling times, and the benchmark dispersion is determined by historical data with the same service type, similar processing requirements, and stable execution status. The state reliability is obtained by multiplying the ratio of the benchmark update interval to the sum of the benchmark update interval and the state lag by the ratio of the benchmark dispersion to the sum of the benchmark dispersion and the fluctuation dispersion.

[0015] Furthermore, at the expected time of resource release events, the task running status, resource occupancy status, and processing component feedback are obtained. Confirmed events and unconfirmed events are distinguished based on task completion information and changes in resource availability. The basic equivalent computing power is calculated based on the actual resource status confirmed before the start of the time-series unit. The marginal equivalent computing power increment of confirmed events is fully included, and the marginal equivalent computing power increment of unconfirmed events is multiplied by the status credibility and included to obtain the reliable equivalent computing power.

[0016] Specifically, the required computing power and expected processing time are determined based on the processing needs; the time interval in the trusted computing power time series where the trusted equivalent computing power is continuously not lower than the required computing power and the duration is not shorter than the expected processing time is determined as the serviceable time period, and the corresponding time domain matching results are generated.

[0017] Specifically, the node scheduling priority is determined based on the expected completion time, status reliability, and the minimum margin of reliable equivalent computing power relative to the required computing power within the serviceable period. The target node is selected according to the node scheduling priority, and if the node scheduling priorities are the same, the candidate node with the earlier expected completion time is selected. A resource reservation record is generated based on the processing stage, deployment time interval, resource type, and resource occupancy corresponding to the target node, and a scheduling instruction containing the target node, the execution order of the processing stage, and the resource reservation time boundary is sent.

[0018] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are: 1. This invention converts the heterogeneous computing resources of candidate nodes into a unified equivalent computing power through the performance mapping relationship corresponding to service types. It also generates computing power change events based on the expected start time, expected completion time, and resource reservation end time of the task to be processed, and constructs the equivalent computing power time sequence within the response time limit. This can uniformly evaluate different candidate nodes composed of resources and reflect the subsequent computing power changes caused by registered tasks, reducing the judgment bias caused by selecting service nodes based solely on the current resource status. 2. This invention determines the reliability of a state by the lag between the state generation time and the current scheduling time and the fluctuation dispersion of the equivalent computing power time sequence. Based on the reliability of the state, it delays and corrects unconfirmed resource release-related computing power change events to obtain a reliable computing power time sequence. This can reduce the overestimation of computing power caused by state data lag, task delays, and failure to release resources on schedule, and improve the accuracy and reliability of the computing power status evaluation of candidate nodes. 3. This invention converts processing requirements into a demand computing power template and performs time-domain matching with trusted computing power time series. It determines node scheduling priority by comprehensively considering response margin, state stability, and computing power surplus. At the same time, it adjusts or regenerates resource reservation records based on the actual occurrence deviation of bottleneck events. This can improve the stability of target node and serviceable time period selection and reduce processing timeouts, resource reservation conflicts, and scheduling failures. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 The diagram illustrates the steps of a service node selection method based on computing power status according to the present invention. Figure 2 The diagram illustrates the steps of the method for forming the equivalent computing power timing sequence of the present invention. Figure 3The diagram illustrates the steps of the method for determining the trusted computing power timing and available service periods according to the present invention. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] like Figure 1 The service node selection method based on computing power status shown includes the following steps: S1. Receive service requests, parse the service type, processing requirements and response time limit, and determine candidate nodes from the nodes with corresponding service capabilities; Upon receiving a service request, the system performs field identification to obtain the service type, processing requirements, and response timeframe. The service type characterizes the business processing category corresponding to the service request and establishes a correspondence with pre-configured service processing information. Service processing information includes the processing stages, processing components, resource types, and execution dependencies between each processing stage required to complete the corresponding service.

[0023] The processing requirements characterize the task scale and processing constraints of this service request, including the amount of data to be processed, input data characteristics, output requirements, and execution methods for each processing stage. The execution method for each processing stage indicates whether multiple processing stages are executed serially, in parallel, or in a combination of both. The response time limit defines the allowed time range from receiving the service request to generating the processing result, and serves as the time boundary for subsequently constructing the equivalent computing power sequence and identifying serviceable periods.

[0024] Query the node capability information corresponding to each node based on the service type. Node capability information includes the service types supported by the node, the configured processing components, the types of computing resources that can be provided, and the node scheduling status. Match the necessary processing components and necessary resource types corresponding to the service type with the processing components and computing resource types of each node.

[0025] Record the service type as ,node The set of supported service types is denoted as Service type The corresponding sets of necessary processing components and sets of necessary resource types are denoted as follows: and ,Festival The set of configured processing components and the set of resource types that can be provided are respectively denoted as and The initial set of candidate nodes is represented as:

[0026] in, This represents the initial set of candidate nodes that have the corresponding service processing capabilities.

[0027] Obtain the scheduling status of the initial candidate nodes and determine if the node is in a state where it can receive new tasks. If a node has communication problems, its resource status cannot be obtained, its necessary processing components are unavailable, or it is in a state where it is prohibited from receiving new tasks, the corresponding node will be excluded from the initial candidate node set. The scheduling can be denoted by the identifier as Finally, the candidate node set is obtained:

[0028] If the service request also includes data storage location, processing component version, data access permission, or node region restrictions, the corresponding execution constraints will continue to be matched with the node capability information of the initial candidate nodes, and nodes that do not meet the execution constraints will be excluded.

[0029] S2. Obtain the resource status, task occupancy status and status generation time of candidate nodes at multiple sampling times. Based on the performance mapping relationship corresponding to the service type, convert heterogeneous computing resources into equivalent computing power. Determine the computing power change event according to the expected start time, expected completion time and resource reservation end time of the task to be processed, and form the equivalent computing power time sequence within the response time limit. For each identified candidate node, the resource status, task occupancy status, and status generation time corresponding to multiple sampling times are obtained. The resource status includes the total amount of various computing resources, the amount occupied, the amount reserved, and the amount available; the task occupancy status includes the running status of registered tasks, resource occupancy type, resource occupancy amount, estimated start time, estimated completion time, and resource reservation end time; the status generation time is the actual time when the resource status and task occupancy status are formed, used to determine the degree of lag of the status data relative to the current scheduling time.

[0030] The state data is sorted according to the sampling time. The set of state data with complete fields and the sampling time closest to the current scheduling time is used as the initial state of the equivalent computing power time series. The state data from the remaining sampling times are retained to reflect the changes in the resource status of candidate nodes over time. When the resource status and task occupancy status of the same candidate node have different state generation times, the earlier state generation time is used as the effective generation time of the corresponding state data to avoid prematurely including resource changes that have not yet been reflected in the task occupancy status into the available resources.

[0031] Based on the performance mapping relationship corresponding to the service type, obtain the resource types and resource consumption per unit task required for each processing stage. Then, determine the candidate nodes at time t. The availability of resource type k is represented as follows: The available amount is obtained by subtracting the occupied and reserved amounts from the total amount of the corresponding resources. The unit task resource consumption of processing stage j for resource type k is expressed as... The resource consumption per unit task is determined through benchmark calibration and request characteristic correction. For each service type, the complete processing flow is broken down into independent processing stages. Benchmark task samples covering different data sizes, data structures, and processing complexities are selected. While ensuring that other resource supplies are not constrained, the available resources to be calibrated are varied. The actual resource usage, effective processing workload, and stage execution time for each benchmark task sample within its corresponding processing stage are recorded. After deducting the fixed resource overhead incurred by processing component initialization, data loading, and result writing back, the initial resource consumption per unit task is formed by the ratio of effective resource usage to effective processing workload. Based on the differences between the input data characteristics of the current service request and the benchmark task samples, data size correction, data complexity correction, and processing component adaptation correction are determined, and the initial resource consumption per unit task is adjusted accordingly. For calibration results with abnormal fluctuations in the same type of samples, the average value is not directly used; instead, the upper bound of resource consumption that meets the preset coverage ratio is selected as the final value. This allows the parameter to reflect both the actual dependence of the processing stage on resource types and to reduce the possibility of overestimating computing power due to occasional low-load samples.

[0032] The set of resource types that processing stage j depends on is represented as The task processing rate that each dependent resource can support is determined by the ratio of available resources to resource consumption per unit task, and the lowest task processing rate is determined as the stage processing rate. :

[0033] The stage processing rate is used to characterize the actual processing capacity of a candidate node in executing the corresponding processing stage under the current resource conditions. By using the minimum processing rate, the resources that limit the execution capacity of a processing stage can be identified as constrained resources, avoiding the evaluation of the service processing capacity of a candidate node based solely on a single computing resource.

[0034] Further obtain the stage baseline workload from the performance mapping relationship. Baseline workload and execution path set The baseline workload represents the standardized workload that the baseline task needs to complete in processing stage j; the baseline task size represents the uniform task size used to evaluate the processing capabilities of different candidate nodes; the execution path set is determined based on the data dependencies between each processing stage, and each execution path includes processing stages that are executed consecutively in the order of dependencies.

[0035] The ratio of the baseline workload to the processing rate of a stage is the processing time of the corresponding processing stage. The processing times of each processing stage within the same execution path are accumulated, and the largest accumulated result across all execution paths is determined as the baseline task processing time. This balances the time accumulation relationship of serial processing stages with the longest completion time of parallel processing branches. The equivalent computing power of a candidate node at time t... Represented as:

[0036] This transforms processing resources, acceleration resources, storage access resources, and data transmission resources into a unified task processing capability oriented towards the current service type, enabling candidate nodes with different resource compositions to be evaluated using the same equivalent computing power unit.

[0037] The computing power change events for each candidate node are determined based on the task occupancy status. For tasks that have completed resource reservation but have not yet started, the expected start time is determined as the resource occupancy increase event; for tasks that are currently executing, the expected completion time is determined as the resource release event; for tasks that have not yet started and have a reserved validity period set, the resource reservation end time is determined as the resource release event. For the same resource occupancy record, the expected completion time or the resource reservation end time is selected as the resource release time based on the task status to avoid the same resource being released repeatedly.

[0038] Candidate nodes The corresponding set of computing power change events is represented as Let any event in the set be denoted as e, and let the expected time of event e be denoted as . Its value is derived from the expected start time, expected completion time, or resource reservation end time in the task's occupancy status. The change in the availability of resource type k caused by event e is expressed as... Among them, resource release events correspond to an increase in available resources, while resource occupancy increase events correspond to a decrease in available resources.

[0039] The computing power change events are sorted according to their expected occurrence time. Using the current scheduling time, the expected occurrence time of each computing power change event, and the response time limit as time boundaries, the response time limit is divided into multiple variable-length time periods. The current scheduling time is represented as... Let the start time of the m-th time interval be denoted as The availability of resource type k during this time period is represented as The available quantity is determined based on the initial available quantity corresponding to the current scheduling time, and the cumulative resource changes that have occurred before the start of the time period:

[0040] Wherein, the formula satisfies The available resources for each time period are substituted into the calculation process of stage processing rate and equivalent computing power to obtain the equivalent computing power for each time period. By associating each time period, its corresponding equivalent computing power, and the computing power change events that caused the resource changes in chronological order, an equivalent computing power time sequence within the response time limit is formed.

[0041] Each time series unit in the equivalent computing power time series includes the start time of the time period, the end time of the time period, the equivalent computing power, the availability of each resource type, and the associated computing power change events. This reflects the current resource status of candidate nodes and the computing power changes caused by the start, completion, or termination of resource reservations for registered tasks within the response time limit, providing a state data foundation for subsequent status reliability determination, equivalent computing power time series correction, and service period matching.

[0042] S3. Determine the baseline update interval based on historical state data. Determine the state reliability based on the state generation time, baseline update interval, and fluctuation dispersion of equivalent computing power. Divide resource release events into confirmed and unconfirmed events. Fully include the marginal equivalent computing power increment of confirmed events and reduce the marginal equivalent computing power increment of unconfirmed events according to state reliability to obtain the reliable computing power time series. Determine the required computing power and expected processing time based on processing needs. Identify serviceable time periods from the reliable computing power time series and generate time-domain matching results. Get candidate nodes Baseline update interval The baseline update interval is determined based on the generation time interval of adjacent state data successfully received by the candidate node within a preset historical range. Within the preset historical range, candidate nodes are updated according to their state generation time. The status data is sorted, retaining only those with complete fields, continuously increasing generation times, and successfully received by the scheduler. The generation time interval between each pair of adjacent status data is calculated sequentially to form a candidate update interval sequence. These candidate update intervals are arranged in ascending order to determine the median of all candidate update intervals. The deviations of the candidate update intervals at both ends of the sequence from this median are compared; the end with the larger deviation is deleted each time until the difference between the maximum and minimum values ​​of the remaining candidate update intervals is no greater than the median. This eliminates excessively long intervals caused by communication interruptions, excessively short intervals caused by status re-reporting, and abnormal intervals formed during node maintenance. The median of the remaining candidate update intervals is then used as the candidate node. Baseline update interval .

[0043] Obtain the state generation time of each candidate node and the equivalent computing power data corresponding to multiple sampling times. Then, select the candidate nodes... The effective state generation time is expressed as The effective state generation time is the earlier of the resource state generation time and the task occupancy state generation time, based on the current scheduling time. With effective state generation time The difference is used to determine the actual duration of the state data, based on the candidate nodes. Baseline update interval As a normal update tolerance: when the actual duration does not exceed At that time, it is considered that the status data is still within the normal reporting cycle, and the status lag is considered to be... Set to zero; when the actual duration exceeds When the excess is determined, the base lag is determined, and then progressively amplified based on the number of complete baseline update intervals included in the actual duration to determine the state lag. The specific formula is:

[0044] in, It indicates rounding down and does not penalize the waiting time for normal status reporting. When a slight overdue occurs, the status lag increases with the overdue time. When a new status is not obtained after crossing multiple baseline update intervals, it reflects the risk of status distortion caused by multiple possible missed reports through progressive amplification. It is used to distinguish between normal reporting jitter, single delay and continuous missed reports.

[0045] The fluctuation dispersion is determined based on the equivalent computing power data corresponding to multiple sampling times. First, the median level of the equivalent computing power data is determined, then the absolute deviation of each sampled data relative to the median level is calculated, and weighted and normalized according to the proximity of the sampling time to the current scheduling time.

[0046] Get Service Type Corresponding benchmark dispersion The reference dispersion is adopted in conjunction with the fluctuation dispersion. The same calculation method is used, based on historical equivalent computing power data of the same service type, similar processing requirements, and stable execution status. Candidate nodes are selected. State credibility is represented as The calculation method is as follows:

[0047] Among these, a smaller state lag indicates that the state data is closer to the current scheduling time; a smaller fluctuation dispersion indicates that the equivalent computing power obtained from continuous sampling is more stable. State reliability is constrained by both state timeliness and computing power stability, which can avoid evaluating candidate nodes based solely on a single state data sample.

[0048] Set of events related to changes in computing power The resource release event is confirmed. When event e is expected to occur... Upon arrival, the system retrieves the running status, resource usage status, and feedback information from the processing components for the corresponding task. If the task is completed and the increase in available resources of the corresponding resource type matches the resource change value for this event, the event is designated as a confirmed resource release event. If no task completion information is obtained, no corresponding resource change is detected, or different status information contradicts each other, the event is designated as an unconfirmed resource release event.

[0049] Let the set of confirmed resource release events that have reached their expected occurrence time within the m-th time period be represented as: The corresponding set of unconfirmed resource release events is represented as Both sets mentioned above are sets of computing power change events. A subset of the event, and dynamically updated based on the event confirmation results.

[0050] The basic equivalent computing power for the m-th time period is represented as: The basic equivalent computing power is determined based on the actual resource status confirmed before the start of the time period and is included in the resource occupation increase event, but the available resource increment generated by the resource release event is not included for the time being.

[0051] Let the marginal equivalent computing power increment generated by resource release event e in the m-th time period be expressed as: The marginal equivalent computing power increment is determined by sequentially adding the resource availability changes of the corresponding events according to their expected occurrence times, and then recalculating using the aforementioned stage processing rate and equivalent computing power conversion relationship. This item-by-item recalculation method avoids the duplication of computing power increments when multiple events release the same constrained resources.

[0052] The trusted equivalent computing power corresponding to the m-th time period Represented as:

[0053] The marginal equivalent computing power increment corresponding to confirmed resource release events is fully included, while the marginal equivalent computing power increment corresponding to unconfirmed resource release events is reduced according to the state credibility. After an unconfirmed event is confirmed, it is adjusted to the set of confirmed resource release events; if it is determined that the corresponding resource has not been released, its marginal equivalent computing power increment is canceled.

[0054] The trusted equivalent computing power is arranged in chronological order according to time periods, and associated with the state credibility, the confirmation status of resource release events, and the corresponding correction results to form a trusted computing power time sequence. This trusted computing power time sequence can reduce the overestimation of computing power caused by state lag, task delays, and resource release failures while preserving the service capabilities formed by potential resource releases.

[0055] Based on the task scale, input data characteristics, and execution method of each processing stage in the processing requirements, obtain the baseline workload and processing rate for each stage corresponding to the service type, and determine the estimated processing time for each stage. For sequentially executed processing stages, sum the estimated processing times of each stage; for parallelly executed processing stages, if resources can support concurrent execution, take the maximum estimated processing time of each parallel branch; if resources cannot support concurrent execution, adjust the execution order of the processing stages according to resource constraints. The maximum estimated processing time of each adjusted execution path is then determined as the estimated processing time. The standardized task volume corresponding to this service request. The ratio of the required computing power to the estimated processing time is determined as the computing power required. .

[0056] The current scheduling time and the moment when the trusted equivalent computing power changes in the trusted computing power time series are determined as candidate start times. For any candidate start time s, the time interval is determined. The entire time series is covered, and the minimum trusted equivalent computing power of each time series is determined as the minimum trusted equivalent computing power for the time interval. When the minimum trusted equivalent computing power is not lower than the required computing power, and The time interval is determined as the serviceable period no later than the end time of the response time limit.

[0057] The start time of the serviceable period is determined as the expected start time; the sum of the expected start time and the expected processing time is determined as the expected completion time; the difference between the response deadline and the expected completion time is determined as the response margin; and the difference between the minimum trusted equivalent computing power and the required computing power is determined as the minimum computing power reserve. The time-domain matching results include the serviceable period, expected start time, expected completion time, response margin, minimum trusted equivalent computing power, minimum computing power reserve, and associated computing power change events.

[0058] S4. Determine the node scheduling priority and target node based on the time-domain matching results, generate resource reservation records, and send scheduling instructions.

[0059] Obtain the available service time periods and corresponding time-domain matching results for each candidate node. For candidate nodes that do not have available service time periods, exclude them from the node selection range for this time; when all candidate nodes do not have available service time periods, generate a scheduling result indicating that the service request cannot meet the response time limit for the time being, and do not generate a resource reservation record.

[0060] When a candidate node has multiple serviceable time slots, the start time of each serviceable time slot is determined as the expected start time, and the corresponding expected completion time is obtained based on the determined expected processing time. The serviceable time slot with the earlier expected completion time is preferentially selected as the target serviceable time slot for that candidate node; when the expected completion times are the same, the serviceable time slot with the larger minimum margin of trusted equivalent computing power relative to the required computing power within its serviceable time slot is selected.

[0061] Candidate nodes The target available service period is represented as ,in, For the expected start time, The estimated completion time is determined by: acquiring all trusted computing power time-series units covered by the target's serviceable time period; determining the minimum trusted equivalent computing power of each time-series unit as the minimum trusted equivalent computing power; and determining the difference between the minimum trusted equivalent computing power and the required computing power as the minimum computing power margin. The difference between the response deadline termination time and the estimated completion time is determined as the response margin.

[0062] The node scheduling priority is determined based on the expected completion time, state reliability, and minimum computing power reserve of each candidate node. Candidate nodes are sorted by expected completion time (earliest to latest), state reliability (highest to lowest), and minimum computing power reserve (largest to smallest), resulting in a completion time sorting number, a state reliability sorting number, and a computing power reserve sorting number. The sum of these three sorting numbers is determined as the node scheduling ranking value. The smaller the node scheduling ranking value, the higher the node scheduling priority.

[0063] Candidate nodes are sorted according to their scheduling priority, and the candidate node with the highest scheduling priority is determined as the target node. If multiple candidate nodes have the same scheduling priority, the candidate node with the earlier expected completion time is determined as the target node. If the expected completion times are still the same, the candidate node with the higher state reliability is determined as the target node.

[0064] After determining the target node, obtain the service processing information determined by S1 and the performance mapping relationship established by S2, and obtain the processing stage corresponding to this service request, the execution dependency relationship between each processing stage, the resource type dependent on each processing stage, and the unit task resource consumption after correction based on the characteristics of the current service request.

[0065] Based on the execution dependencies between processing stages, the deployment time interval for each processing stage is determined within the target serviceable time period. For processing stages executed sequentially, the end time of the previous processing stage is taken as the earliest start time of the next processing stage. For processing stages executed in parallel, the corresponding processing stage is set to be deployed in parallel when all its preceding processing stages are completed and the corresponding resources of the target node can support simultaneous execution. When the corresponding resources cannot support simultaneous execution, the start time of the relevant processing stages is sequentially postponed according to the execution dependencies of the processing stages, ensuring that the adjusted expected completion time does not exceed the response time limit termination time.

[0066] Based on the task size, input data characteristics, processing stage execution methods, and performance mapping relationships in the processing requirements, the resource types and resource consumption amounts required by each processing stage within the corresponding deployment time interval are determined. The resource consumption amounts use the same resource metering method as in S2 when determining the resource consumption per unit task, and are adjusted based on the task size and input data characteristics of the current service request relative to the baseline task.

[0067] For multiple processing stages that are parallel and whose deployment time intervals overlap, the resource usage of the same resource type required by them is summed. For processing stages whose deployment time intervals do not overlap, the same resource type is reserved according to its respective deployment time interval. The resource usage corresponding to each processing stage is compared with the resource availability of the target node in the corresponding time period. The target node is determined to meet the resource reservation condition only if the resource usage of each resource type does not exceed the availability of the corresponding time period, and the trusted equivalent computing power of that time period is not lower than the required computing power.

[0068] When the target node does not meet the resource reservation conditions, the current target node is abandoned, and the next candidate node is selected according to the node scheduling priority. The target service time period and corresponding resource reservation conditions are re-determined until a target node that meets the resource reservation conditions is obtained or all candidate nodes have completed the verification.

[0069] When the target node meets the resource reservation conditions, a resource reservation record is generated based on the target node identifier, service request identifier, processing stage, deployment time interval, resource type, and resource usage. The resource reservation record includes at least the service request identifier, target node identifier, processing stage identifier, resource type, resource usage, resource reservation start time, resource reservation end time, and reservation status.

[0070] For resources that have been written into the resource reservation record, the corresponding resource usage will be included in the reservation amount of the target node, so that when subsequent service requests form an equivalent computing power sequence, the resource can be identified as having been reserved, thus avoiding the duplicate allocation of the same resource within the same time interval.

[0071] After the resource reservation record is generated, a scheduling instruction is sent to the target node. The scheduling instruction includes a service request identifier, a target node identifier, the execution order of processing stages, the resource type and resource occupancy corresponding to each processing stage, the resource reservation time boundary, and the response time limit termination time. Upon receiving the scheduling instruction, the target node executes the service request according to the execution order of the processing stages and the corresponding resource reservation time boundary, and updates the resource occupancy status based on the task execution status, providing status data for forming a new equivalent computing power sequence in subsequent scheduling cycles.

[0072] 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.

[0073] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific embodiments described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A service node selection method based on computing power status, characterized in that, include: S1. Receive service requests, parse the service type, processing requirements and response time limit, and determine candidate nodes from the nodes with corresponding service capabilities; S2. Obtain the resource status, task occupancy status and status generation time of candidate nodes at multiple sampling times. Based on the performance mapping relationship corresponding to the service type, convert heterogeneous computing resources into equivalent computing power. Determine the computing power change event according to the expected start time, expected completion time and resource reservation end time of the task to be processed, and form the equivalent computing power time sequence within the response time limit. S3. Determine the baseline update interval based on historical state data. Determine the state reliability based on the state generation time, baseline update interval, and fluctuation dispersion of equivalent computing power. Divide resource release events into confirmed and unconfirmed events. Fully include the marginal equivalent computing power increment of confirmed events and reduce the marginal equivalent computing power increment of unconfirmed events according to state reliability to obtain the reliable computing power time series. Determine the required computing power and expected processing time based on processing needs. Identify serviceable time periods from the reliable computing power time series and generate time-domain matching results. S4. Determine the node scheduling priority and target node based on the time-domain matching results, generate resource reservation records, and send scheduling instructions.

2. The service node selection method based on computing power status according to claim 1, characterized in that, Initial candidate nodes are selected based on the necessary processing components and necessary resource types corresponding to the service type; the scheduling status of the initial candidate nodes is obtained, and nodes with communication abnormalities, unavailable resource status, unavailable necessary processing components, or prohibited from receiving new tasks are excluded to obtain candidate nodes.

3. The service node selection method based on computing power status according to claim 1, characterized in that, The performance mapping relationship includes processing stages, resource dependencies, and resource consumption per unit task. The initial resource consumption per unit task is obtained through benchmark task calibration and is corrected according to the input data scale, data complexity, and processing component adaptation. The upper bound of resource consumption that meets the preset coverage ratio is taken as the resource consumption per unit task.

4. The service node selection method based on computing power status according to claim 3, characterized in that, The minimum value among the ratios of the available resources for each processing stage to the resource consumption per unit task is determined as the stage processing rate. The stage processing time is determined based on the stage baseline workload and stage processing rate. The baseline processing time is determined by the maximum cumulative stage processing time of each execution path. The ratio of the baseline workload to the baseline processing time is determined as the equivalent computing power.

5. The service node selection method based on computing power status according to claim 1, characterized in that, The expected start time is defined as the resource occupancy increase event, and the expected completion time or the end time of resource reservation is defined as the resource release event. The time period is divided by the current scheduling time, the expected occurrence time of each computing power change event and the response time limit. The available resources in each time period are cumulatively updated and converted into equivalent computing power to form the equivalent computing power time sequence.

6. The service node selection method based on computing power status according to claim 1, characterized in that, Outliers are removed from the generation time intervals of historical adjacent state data, and the median of the remaining generation time intervals is determined as the baseline update interval. The earlier of the resource state generation time and the task occupation state generation time is taken as the effective state generation time. When the actual duration exceeds the baseline update interval, the excess is progressively amplified according to the number of complete baseline update intervals crossed to obtain the state lag.

7. The service node selection method based on computing power status according to claim 6, characterized in that, The fluctuation dispersion is determined by the weighted absolute deviation of the equivalent computing power relative to the median level at multiple sampling times. The baseline dispersion is determined by historical data with the same service type, similar processing requirements, and stable execution status. The state reliability is obtained by multiplying the ratio of the baseline update interval to the sum of the baseline update interval and the state lag by the ratio of the baseline dispersion to the sum of the baseline dispersion and the fluctuation dispersion.

8. The service node selection method based on computing power status according to claim 1, characterized in that, At the expected time of resource release events, the task running status, resource occupancy status, and processing component feedback are obtained. Confirmed events and unconfirmed events are distinguished based on task completion information and changes in resource availability. The basic equivalent computing power is calculated based on the actual resource status confirmed before the start of the time-series unit. The marginal equivalent computing power increment of confirmed events is fully included, and the marginal equivalent computing power increment of unconfirmed events is multiplied by the status credibility and included to obtain the reliable equivalent computing power.

9. The service node selection method based on computing power status according to claim 8, characterized in that, Determine the required computing power and expected processing time based on the processing needs; determine the time interval in the trusted computing power time series where the trusted equivalent computing power is continuously not lower than the required computing power and the duration is not shorter than the expected processing time as the serviceable time period, and generate the corresponding time domain matching results.

10. The service node selection method based on computing power status according to claim 9, characterized in that, The node scheduling priority is determined based on the expected completion time, the reliability of the status, and the minimum margin of the reliable equivalent computing power relative to the required computing power within the serviceable period. The target node is selected according to the node scheduling priority. If the node scheduling priorities are the same, the candidate node with the earlier expected completion time is selected. A resource reservation record is generated based on the processing stage, deployment time interval, resource type, and resource occupancy of the target node. A scheduling instruction containing the target node, the execution order of the processing stage, and the resource reservation time boundary is sent.

Images