An intelligent networked vehicle cloud service platform elastic scaling method and system
By employing multi-dimensional predictive models and refined resource scheduling, the scaling issues of the intelligent connected vehicle cloud service platform under business and regional differences have been resolved, achieving efficient resource utilization and service stability.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- DEEPAL AUTOMOBILE TECH CO LTD
- Filing Date
- 2026-04-30
- Publication Date
- 2026-06-19
AI Technical Summary
Existing scaling technologies for intelligent connected vehicle cloud service platforms cannot adapt to differences in business needs and geographical locations, resulting in response delays, resource waste, and service instability.
By acquiring historical sequence features, time features, and event features of QPS by region, business, and application, a multi-dimensional prediction model is used to accurately predict the QPS value of each region and business. Combined with the resource redundancy required by the business and the instance processing capacity, refined resource scheduling and elastic scaling are achieved by region and business.
It enables precise scheduling of resources on the intelligent connected vehicle cloud service platform, adapts to business and regional differences, avoids response delays and resource waste, and ensures service stability and efficient utilization.
Smart Images
Figure CN122247996A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of cloud computing and vehicle networking technology, specifically relating to a method and system for elastic scaling of an intelligent connected vehicle cloud service platform. Background Technology
[0002] With the increasing prevalence of intelligent connected vehicles, vehicle-to-everything (V2X) cloud service platforms need to process massive amounts of real-time vehicle data. Currently, there are two main automatic scaling technologies.
[0003] The first method is reactive scaling based on resource thresholds: it automatically scales up when CPU utilization exceeds 80% and automatically scales down when it falls below 30%. This method has a significant response delay. In scenarios with sudden traffic spikes, it takes 3-5 minutes from monitoring the resource threshold to completing the scaling up, during which time a large number of vehicle requests fail. Real-world testing data shows that during peak morning hours, the request failure rate due to scaling delays can reach 8%-15%.
[0004] The second method is pre-scaling and scaling up based on a fixed schedule: scaling up in advance during peak hours based on historical experience. This method cannot adapt to sudden changes in traffic, resulting in insufficient matching between the pre-scaling strategy and actual business volume. Statistics show that this method achieves a resource utilization rate of only 40%-50%, indicating significant resource waste.
[0005] In addition, both of the above scaling-up and scaling-down technologies have the following problems: (1) Lack of awareness of business differences: Different vehicle networking services have different QoS (Quality of Service) requirements. For example, remote vehicle control requires a latency of <100ms and availability of >99.9%; entertainment services can tolerate a latency of 500ms-1s and availability of >99%; data reporting allows for a latency of seconds and availability of >95%. The above two scaling-up and scaling-down technologies adopt a unified scaling-up and scaling-down strategy, which cannot meet the differentiated needs. (2) Insufficient consideration of regional characteristics: There are differences in traffic peak hours and holiday patterns in different cities. Taking Shanghai as an example, the morning peak is 7:00-9:00 and the evening peak is 17:00-19:00; while in Chengdu, the morning peak is 7:30-9:30 and the evening peak is 17:30-19:30. The above two scaling-up and scaling-down technologies cannot adapt to this regional difference. Summary of the Invention
[0006] In view of the shortcomings of the prior art, the purpose of this application is to provide a method and system for elastic scaling of intelligent connected vehicle cloud service platform, so as to quickly and accurately expand and shrink the application layer, and adapt to business differences and regional differences.
[0007] Firstly, this application provides a method for elastic scaling of an intelligent connected vehicle cloud service platform, comprising:
[0008] Obtain historical sequence features of region-business-QPS (i.e., requests per second), time features, spatial features, and event features of each prediction target time, as well as the current number of instances in each region, and perform preprocessing.
[0009] The preprocessed regional-business-QPS historical sequence features, as well as the temporal, spatial, and event features of each prediction target time, are input into the regional-business-QPS prediction model. The model then outputs the predicted QPS values for each region and business at each prediction target time (i.e., multiple regional-business-QPS sequences).
[0010] Based on the QPS prediction value (i.e., the QPS prediction value of each region and each service at each prediction target time), the preset single instance processing capacity, and the resource redundancy required by the service, the predicted number of instances required by each region and each service at each prediction target time is determined.
[0011] Calculate the predicted total number of instances required for each region at each prediction target time.
[0012] If the predicted total number of instances required for a certain region at a certain prediction target time is greater than the current number of instances in that region, then application layer scaling will be performed for that region.
[0013] By combining historical sequence features with time, space, and event features for joint prediction, and refining resource accounting by region and business, the intelligent connected vehicle cloud service platform achieves precise scheduling and efficient utilization of resources, adapting to business and regional differences. Through multi-dimensional prediction by integrating historical sequence features of region, business, and QPS with time, space, and event features, it accurately captures the fluctuation patterns of access volume in different regions and businesses, providing a highly reliable quantitative basis for resource planning and avoiding resource shortages or waste caused by prediction bias. Based on QPS predictions, single-instance processing capacity, and resource redundancy required by the business, the predicted number of instances is calculated, enabling refined accounting of resource needs. This ensures the platform has sufficient processing capacity during peak business periods while balancing service stability and resource economy. By statistically analyzing the total number of instances by region and comparing it with the current number of instances in real time, it can quickly identify resource bottleneck regions and trigger application-layer expansion, achieving on-demand, timely, and accurate elastic scaling, effectively avoiding service response delays, congestion, or even downtime caused by sudden increases in request volume.
[0014] Optionally, the method for determining the number of instances required for each region and each service at each prediction target time includes:
[0015] For the target time t, region d, and service s, obtain the QPS prediction value of service s in region d at that time, and calculate the target QPS prediction value after considering the redundancy requirements of service s, based on the resource redundancy required by service s.
[0016] Divide the target QPS prediction by the processing capacity of a single instance in region d for service s to obtain the raw number of instances.
[0017] The original value of the number of instances is rounded up to obtain the predicted number of instances required for service s in region d at time t.
[0018] By incorporating resource redundancy into the computational model, reasonable resource reserves can be reserved to effectively withstand fluctuations in business traffic and sudden access surges, ensuring the continuous, stable, uncongested, and undegraded operation of intelligent connected vehicle cloud services. Employing independent computation methods based on different regions and businesses allows for precise matching of the differentiated processing capabilities and load characteristics of different regions and businesses, avoiding resource over- or under-allocation caused by uniform configuration. Rounding up ensures that the calculated number of instances is the smallest integer sufficient to meet processing requirements, maximizing resource utilization while guaranteeing adequate service capacity. This provides a rigorous, reliable, and directly executable quantitative basis for subsequent elastic scaling decisions.
[0019] Optionally, the predicted total number of instances required by each region at each prediction target time is calculated by summing the predicted number of instances required by all services in region d at time t to obtain the predicted total number of instances required by region d at time t. This aggregation calculation method can accurately reflect the overall service pressure of each region, providing a comprehensive and accurate total basis for regional elastic scaling decisions. The direct accumulation calculation logic is simple and efficient, which helps reduce the computational load of the cloud platform scheduling system, improves the real-time performance and execution efficiency of resource elastic scaling decisions, ensures that the overall platform resource configuration matches the actual total load, and thus guarantees the stability and rationality of resource allocation for intelligent connected vehicle cloud services.
[0020] Optionally, the resource redundancy required by the service s The method for obtaining this is as follows: Based on the business 's', query the preset business level table to obtain the business level to which business 's' belongs. Then, based on the business level to which business 's' belongs, query the preset correspondence table between business levels and resource redundancy to obtain the corresponding required resource redundancy. Resource redundancy is configured based on business level, which can differentiate the security levels of core and ordinary businesses. Higher redundancy is allocated to higher-level businesses to ensure uninterrupted service, while resources are allocated reasonably to lower-level businesses to improve overall utilization. Values are retrieved from a pre-defined level table and a pre-defined mapping table between business levels and resource redundancy, eliminating the need for complex real-time calculations. The logic is simple and the response is rapid, effectively reducing the platform's decision-making load while ensuring uniform and consistent resource redundancy configuration.
[0021] Optionally, in the preset correspondence table between business levels and resource redundancy, business levels are divided into P, A, and B levels. If the business level is P, the resource redundancy is 25%–35%; if the business level is A, the resource redundancy is 10%–20%; and if the business level is B, the resource redundancy is 0. By dividing business levels into P, A, and B and configuring three-tiered resource redundancy levels of 25%–35%, 10%–20%, and 0 respectively, a differentiated resource guarantee mechanism is achieved for high reliability of core intelligent connected vehicle cloud services, high efficiency of ordinary services, and low cost of non-critical services. Configuring the highest redundancy ratio for high-priority P-level services effectively resists traffic surges and instance failures, ensuring the continuous and stable operation of critical vehicle networking services. Configuring medium redundancy for A-level services ensures service quality while considering resource economy, avoiding waste caused by excessive redundancy. Setting zero redundancy for B-level services maximizes resource utilization and adapts to non-real-time, non-critical business scenarios.
[0022] Optionally, the time features include date, day of the week, whether it is a weekday / holiday, and quarter; the spatial features include the number of vehicles online per region, traffic volume in key areas, and P99 latency; the event features include severe weather codes, large-scale event / sport event pedestrian flow index, and traffic accident impact index. Specifically, the severe weather code is a corresponding severe weather digital code mapped from weather types obtained from a meteorological API; the large-scale event / sport event pedestrian flow index is an estimated pedestrian flow index based on the scale of the large-scale event and / or sporting event, obtained from an external event data service; and the traffic accident impact index is a value mapped from the real-time road congestion index obtained from a traffic situation API.
[0023] By fusing multi-dimensional and multi-source heterogeneous features, a predictive input system is constructed. Preprocessing temporal, spatial, and event features comprehensively covers key factors influencing QPS fluctuations in intelligent connected vehicle cloud services, significantly improving the accuracy and generalization ability of the predictive model. Introducing temporal features such as date, weekday, weekday / holiday, and quarterly data accurately depicts the periodicity and seasonality of business traffic, enabling the model to adaptively learn from regular traffic trends. Collecting spatial features such as regional vehicle online counts, traffic heat in key areas, and P99 latency allows for real-time reflection of actual load levels and service pressure in different regions, achieving precise matching of resource demand and spatial distribution. Digitizing external events such as severe weather, large-scale events, and traffic accidents into codes, traffic flow indices, and impact indices quantifies the disturbance impact of sudden environmental and social events on vehicle network traffic, effectively improving the model's robustness in predicting abnormal peaks and sudden traffic surges.
[0024] Optionally, when scaling up the application layer for a specific region, if the resource pool for that region is insufficient, the application layer scaling will be performed according to the preset business scaling priority from high to low. Implementing application layer scaling based on the preset business scaling priority when the target region's resource pool is insufficient enables orderly and differentiated execution of scaling scheduling in resource-constrained scenarios. This ensures that high-priority businesses receive resources first, avoiding resource contention and scheduling chaos. It also ensures that core businesses and critical services can still complete scaling up under resource-constrained conditions, maintaining the service continuity and operational stability of the intelligent connected vehicle cloud platform.
[0025] Optionally, the method for scaling up the application layer for a specific region is as follows: if application layer scaling for region d is required at time t, then during the time interval from time t0 to time t, the number of instances increases by 1 every time interval T; where, , This represents the predicted total number of instances required for region d at time t. Let t represent the number of instances in region d at time t0, where t - t0 = T. th T th This indicates the preset duration.
[0026] Based on the number of instances required at time t ( ), number of instances at time t0 ( By constructing a formula for calculating the expansion interval, the expansion rhythm can be precisely quantified and controlled, ensuring that the expansion process is completed at a uniform and controllable speed before reaching the predicted target time, thus guaranteeing that resources are gradually in place and do not concentrate on impacting the platform.
[0027] Optionally, if application-layer scaling for region d is required at time t, the number of instances will be dynamically adjusted using proportional-integral (PI) control during the period from time t until the application-layer scaling is maintained. Dynamically adjusting the number of instances using PI control during the scaling maintenance phase enables real-time tracking and closed-loop correction of platform load changes, improving the adaptability and stability of elastic scaling. Dynamically correcting the number of instances through PI control effectively eliminates resource deviations caused by QPS prediction errors, actual traffic fluctuations, and external sudden factors, avoiding resource insufficiency or excessive redundancy, and ensuring accurate matching of service capacity with actual load.
[0028] Optionally, after scaling up the application layer for a specific region, if scaling down the application layer for that region is required (e.g., after a traffic peak), the number of instances will be gradually reduced according to a preset scaling-down gradient until it returns to the baseline level. Using a preset scaling-down gradient to gradually reduce the number of instances after application layer scaling up ensures a smooth transition and orderly convergence of resource release, avoiding sudden increases in service pressure, connection interruptions, or system instability caused by instantaneous batch scaling down. By gradually reducing instances according to a gradient and returning to the baseline level, the gradual pattern of traffic decline can be fully adapted, slowly releasing idle resources while ensuring service processing capacity redundancy, thus balancing business stability and resource utilization economy.
[0029] Secondly, this application provides an elastic scaling system for an intelligent connected vehicle cloud service platform, including a memory and a processor. The memory stores a computer program, and the processor is configured to execute the computer program to implement the aforementioned elastic scaling method for the intelligent connected vehicle cloud service platform. Attached Figure Description
[0030] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the embodiments of this application will be described below.
[0031] Figure 1 This is a schematic diagram of the elastic scaling system of the intelligent connected vehicle cloud service platform in the embodiments of this application.
[0032] Figure 2 This is a flowchart of the elastic scaling method for the intelligent connected vehicle cloud service platform in this application embodiment. Detailed Implementation
[0033] The embodiments of this application will be described in further detail below with reference to the accompanying drawings and examples. The detailed description of the following embodiments and the accompanying drawings are used to illustrate the principles of this application by way of example, but should not be used to limit the scope of this application, that is, this application is not limited to the described embodiments.
[0034] like Figure 1 As shown, the intelligent connected vehicle cloud service platform elastic scaling system includes a memory and a processor. The memory stores a computer program, and the processor is configured to execute the computer program to implement the intelligent connected vehicle cloud service platform elastic scaling method described in the embodiments of this application. The intelligent connected vehicle cloud service platform elastic scaling system adopts a microservice architecture and can be deployed in mainstream public cloud or private cloud environments (such as Alibaba Cloud, AWS, and Tencent Cloud). It mainly consists of the following four core services.
[0035] Data Acquisition and Feature Engineering Services: Responsible for real-time acquisition of raw data from the vehicle networking platform, including vehicle GPS trajectory, CAN signal, service request log, network quality, number of online vehicles, traffic intensity in key areas, P99 latency, etc., and cleaning and feature extraction to generate time-series feature vectors that can be used by prediction models.
[0036] Multi-dimensional time series prediction service: It carries a pre-trained region-business-QPS prediction model, receives feature data, and performs predictions at multiple time granularities such as the next 5 minutes, 30 minutes, and 1 hour.
[0037] Business-aware elastic decision-making service: Integrates a business strategy library to generate elastic scaling plans that include specific resource types, quantities, and time points based on forecast results, current resource status, and business rules.
[0038] Scaling and scaling execution and monitoring service: As an executor, it calls the APIs of cloud platforms (such as AWS Auto Scaling, Kubernetes Cluster Autoscaler) to perform application-layer scaling actions and monitors the execution status and business metrics in real time, forming a feedback loop.
[0039] like Figure 2 As shown in the embodiments of this application, the elastic scaling method of the intelligent connected vehicle cloud service platform includes the following steps:
[0040] S1. Obtain the historical sequence features of region-business-QPS (i.e., requests per second), the time features, spatial features and event features of each prediction target time, and the current number of instances in each region, and perform preprocessing, and then execute S2.
[0041] As an example, the region-business-QPS (requests per second) historical sequence feature is a sequence feature composed of the actual QPS values of each region and each business at various times over the past 7 days. As an example, the prediction target time could be any time within the next 5 minutes or the next 30 minutes. The number of instances refers to the number of independent service nodes / server replicas running business services in the cloud platform. The number of instances is the smallest scheduling unit for carrying business traffic, processing interface requests, and absorbing QPS load.
[0042] In one possible implementation, time features include date, day of the week, whether it is a weekday / holiday, and quarter. The date and day of the week use sine-cosine encoding for easy identification and improved prediction accuracy. By introducing time features such as date, day of the week, weekday / holiday, and quarter, the periodicity and seasonality of business traffic can be accurately characterized, enabling the model to adaptively learn from regular traffic trends.
[0043] In one possible embodiment, spatial features include the number of online vehicles at the regional level, traffic intensity in key areas, and P99 latency. These spatial features—regional online vehicle count, traffic intensity in key areas, and P99 latency—can reflect the actual load levels and service pressure in different regions in real time, achieving precise matching of resource demand and spatial distribution. These spatial features are obtained from the vehicle-to-everything (V2X) platform.
[0044] In one possible implementation, event features include severe weather codes, large-scale event / sport traffic flow indices, and traffic accident impact indices. By digitizing external events such as severe weather, large-scale events, and traffic accidents into codes, traffic flow indices, and impact indices, the disruptive impact of sudden environmental and social events on vehicle-to-everything (V2X) traffic can be quantified, effectively improving the model's robustness in predicting abnormal peaks and sudden traffic surges.
[0045] Severe weather codes are numerical codes mapped to the weather types in the meteorological API. For example, if the weather type at the target time is predicted to be light rain, the severe weather code is 1; if the weather type at the target time is predicted to be heavy rain, the severe weather code is 2; if the weather type at the target time is predicted to be light snow, the severe weather code is 3; if the weather type at the target time is predicted to be blizzard, the severe weather code is 4; otherwise, the severe weather code is 0.
[0046] The large-scale event / sports event attendance index is an estimated attendance index based on the scale of the large-scale event and / or sports event, obtained from an external event data service. For example, if the attendance of a large-scale event and / or sports event is less than 10,000 people, the attendance index is 1; if the attendance of a large-scale event and / or sports event is between 10,000 and 50,000 people, the attendance index is 2; and if the attendance of a large-scale event and / or sports event is greater than 50,000 people, the attendance index is 3.
[0047] The traffic accident impact index is a value mapped from the real-time road congestion index obtained through the traffic situation API. It connects to external traffic data services, such as the traffic situation API of Gaode / Baidu Maps, and maps the data to a value from 1 to 10.
[0048] S2. Input the preprocessed regional-business-QPS historical sequence features and the time, spatial and event features of each prediction target time into the regional-business-QPS prediction model. After prediction, output the QPS prediction values of each region and each business at each prediction target time (i.e., multiple regional-business-QPS sequences). Then execute S3.
[0049] As an example, the region-business-QPS prediction model is obtained through training. The training process is as follows:
[0050] 1. Model Architecture Design: Multi-task Learning Based on LSTM
[0051] To efficiently handle massive "region-business" combination prediction tasks and achieve knowledge sharing, the system employs a multi-task learning framework based on Long Short-Term Memory (LSTM) networks for modeling. This framework consists of two parts:
[0052] The first part, the shared LSTM encoder, is a multi-layered shared LSTM network that serves as the backbone of the model. Temporal feature data from all "region-business" tasks are input into this encoder. Its function is to learn common spatiotemporal dependency patterns of intelligent connected vehicle business loads from global data, such as daily / weekly cycles, holiday effects, and the general impact of weather on vehicle network traffic.
[0053] The second part consists of multi-task-specific prediction heads: Each "region-business" task requiring prediction (such as "Beijing-remote vehicle control" or "Shanghai-online navigation") corresponds to an independent small neural network (shallow LSTM). Each head receives a general context vector from the output of the same shared LSTM encoder and learns the personalized patterns for that specific task based on this vector, ultimately outputting a prediction sequence specific to that task.
[0054] 2. Data Sample Construction: Supervised Learning
[0055] Using historical data from the past 6 months, supervised learning samples are constructed with a 7-day rolling window and a 1-hour step size. Each sample is an (X,y) pair.
[0056] Input feature X: A matrix of shape [n, d]. Here, n represents the number of historical time points (e.g., 120), and d represents the feature dimension (e.g., 50-dimensional, including: encoded temporal features, spatial features, event features, and a region-service-QPS sequence feature composed of the actual QPS values for each region and service at each time point). Specifically, it includes feature vectors from the past n time points: .
[0057] Target label y: A vector of length m, representing the QPS (Queries Per Second) for each region and service to be predicted at the next m time points (e.g., the next 30 minutes). The model output is a sequence of region-service-QPS for the next m time points. .
[0058] 3. Feature preprocessing: standardization
[0059] Standardize all numerical features. For example, for numerical features... Standardization is performed to eliminate differences in units and ranges among different features, ensuring a mean of 0 and a standard deviation of 1, thereby accelerating model training convergence. The standardization formula is as follows:
[0060]
[0061] This represents the arithmetic mean of all sample values for a certain feature in the training set. This represents the standard deviation of all sample values for this feature in the training set, measuring the dispersion of the feature values. This represents the standardized eigenvalues. This represents the original feature values that need to be standardized.
[0062] 4. Training and Optimization Strategies
[0063] The model training employs the following strategies to ensure the accuracy and robustness of predictions:
[0064] Loss function: The Huber loss function is adopted to balance the accuracy of prediction and robustness to outliers.
[0065]
[0066] in, Let e represent the Huber loss function, and e represent the residual. This represents the threshold parameter.
[0067] Optimization and Validation: The training process uses time-series cross-validation to prevent data leakage.
[0068] Optimizer: The AdamW optimizer is used in conjunction with a learning rate scheduler for training.
[0069] Overfitting prevention: Apply early stopping on the validation set to prevent the model from overfitting.
[0070] In addition, the system integrates Prometheus and Grafana for comprehensive monitoring, with key metrics including: prediction error of each service, scaling latency, resource utilization, P99 latency, and error rate. These metrics can not only be used for alerts but also be summarized weekly to retrain the prediction model, optimize model parameters, and achieve system self-iteration and continuous optimization.
[0071] The online prediction process using the region-business-QPS prediction model is as follows: The prediction service runs once per minute. The system inputs the pre-processed historical sequence features of region-business-QPS, along with the temporal, spatial, and event features of each prediction target time, into the trained region-business-QPS prediction model. The model performs parallel inference and outputs the predicted QPS values for all monitored "region-business" tasks at multiple 30-minute time points in the future. The model's output is a structured set of prediction results, for example: {"Beijing - Remote Vehicle Control": [850, 920, 1000, ...], "Shanghai - Data Reporting": [12000, 12500, ...], ...}. Each prediction sequence is explicitly associated with its business type and geographical region. These prediction results with clear business and region labels are published in real time to a message queue (such as Kafka) for subsequent elastic decision-making subscription and consumption.
[0072] S3. Based on the QPS forecast values for each region and each service at each forecast target time, the preset single instance processing capacity, and the resource redundancy required by the service, determine the forecast value of the number of instances required for each region and each service at each forecast target time, and then execute S4.
[0073] In one possible embodiment, the method for determining the predicted number of instances required for each region and each service at each prediction target time includes:
[0074] For the target time t, region d, and service s, obtain the QPS prediction value of service s in region d at that time, and calculate the target QPS prediction value after considering the redundancy requirements of service s, based on the resource redundancy required by service s.
[0075] Divide the target QPS forecast, which takes redundancy requirements into the processing capacity of a single instance in region d for service s, to obtain the raw number of instances.
[0076] The original value of the number of instances is rounded up to obtain the predicted number of instances required for the service s in region d at time t.
[0077] As an example, using the formula: Calculate the predicted number of instances required for service s in region d at time t (which is within the target prediction time). ;in, This represents the processing capacity of a single instance in region d for business s. This represents the predicted QPS value of service s in region d at time t. This indicates the resource redundancy required by service s. This indicates rounding up to the nearest integer.
[0078] By incorporating resource redundancy into the computing model, reasonable resource reserves can be reserved to effectively withstand fluctuations in business traffic and sudden access surges, ensuring the continuous, stable, uncongested, and undegraded operation of intelligent connected vehicle cloud services. Adopting independent computing methods based on different regions and businesses allows for precise matching of the differentiated processing capabilities and load characteristics of different regions and businesses, avoiding resource overabundance or underabundance caused by uniform configuration.
[0079] In one possible implementation, the resource redundancy required by service s The method for obtaining the resource redundancy is as follows: First, query the preset service level table based on service 's' to obtain the service level to which service 's' belongs. Then, query the preset correspondence table between service level and resource redundancy based on the service level to which service 's' belongs to obtain the corresponding required resource redundancy. Resource redundancy is configured based on business level, which can differentiate the security levels of core and ordinary businesses. Higher redundancy is allocated to higher-level businesses to ensure uninterrupted service, while resources are allocated reasonably to lower-level businesses to improve overall utilization. Values are retrieved from a pre-defined level table and a pre-defined mapping table between business levels and resource redundancy, eliminating the need for complex real-time calculations. The logic is simple and the response is rapid, effectively reducing the platform's decision-making load while ensuring uniform and consistent resource redundancy configuration.
[0080] In one possible embodiment, in the preset correspondence table between business level and resource redundancy, the business level is divided into P level (safety critical), A level (experience core), and B level (data reporting).
[0081] If the service level is P (e.g., remote door unlocking, engine start / stop, requiring a response time of <100ms), then the resource redundancy should be 25%–35%. For example, 30% could be used.
[0082] If the service level is A (e.g., online navigation, voice interaction, requiring a response time of <500ms), then the resource redundancy should be 10%–20%. As an example, it could be 15%.
[0083] If the service level is B (such as vehicle status reporting, log uploading, allowing for second-level delays, and batch processing), then the resource redundancy is 0.
[0084] By classifying business levels into P, A, and B and configuring three-tiered resource redundancy levels of 25%–35%, 10%–20%, and 0 respectively, a differentiated resource guarantee mechanism is achieved for the core business of intelligent connected vehicle cloud services, ensuring high reliability, high efficiency of ordinary business, and low cost of non-critical business.
[0085] S4. Calculate the predicted total number of instances required for each region at each prediction target time, and then execute S5.
[0086] In one possible implementation, the predicted total number of instances required by each region at each prediction target time is calculated by summing the predicted number of instances required by all services in region d at time t to obtain the predicted total number of instances required by region d at time t. This aggregation calculation method can accurately reflect the overall service pressure in each region, providing a comprehensive and accurate total basis for regional-level elastic scaling decisions.
[0087] S5. Determine whether the predicted total number of instances required for a certain region at a certain prediction target time is greater than the current number of instances in that region. If yes, execute S6; otherwise, return to execute S1.
[0088] S6. Expand the application layer for this region, and then execute S7.
[0089] As an example, upon receiving a predicted sequence of "Beijing - Remote Vehicle Control," the service s is determined to be at level P, and policy parameters such as the required resource redundancy of 30% and the processing capacity of a single instance for service s are obtained. Assuming that the predicted remote vehicle control requests in Beijing will increase to 1000 QPS in 30 minutes, and considering the current number of instances (5) and the SLA (i.e., each Pod's processing capacity is 100 QPS) for service s, an elasticity plan is generated: "In the next 30 minutes, the vehicle control service for the Beijing area will be expanded by 8 instances (Pods)."
[0090] In one possible implementation, when scaling up the application layer for a specific region, if the resource pool for that region is insufficient, the application layer scaling is performed in descending order of preset business scaling priorities. The program pre-sets the scaling priorities for each business; if the resource pool for that region is insufficient (i.e., the remaining number of instances that can be created is less than the sum of all pending scaling plans' current and near-term resource requirements), the application layer scaling is performed in descending order of business scaling priorities.
[0091] In one possible implementation, if application layer scaling for region d is required at time t, then during the time interval from time t0 to time t, the number of instances increases by 1 every time interval T; where, , This represents the predicted total number of instances required for region d at time t. Let t represent the number of instances in region d at time t0, where t - t0 = T. th T th This indicates the preset duration. For example, T... th =15min. Suppose that at time t0, the number of instances in region d is 5 (i.e., ... A peak period is expected in the next 30 minutes, requiring the number of instances to be increased to 20 (i.e., If the current time is t0 (i.e., within the next 15 minutes), the capacity will not be expanded. Within the 15 minutes starting from t0, the capacity expansion method is as follows: every minute, the number of instances increases by 1, so that the number of instances will be expanded to 20 by the time the next 30 minutes are reached.
[0092] Based on the number of instances required at time t ( ), number of instances at time t0 ( By constructing a formula for calculating the expansion interval, the expansion rhythm can be precisely quantified and controlled, ensuring that the expansion process is completed at a uniform and controllable speed before reaching the predicted target time, thus guaranteeing that resources are gradually in place and do not concentrate on impacting the platform.
[0093] In one possible implementation, if application layer scaling for region d is required at time t, the number of instances is dynamically adjusted using proportional-integral control during the time period from time t until the application layer scaling is maintained. The adjustment amount of proportional-integral control... satisfy: ;in, Indicates proportional gain. Indicates integral gain. Indicates deviation, , This represents the actual total number of instances required for the current region d (i.e., the actual total number of instances required for region d). (Through...) This is to quickly offset the impact of prediction errors.
[0094] S7. Determine whether application-layer scaling down is needed for this region. If yes, proceed to S8; otherwise, continue with S7.
[0095] As an example, if the total number of instances required for a region is less than or equal to the baseline number of instances for that region after a peak traffic period, it means that application-layer scaling down is required for that region.
[0096] In one possible implementation, the scaling down method involves gradually reducing the number of instances according to a preset scaling down gradient until it returns to the baseline level. For example, reducing redundant instances by 20% every 30 minutes until it returns to the baseline level, preventing frequent scaling down caused by short-term traffic fluctuations.
[0097] S8. Gradually reduce the number of instances according to the preset scaling gradient until it returns to the baseline level, and then return to execute S1.
[0098] By combining historical sequence features with time, space, and event features for joint prediction, and by refining resource accounting by region and business, and by expanding capacity elastically on demand, the intelligent connected vehicle cloud service platform achieves precise scheduling and efficient utilization of resources, adapting to business and regional differences.
[0099] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.
Claims
1. A smart connected vehicle cloud service platform elasticity method, characterized in that, include: Obtain historical sequence features of region-business-QPS, temporal features, spatial features, and event features at each prediction target time, as well as the current number of instances in each region, and perform preprocessing; The preprocessed regional-business-QPS historical sequence features, as well as the temporal, spatial, and event features of each prediction target time, are input into the regional-business-QPS prediction model. The model then outputs the predicted QPS values for each region and each business at each prediction target time. Based on the predicted QPS value, the preset single instance processing capacity, and the resource redundancy required by the business, the predicted number of instances required by each region and each business at each prediction target time is determined. Calculate the predicted total number of instances required for each region at each prediction target time; If the predicted total number of instances required for a certain region at a certain prediction target time is greater than the current number of instances in that region, then application layer scaling will be performed for that region.
2. The elastic scaling method for the intelligent connected vehicle cloud service platform according to claim 1, characterized in that: The methods for determining the predicted number of instances required for each region and each service at each prediction target time include: For the target time t, region d, and service s, obtain the QPS prediction value of service s in region d at that time, and calculate the target QPS prediction value after considering the redundancy requirements of service s by combining the resource redundancy required by service s. Divide the target QPS prediction by the processing capacity of a single instance in region d for service s to obtain the raw number of instances. The original value of the number of instances is rounded up to obtain the predicted number of instances required for service s in region d at time t; The method for calculating the predicted total number of instances required by each region at each prediction target time is as follows: add up the predicted number of instances required by all services in region d at time t to obtain the predicted total number of instances required by region d at time t.
3. The elastic scaling method for the intelligent connected vehicle cloud service platform according to claim 2, characterized in that, The resource redundancy required by the service s The obtaining manner is that: The business level to which business s belongs is obtained by querying a preset business level table based on the business s. Based on the business level to which business 's' belongs, query the preset correspondence table between business level and resource redundancy to obtain the corresponding required resource redundancy. .
4. The elastic scaling method for an intelligent connected vehicle cloud service platform according to claim 3, characterized in that, In the preset correspondence table between service level and resource redundancy, the service level is divided into P level, A level and B level; if the service level is P level, the resource redundancy is 25% to 35%; if the service level is A level, the resource redundancy is 10% to 20%; if the service level is B level, the resource redundancy is 0.
5. The elastic scaling method for an intelligent connected vehicle cloud service platform according to claim 1, characterized in that: The time characteristics include date, day of the week, whether it is a weekday / holiday, and quarter; The spatial characteristics include the number of vehicles online per region, traffic volume in key areas, and P99 latency; The event characteristics include severe weather codes, large-scale event / sport crowd flow index, and traffic accident impact index; The severe weather code is a corresponding severe weather digital code mapped from the weather type of the meteorological API; the large-scale event / sport traffic flow index is a traffic flow index estimated based on the scale of the large-scale event and / or sports event; and the traffic accident impact index is a value mapped from the real-time road congestion index obtained from the traffic situation API.
6. The elastic scaling method for the intelligent connected vehicle cloud service platform according to claim 1, characterized in that: When expanding the application layer for a specific region, if the resource pool for that region is insufficient, the application layer expansion will be carried out in descending order of the preset business expansion priority.
7. The elastic scaling method for an intelligent connected vehicle cloud service platform according to claim 1, characterized in that, The method for scaling up the application layer for a specific region is as follows: If application layer scaling for region d is required at time t, then during the time interval from time t0 to time t, the number of instances increases by 1 every time interval T; where, , This represents the predicted total number of instances required for region d at time t. This represents the number of instances in region d at time t0, where t - t0 = T. th T th This indicates the preset duration.
8. The elastic scaling method for an intelligent connected vehicle cloud service platform according to claim 7, characterized in that: If application layer scaling is required for region d at time t, the number of instances will be dynamically adjusted through proportional-integral control during the period from time t until the application layer scaling is maintained.
9. The elastic scaling method for an intelligent connected vehicle cloud service platform according to any one of claims 1 to 8, characterized in that: After scaling up the application layer for a region, if it is necessary to scale down the application layer for that region, the number of instances will be gradually reduced according to the preset scaling down gradient until it returns to the baseline level.
10. A flexible scaling system for an intelligent connected vehicle cloud service platform, comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: The processor is configured to execute the computer program to implement the elastic scaling method for the intelligent connected vehicle cloud service platform as described in any one of claims 1 to 9.