An online examination intelligent regulation and control method and system based on question granularity

Abstract

The application discloses an online examination intelligent regulation and control method and system based on a question granularity, the method builds a receiving and processing asynchronous decoupling architecture; a granular load quantification system based on question time consumption is constructed to accurately count multi-dimensional business loads; a dynamic computing power mapping table with full working condition boundary calibration is established to realize normalized matching of business loads and hardware computing power; global congestion is calculated by combining real-time loads and terminal cache pressure, composite risk coefficients are dynamically corrected by combining examination time sequence progress and concurrent scale, the maximum allowable processing load volume is adapted through reverse search of the dynamic mapping table, and the thread scale range of the processing module is accurately regulated and controlled; dual safety constraints are realized through computing power compliance constraints and real-time CPU hardware bottoming; the terminal distinguishes busy and idle working conditions to adaptively adjust reporting intervals, uploading granularity and cache thresholds, and peak shaving is realized from the source.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of online education and computer technology, specifically relating to an intelligent control method and system for online examinations based on question granularity. Background Technology

[0002] Currently, in high-concurrency scenarios for online examinations, the mainstream solutions in the industry mainly include four common technical means: server cluster expansion and deployment, asynchronous business splitting and processing, simple resource scheduling based on hardware utilization, and static traffic limiting for terminals. The above solutions can cope with the instantaneous access pressure brought by login and centralized submission to a certain extent, and are also the basic architecture mode commonly used by current online examination systems.

[0003] Horizontal scaling using server clusters is the most common approach. This method involves pre-configuring hardware based on the peak concurrency of the exam and using load balancing to distribute requests and handle simultaneous access from a large number of users. However, online exams exhibit typical tidal traffic patterns, with high concurrency occurring only during a few peak periods. For the majority of the time, the system operates under relatively low pressure, leaving server resources deployed according to peak standards idle for extended periods. Overall, resource utilization is low, and cluster setup, daily operations, and equipment maintenance all incur significant costs. Consequently, the resource investment does not align well with actual business needs.

[0004] The industry commonly uses an asynchronous processing architecture, which separates the reception of candidate answer submissions from subsequent data verification, parsing, and storage operations. It relies on message queues for task relay, prioritizing and quickly responding to user submission requests, effectively improving instantaneous concurrent reception capabilities. However, this architecture lacks a dynamic computing power allocation mechanism between the receiving service and the backend processing service. During peak periods, large-scale data processing tasks can easily overwhelm the resources of the frontend receiving module.

[0005] Most existing systems simply collect server CPU utilization for resource allocation, passively adjusting service resources based on hardware load status. This approach is relatively simple and crude. Such adjustments can only be made after the system has already experienced high load and lag, lacking traffic prediction capabilities and the ability to anticipate peak periods based on exam business patterns. The adjustments are significantly delayed and fail to proactively avoid system congestion.

[0006] At the terminal control level, most existing solutions use static rate limiting methods with fixed thresholds and fixed reporting frequencies to restrict the data upload pace of candidates. This one-size-fits-all control model has poor flexibility. When the system load is low, it still strictly restricts data synchronization, resulting in untimely updates of answer status. When the system pressure surges, the rate limiting is insufficient and cannot effectively reduce the uplink traffic. This not only affects the normal user experience of candidates, but also makes it difficult to stably control the synchronization pace of exam data across the entire network.

[0007] Overall, the existing control measures operate independently, with no coordinated coordination between server-side resource allocation and terminal traffic management. Furthermore, current controls rely solely on hardware load conditions, failing to integrate with the actual online examination scenario. This lack of consideration for factors such as the number of concurrent terminals, question types, differences in processing different questions of the same type, parsing time, and data backlog. Consequently, the system cannot flexibly adapt to the computational consumption characteristics of different examination scenarios, and it lacks a dynamic adaptive adjustment and continuous optimization mechanism. This results in insufficient allocation and utilization of system resources, hindering the maximization of system performance. Summary of the Invention

[0008] The purpose of this invention is to provide an intelligent control method for online examinations based on question granularity. This method aims to construct a dynamic quantization system for question granularity computing power normalization in online examination scenarios, reconstruct load quantization and hardware mapping standards, realize intelligent allocation and full utilization of resources, and maximize operational efficiency.

[0009] The technical solution to achieve the purpose of this invention is: an intelligent control method for online examinations based on question granularity, comprising the following steps:

[0010] S1. Deploy an asynchronous decoupled infrastructure, splitting the server into independent receiving and processing modules. The receiving and processing modules are asynchronously isolated through message queues. The receiving module, as a high-priority module, is responsible for receiving, caching, and verifying the integrity of data, and reserves computing resources to cope with peak pressure based on load conditions. The processing module, as a flexible and adjustable module, is responsible for data parsing, logical verification, and database storage, and dynamically adjusts the number of threads and computing power ratio based on real-time load conditions and historical operating data.

[0011] S2. Construct a unified load quantization system based on the granularity of the question, and use the unified load quantization system to quantify and iteratively update multi-dimensional core operating indicators;

[0012] S3. Construct and iterate in real time a dynamic mapping table of problem granularity load and hardware computing power. The dynamic mapping table uses the load ratio coefficient K as the index dimension and the computing power benchmark CP as the index dimension. K For storage dimensions, it achieves normalization and unification of computing power under different volumes and the same ratio of working conditions; it performs minimum distance matching lookup based on the load ratio coefficient K, and completes real-time iterative updates of the dynamic mapping table by combining historical computing power benchmarks and the current average computing power mapping;

[0013] S4. Based on the server's global carrying capacity, quantify the real-time load pressure and dynamically correct the risk of the final flood by combining the time sequence characteristics of the exam progress. Based on the quantitative computing power benchmark of the dynamic mapping table of question granular load and hardware computing power, construct a linkage control mechanism between the receiving and processing modules. Constrain the scheduling behavior of the processing module with the priority principle of the receiving module to realize the collaborative cooperation and dynamic adaptation of the two modules, and complete the risk prediction and dynamic resource matching in high-concurrency exam scenarios.

[0014] S5. During the exam, the load sample and mapping data are updated in real time. After the exam ends, the unprocessed answer data in the processing module is completed, and the question time parameters, time sequence data and dynamic mapping table data are updated iteratively.

[0015] This invention also proposes an online examination intelligent control system based on question granularity, used to execute an online examination intelligent control method based on question granularity, including:

[0016] The architecture is layered and modular, used to deploy an asynchronous decoupled architecture between the receiving module and the processing module. A message queue is used to achieve asynchronous isolation and priority scheduling between the receiving module and the processing module.

[0017] The question-granular load quantification module is used to quantify and iteratively update multi-dimensional core operating indicators, including: obtaining the actual reception time, the actual processing time, the real-time CPU utilization, and the number of current online terminals; quantifying CPU utilization, the number of concurrent terminals, reception load, processing load, queue backlog load, and terminal cache load; and constructing a unified question-granular load evaluation system.

[0018] The dynamic computing power mapping module is used to construct, calibrate the boundaries, and iterate the problem-granularity load-hardware computing power dynamic mapping table in real time. It achieves working condition lookup matching and computing power benchmark calibration update through minimum distance matching.

[0019] The pressure prediction and risk correction module is used to calculate the global receiving load congestion, generate the progress correction coefficient and combine it to obtain the dynamic risk coefficient, so as to realize the timing risk correction and the early prediction of flood peak risk.

[0020] The terminal traffic control module, deployed on the candidate's terminal, is used to adaptively adjust the reporting interval, upload granularity, and cached reporting threshold to achieve peak reduction at the source.

[0021] The receiving and processing module linkage control module is used to complete the receiving module's operating condition calibration, mapping table iteration, and receiving guarantee volume locking, and to realize the processing module's computing power amplitude adjustment and hardware security control through quantitative reverse retrieval.

[0022] The end-to-end iterative optimization module is used to iteratively update the load and mapping data in real time during the exam, and to iteratively optimize the question parameters and mapping table parameters across the entire domain after the exam.

[0023] Compared with the prior art, the significant advantages of this invention are:

[0024] (1) This invention constructs a unified load quantification system for questions, abandons the traditional whole machine homogeneous load statistics method, and uses the actual transmission time and processing time of different questions as the benchmark to quantify the examination business data with differentiated resource consumption and unify the load evaluation scale of the examination system.

[0025] (2) Based on load quantization, this invention constructs a dynamic mapping table of question granular load and hardware computing power that can be updated in real time, which solves the technical problem that abstract business load cannot be directly matched with hardware computing power occupation, and significantly improves the accuracy of resource scheduling and the stability of examination system operation in large-scale online examination scenarios.

[0026] (3) This invention constructs a multi-dimensional flood peak risk prediction and dynamic correction model to realize advance prediction and resource reservation for the flood peak at the end of the exam. This invention uses the server's global maximum receiving load as the upper limit benchmark, and calculates the global receiving load congestion by combining the real-time receiving load and the terminal cache potential load, simultaneously characterizing the current instantaneous pressure and short-term pre-burst pressure of the exam system. At the same time, it generates a basic correction increment by combining the time sequence relationship between the remaining exam time and the total time, and makes differentiated fine adjustments by matching the number of active concurrent terminals across the network to generate a dimensionless progress correction coefficient. The closer to the end of the exam and the larger the scale of concurrent terminals, the higher the magnitude of the time sequence risk correction. The progress correction coefficient and the global load congestion are combined to obtain a composite risk coefficient, so that the scheduling decision integrates the multi-dimensional characteristics of load volume, time sequence risk, and concurrent impact intensity, making up for the technical shortcomings of traditional solutions that rely only on real-time load and lack advance prediction capabilities.

[0027] (4) The server of this invention adopts a receiving-priority and flexible-adaptive linkage control mechanism between the receiving and processing modules to achieve refined collaborative scheduling of computing resources. The receiving module has the highest scheduling priority. Based on the composite risk coefficient, the guaranteed volume of the receiving module that cannot be squeezed out is calculated, which defines the resource boundary for the control of the processing module. The processing module adopts a quantitative reverse retrieval mechanism to match the computing power benchmark that meets the computing power security requirements, accurately obtains the maximum allowable processing load volume that can be adapted, and dynamically adjusts the scale of processing threads, which not only ensures the stability of the upload of answer data, but also maximizes the utilization rate of the whole machine resources.

[0028] (5) This invention sets up a terminal busy / idle adaptive traffic control mechanism to achieve peak shaving at the source. By adjusting the reporting time interval, single upload granularity of questions, and local cache reporting threshold of the candidate terminal in combination with the busy / idle status and the exam schedule, the source traffic control achieves real-time maintenance during idle times and peak shaving during busy times.

[0029] This invention effectively solves industry pain points such as high concurrency surges and lag, inefficient resource scheduling, and low resource utilization in large-scale online examination scenarios, and significantly improves the concurrent carrying capacity, operational stability, and refined resource control level of online examination systems.

[0030] The present invention will now be described in further detail with reference to the accompanying drawings. Attached Figure Description

[0031] Figure 1 This is a flowchart illustrating a specific embodiment of an online examination intelligent control method based on question granularity.

[0032] Figure 2 This is a flowchart illustrating the title granularity load quantization and computing power mapping process in an embodiment of the present invention.

[0033] Figure 3 This is a flowchart illustrating the collaborative load control process of the receiving module and processing module in risk prediction and edge-cloud linkage according to an embodiment of the present invention. Detailed Implementation

[0034] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0035] A method for intelligent control of online examinations based on question granularity is proposed. This method establishes an asynchronous decoupled architecture for receiving and processing, ensuring high-priority and stable transmission from the receiving module and flexible adjustment of the processing module. It constructs a granular load quantification system based on question duration to accurately calculate multi-dimensional business load. A dynamic computing power mapping table with full-condition boundary calibration is established to achieve normalized matching between business load and hardware computing power. Global congestion is calculated by combining real-time load and terminal cache pressure, and a composite risk coefficient is dynamically generated by integrating examination timing and concurrency scale. The maximum allowable processing load is adapted to the dynamic mapping table through reverse retrieval, accurately controlling the thread scale of the processing module. Dual security constraints are achieved through computing power compliance constraints and real-time CPU hardware safeguards. The terminal adaptively adjusts the reporting interval, upload granularity, and cache threshold according to busy / idle conditions, thus reducing peak loads at the source. The method includes:

[0036] S1. Deploy an asynchronous, decoupled infrastructure: The server is split into an independent receiving module and a processing module, isolated by a message queue. The receiving module is responsible for receiving, caching, and verifying the integrity of student answer data. As a high-priority module, it can reserve computing resources to handle peak loads. The processing module is responsible for data parsing, logical verification, and data storage. As a flexible and adjustable module, it can dynamically adjust the number of threads and computing power allocation based on real-time load and historical data. Specifically, the message queue is the data receiving queue used in the asynchronous processing architecture, established using the RabbitMQ message middleware.

[0037] S2. Construct a unified load quantification system based on question granularity to obtain the actual reception time, actual processing time, real-time CPU utilization, and number of currently online terminals. The unified load quantification system quantifies core operational metrics such as CPU utilization, concurrent terminals, reception load, processing load, queue backlog load, and terminal cache load based on the acquired data. Unlike traditional homogenized load statistics methods, this step uses the average reception time and average processing time of each question as a weighted benchmark to perform equivalent quantification on question data with different computing power and network consumption, thus unifying the load measurement scale. Specific indicator definitions and quantification methods are as follows:

[0038] Average reception time T for the question R : Calculate the average time taken for each question from receiving the question data from the server receiving module to completing the data reception, caching, and integrity verification. The value is obtained by iterative update, that is, the average time taken at the current moment is the average of the previous historical average time and the actual time taken at the current moment. The actual time taken at the current moment is used directly for the first run.

[0039] Average processing time T for each problem P The system calculates the average time taken for each question from receiving the question data from the server-side processing module to completing data parsing, logical verification, and data storage. The iteration rule for the value is consistent with the average time taken to receive the question, thus achieving continuous convergence of the question's computing power consumption characteristics.

[0040] CPU average utilization C: Real-time CPU utilization of the data server is collected, and mean smoothing and noise reduction are performed through a fixed time window to eliminate instantaneous sampling jitter interference and characterize the steady-state hardware computing power occupancy of the server.

[0041] Concurrent Terminals N: This counts the number of active (answering questions and not offline) candidate terminals, used to help assess the scale of upstream concurrency.

[0042] Receive load R: Real-time load volume at the granular level of the receiving side. It is obtained by statistically analyzing the number of various types of questions currently received in real time, matching the average receiving time of the corresponding questions, and weighting the sum. It reflects the receiving load pressure of the server in the receiving process.

[0043] Processing load P: The granular load volume of the questions being processed by the server. It is obtained by weighted summation of the number of various types of questions being processed by the server and the average processing time of the corresponding questions, reflecting the current processing load pressure of the server.

[0044] Waiting backlog load Q: The granular load volume of questions waiting to be processed in the server queue. It is obtained by weighted sum of the number of various types of questions waiting in the message queue and the average processing time of the corresponding questions, reflecting the backlog load pressure in the server queue.

[0045] Terminal cache load B: The load volume of questions cached locally on the terminal. The terminal status feedback module collects and reports the number of questions cached locally. The server counts the number of various types of questions that have not been uploaded in the terminal's local cache. The load is obtained by weighted summing the average receiving time of the corresponding questions. This is used to predict the uplink surge pressure that will occur in the near future.

[0046] S3. Construct a dynamic mapping table between question-level load and hardware computing power to achieve a quantitative mapping between business load and hardware computing power. Essentially, the dynamic mapping table is a mapping record of historical operating data. To solve the technical problem that abstract question-level load cannot directly correspond to server hardware computing power occupancy, this invention constructs a dynamic mapping table between question-level load and hardware computing power that can be updated in real time, providing a basis for decision-making regarding computing power assurance and processing thread control.

[0047] This mapping table uses the load ratio coefficient K as the index dimension and the computing power benchmark CP as the benchmark. K For storage, this achieves normalized matching of computing power under different volumes but with the same allocation ratio. The load allocation coefficient is the characteristic of the ratio between processing and receiving loads, reflecting the operating condition distribution of the examination system; the computing power benchmark is the steady-state CPU computing power occupancy corresponding to the receiving load.

[0048] A preset load fluctuation tolerance threshold Kmax is established, defining the effective index range [0, Kmax] to avoid lookup failures caused by extreme load imbalance or no load. Based on the boundary calibration rules of the load ratio coefficient and computing power benchmark, the value range of the load ratio system is standardized within the effective index range. These boundary calibration rules are set as follows:

[0049] When there is neither receiving load nor processing load, i.e., P=0 and R=0, take K=0, CP K =C;

[0050] When the processing load is greater than the receiving load or there is no receiving load, i.e., P / R > Kmax or (P ≠ 0 and R = 0), take K = Kmax, CP K =C / P;

[0051] When the processing load is less than the receiving load or there is no processing load, i.e., P / R < 1 / Kmax or (P = 0 and R ≠ 0), take K = 1 / Kmax, CP K =C / R;

[0052] When there is both receiving and processing load, i.e., P≠0 and R≠0, take K=P / R, CP K =C / R;

[0053] Among them, the load fluctuation tolerance threshold Kmax is an empirically calibrated threshold adapted to the load fluctuation characteristics of online examination services, used to limit the effective range of load ratio under normal steady-state operating conditions of the examination system. This value can be obtained by statistically fitting the maximum steady-state load imbalance multiple obtained from historical operating data of online examinations. The larger the value, the more complete the normal operating conditions covered by the effective index range of load ratio, and the higher the accuracy of the dynamic mapping table.

[0054] Mapping table lookup and iterative update rules: The computing power benchmark data in the mapping table is a dynamic average, updated in real-time online iteratively. Using the amplitude-limited standardized K as the unique retrieval index, the historical index node with the highest fit to the current operating conditions is selected based on the minimum distance matching principle, and the corresponding historical computing power benchmark data is read. The average value of the historical computing power benchmark value and the current computing power mapping field is used as the calibration of the computing power benchmark value, and updated to the mapping table in real time. The minimum distance can be either Euclidean distance or Manhattan distance, with Euclidean distance being preferred.

[0055] S4. Load control of the receiving and processing modules through global pressure prediction, dynamic progress correction, and end-to-cloud linkage. Based on the server's global carrying capacity, the real-time load pressure of the server is quantified. Combined with the exam progress time-series characteristics, the risk of peak load at the end of the exam is dynamically corrected. Relying on the quantified computing power benchmark of the dynamic mapping table between question-level load and hardware computing power, a linkage control mechanism for the receiving and processing modules is constructed. The scheduling behavior of the processing module is constrained by the priority principle of the receiving module, realizing the collaborative cooperation and dynamic adaptation of the two modules. This completes the risk prediction and dynamic resource matching in high-concurrency exam scenarios. The specific steps are as follows:

[0056] (1) Global Receiver Load Pressure Quantification and Congestion Calculation. Using the server's maximum global receive load as a unified benchmark, the global receive load congestion is calculated by combining the receive load R, representing instantaneous pressure, and the terminal buffer load B, representing short-term potential pressure. The global receive load congestion is a normalized pressure index, specifically calculated as: Global Receiver Load Congestion = (R + B) / S, where S is the server's maximum global receive load, representing the maximum total receive load at the question granularity that the server can handle. This can be determined based on the computing power benchmark CP corresponding to the extreme condition K = 1 / Kmax in the dynamic mapping table. K1 Calculate S = Tout / CP K Tout represents the maximum network request timeout allowed by the examination system. Global receive load congestion can simultaneously characterize the current actual resource occupancy status of the examination system and the upcoming pre-occupancy pressure, enabling a comprehensive assessment of load pressure.

[0057] (2) Dynamic Risk Correction of Examination Progress. A basic correction increment is generated based on the ratio of the remaining examination time to the total examination time. The correction increment is then fine-tuned based on the number of active concurrent terminals N across the entire network to obtain a dimensionless progress correction coefficient, which represents the examination timing risk and the uplink concurrency impact density across the entire network. The closer to the end of the examination and the larger the scale of concurrent terminals, the higher the magnitude of the progress correction coefficient. The basic correction increment is set to different values ​​according to the current proportion of the examination time. The closer to the end of the examination, the larger the basic correction increment. For example, when the basic correction increment is between 1.0 and 2.0, it is set to 2.0 if the remaining examination time is less than 5 minutes, 1.5 if it is less than 10 minutes, and 1.0 otherwise. The progress correction coefficient is determined by multiplying the basic correction increment by the ratio of the number of concurrent terminals N to the maximum number of concurrent terminals set by the server. The progress correction coefficient is multiplied by the global receiving load congestion to obtain the final composite risk coefficient used for resource scheduling of the receiving and processing modules.

[0058] Among them, the global receiving load congestion characterizes the current and short-term potential overall uplink load volume and hardware saturation status of the examination system, forming the basic benchmark for pressure assessment; the concurrent terminal scale is used to quantify the concurrent impact intensity of uplink requests, which can effectively distinguish the severity of different end-of-term peaks under the same load volume; the progress correction coefficient completes the dynamic risk correction of the receiving load. This composite correction method enables the composite risk coefficient to simultaneously possess the basic load pressure, the time-series evolution risk, and the characteristics of concurrent operating conditions. It can proactively amplify the potential submission peak risk at the end of the examination, realize the advance prediction of peak pressure and advance resource reservation, and make up for the shortcomings of traditional load control schemes that only rely on real-time load assessment, lack advance prediction capabilities, and cannot distinguish the differences in concurrent impact intensity under the same load.

[0059] (3) Server-side load control between receiving and processing modules. Based on the quantified receiving load pressure and the corrected peak risk, a dynamic computing power mapping table is used to construct the linkage control logic between the receiving and processing modules: the receiving module prioritizes locking dedicated computing power and cache resources to ensure stable uploading of exam data; the processing module, under the premise that the receiving module has sufficient resources, adapts to the maximum allowable processing load and controls the number of threads. The two modules constrain each other and adapt in a linkage manner to achieve refined utilization of computing power resources in the exam system.

[0060] (a) Priority is given to ensuring high-priority resources for the receiving module. The current receiving load R and processing load P are collected in real time. According to the dynamic mapping table lookup and matching and iterative update rules, the hardware computing power characteristics under the current load ratio are calibrated and the dynamic mapping table is iteratively updated. Then, the receiving guarantee volume Rq = R + R × composite risk coefficient is predicted based on the global receiving load congestion and composite risk coefficient. Rq is the minimum guarantee benchmark that the receiving module cannot be squeezed out, and it sets the constraint boundary for the flexible adjustment of subsequent processing modules.

[0061] (b) Flexible thread-based iterative linkage control of the processing module. The processing module is a flexible and adjustable module. Under the premise of reserving a guaranteed reception capacity Rq, it carries out adaptive control of processing computing power, which must simultaneously meet the dual conditions of predicted computing power compliance and actual hardware load security.

[0062] The regulation of computing power compliance adopts a quantitative reverse retrieval mechanism to determine the regulation range. The regulation process is as follows: First, based on the received guaranteed volume Rq and the processing load P, a load ratio Kq after limiting and standardization is constructed. According to Kq, the computing power benchmark mapping value CPq is obtained by retrieving the mapping table and the CPU utilization Cq is calculated in reverse as a prediction of the computing power benchmark occupancy trend. Then, Cq is compared with the hardware computing power safety threshold. When Cq exceeds the hardware computing power safety threshold, the mapping table is retrieved in reverse along the direction of decreasing load ratio to match the computing power benchmark that meets the computing power safety requirements. The maximum allowable processing load volume Pnew is accurately obtained. The processing volume and thread computing power occupancy are iteratively reduced, and the number of processing threads is actively reduced to release computing power resources until the processing load is less than the maximum allowable processing load volume. When Cq is within the hardware computing power safety threshold, the mapping table is retrieved in reverse along the direction of increasing load ratio to match the computing power benchmark that meets the computing power safety requirements. The scale of processing threads is appropriately expanded, and the number of processing threads is actively increased while the processing load does not exceed the maximum allowable processing load volume.

[0063] The actual hardware load security control process is as follows: the average CPU utilization Ch after smooth noise reduction on the server side is used as the hardware security benchmark. When Ch exceeds the hardware computing power security threshold, the processing volume and thread computing power occupancy on the processing side are iteratively reduced to provide a final safety net for all control behaviors.

[0064] The aforementioned hardware computing power security threshold is the upper boundary of hardware resource security for steady-state operation of the server. The hardware computing power security threshold = upper limit of steady-state security utilization rate - reserved computing power margin; where the upper limit of steady-state security utilization rate is the rated utilization rate of the server CPU for long-term stable operation, which is usually taken as 70% to 80%; the reserved computing power margin is configured at 10% to 20% of the total computing power of the machine to provide a safety net for instantaneous traffic surges. The hardware computing power security threshold is fixed with the server hardware configuration or updated iteratively according to monthly operation and maintenance parameters.

[0065] (4) Terminal-level adaptive traffic control. Based on the coordinated control of the server-side receiving and processing modules, a client-side traffic control mechanism is added to mitigate the impact of instantaneous peak traffic at the source, achieving joint control of front-end traffic and back-end computing power between the terminal and the cloud. Each candidate's terminal synchronizes the server's busy / idle status and exam schedule with the server, dynamically and adaptively adjusting the reporting interval, single-upload question granularity, and local cache reporting threshold based on the local cache question granularity load.

[0066] The busy / idle state is predicted by the server based on the guaranteed reception volume Rq and the maximum carrying reception load S. When the ratio of the guaranteed reception volume to the maximum carrying reception load is greater than an empirical threshold, the server is in a busy state.

[0067] When the remaining time for the exam is less than a given threshold or the server is busy, the reporting interval is lengthened, the granularity of a single upload is increased, and the cache reporting threshold is raised to achieve peak traffic reduction.

[0068] The control logic differentiates between busy and idle working conditions for adaptive adaptation: In the early stages of the exam, when the server is idle and the overall network load is stable, the terminal adopts the regular reporting interval and standard upload granularity to ensure the real-time data upload; at the end of the exam, when the server is under high load and the overall network concurrency pressure surges, the terminal automatically lengthens the reporting interval, increases the single upload granularity, and raises the local cache reporting threshold. By staggering the upload times, it suppresses the synchronous centralized submission behavior of massive terminals, achieves peak traffic shaping at the source, and significantly reduces the instantaneous concurrency pressure on the server.

[0069] S5. Full-process online iterative optimization and post-exam update. During the exam, the load sample and computing power mapping data are updated in real time to ensure that the lookup benchmark is in line with the changes in working conditions. After a single exam, the remaining data in the queue is processed, and the entire running data is summarized to iteratively update the question time parameters, time-series working condition data, and dynamic mapping table data, so as to continuously optimize the control accuracy of the exam system.

[0070] This invention also provides an online examination intelligent control system based on question granularity, used to execute the above-mentioned method. This system includes a layered architecture module, a question granularity load quantification module, a dynamic computing power mapping module, a pressure prediction and risk correction module, a receiving and processing module linkage control module, a terminal traffic control module, and a full-process iterative optimization module. These modules work collaboratively to support a complete intelligent control closed loop, from high-concurrency data reception and precise computing power allocation to peak shaving at the terminal source.

[0071] The layered architecture is used to deploy an asynchronous and decoupled infrastructure for receiving and processing data. It separates high-priority receiving modules from flexible and adjustable processing modules, and uses message queues to achieve asynchronous isolation of the data receiving and processing processes, supporting differentiated and prioritized scheduling of the receiving and processing modules.

[0072] The question-granular load quantization module is used to collect and iteratively update the average question reception time and average question processing time, and to statistically analyze the average CPU utilization, number of active concurrent terminals, reception load, processing load, queue waiting backlog load and terminal cache load in real time. Based on question-granular weighted quantization, a unified load evaluation system is formed to provide accurate data support for subsequent pressure assessment and resource scheduling.

[0073] The dynamic computing power mapping module is used to construct and iterate the dynamic mapping table of problem-level load and hardware computing power in real time. It completes the load ratio coefficient and computing power benchmark calibration for extreme and normal operating conditions according to preset boundary rules. It completes the table lookup matching for operating conditions and updates the computing power benchmark calibration through the minimum distance matching principle, realizing the normalized and quantized mapping of abstract business load and hardware computing power. The minimum distance can be either Euclidean distance or Manhattan distance, with Euclidean distance being preferred.

[0074] The pressure prediction and risk correction module is used to calculate the global receiving load congestion. It combines the remaining exam time and the number of active concurrent terminals in the network to generate a progress correction coefficient. The combined coefficients are used to obtain a dynamic risk coefficient that adapts to the characteristics of the end-of-term flood peak, thereby realizing load pressure quantification, time-series risk correction, and flood peak risk prediction in advance.

[0075] The receiving and processing linkage control module is the core scheduling module, used to complete the receiving module's operating condition calibration, mapping table iterative update, and guarantee volume Rq calculation and locking. At the same time, it completes the computing power trend prediction based on the predicted load ratio lookup table, accurately determines the maximum allowable processing load volume through the quantitative reverse retrieval mechanism, controls the adjustment range of processing threads, and completes hardware safety fallback control in combination with real-time smooth CPU utilization, realizing intelligent linkage scheduling of receiving and processing modules with receiving priority and flexible processing adaptation.

[0076] The terminal traffic control module is deployed on each candidate's terminal to synchronize the server's busy / idle status with the exam schedule. It adaptively adjusts the data reporting interval, the granularity of a single uploaded question, and the local cache reporting threshold based on the local cache load, thereby achieving source traffic control that ensures real-time performance during idle periods and reduces peak traffic during busy periods.

[0077] The full-process iterative optimization module is used to update the load sample and mapping table data in real time during the examination. After the examination, it summarizes all the running data and iteratively updates the question time parameters, time sequence parameters and dynamic mapping table to achieve continuous optimization of the examination system's control accuracy.

[0078] This invention discloses an intelligent control method and system for online examinations based on question granularity, which is applicable to large-scale, high-concurrency online examination scenarios.

[0079] This invention adopts an edge-cloud collaborative layered architecture. The server-side employs an asynchronous, decoupled architecture for receiving and processing, separating the receiving and processing modules. These two modules are asynchronously isolated through message queues, forming a differentiated priority scheduling system. The receiving module, as a high-priority guarantee module, is responsible for receiving, caching, and verifying the integrity of exam answer data. It can reserve dedicated computing power and caching resources based on the global load situation, ensuring stable data upload without loss or timeout during peak scenarios. The processing module, as a flexible and adjustable module, is responsible for data parsing, business logic verification, and database storage. It can dynamically adjust the number of worker threads and computing power allocation based on real-time load conditions and historical steady-state data, providing an architectural foundation for fine-grained computing power scheduling. This overall architecture and edge-cloud collaborative control logic correspond to the overall technical framework of this invention and are adapted to the entire system's operational logic.

[0080] This invention constructs a unified load quantification system at the question granularity level, abandoning the traditional method of uniform load statistics for the entire system. It uses the actual transmission and processing times of different questions as benchmarks to quantify the differentiated resource consumption of examination business data, thus unifying the load evaluation scale of the examination system. It collects and iteratively updates multi-dimensional core operational indicators in real time, including the average question reception time T. R Average processing time T for each question P The metrics include server-side average CPU utilization (C), active concurrent terminals (N), receiving load (R), processing load (P), queue backlog load (Q), and terminal cache load (B). The question duration parameter is updated using an iterative average method to ensure continuous convergence of the question's computational power consumption characteristics. CPU utilization is smoothed and denoised using a time window to eliminate instantaneous sampling jitter. All load metrics are obtained by weighted summation of the corresponding question durations based on the number of questions. This comprehensive approach, encompassing the server's receiving side, processing side, queue backlog side, and terminal pre-upload side, comprehensively characterizes instantaneous load and potential peak pressure, achieving refined load quantification modeling.

[0081] Based on load quantization, this invention constructs a dynamic mapping table of question-granular load and hardware computing power that can be updated in real time, solving the technical problem that abstract business load cannot be directly mapped to hardware computing power occupancy. The mapping table uses the load matching coefficient K as the retrieval index and the computing power benchmark CP as the retrieval benchmark. K To achieve unified computing power across different load volumes under the same load ratio, this invention, based on historical steady-state conditions from online examinations, obtains a load fluctuation tolerance threshold Kmax, defines an effective lookup table interval, and sets specific boundary calibration rules for four types of operating conditions: no load, extreme load imbalance, and normal balanced load. This completely avoids lookup table failures and computing power benchmark distortion caused by extreme operating conditions. Simultaneously, the mapping table employs an online iterative update mechanism, performing minimum distance matching lookups based on the load ratio coefficient after limit calibration. Real-time calibration updates are completed by combining historical computing power benchmarks with the current average computing power mapping value, ensuring that the computing power mapping benchmark always closely matches real-time operating condition changes.

[0082] This invention constructs a multi-dimensional flood risk prediction and dynamic correction model to achieve advance prediction and resource reservation for the final exam submission flood. Using the server's maximum global receiving load as the upper limit benchmark, this invention calculates the global receiving load congestion by combining the receiving load with the terminal cache's potential load, simultaneously characterizing the current instantaneous pressure and short-term pre-burst pressure of the examination system. Simultaneously, it generates a basic correction increment based on the temporal relationship between the remaining exam time and the total exam time, and performs differentiated fine-tuning according to the number of active concurrent terminals across the network, generating a dimensionless progress correction coefficient. The closer to the end of the exam and the larger the scale of concurrent terminals, the higher the temporal risk correction amplitude. The progress correction coefficient and the global load congestion are combined to obtain a composite risk coefficient, enabling scheduling decisions to simultaneously integrate multi-dimensional characteristics such as load volume, temporal risk, and concurrent impact intensity, overcoming the technical shortcomings of traditional solutions that rely solely on real-time load and lack advance prediction capabilities.

[0083] This invention employs a coordinated control mechanism between the receiving and processing modules, prioritizing reception and flexibly adapting to processing, to achieve refined collaborative scheduling of computing resources. The receiving module has the highest scheduling priority, independent of instantaneous CPU load fluctuations. In each control cycle, it prioritizes completing operational condition calibration and iterative updates to the mapping table. Based on a composite risk coefficient, it calculates the unallocated capacity of the receiving module, defining resource boundaries for processing module control. While ensuring sufficient resource reservation for the receiving module, the processing module implements flexible adaptive computing power control. It uses a quantitative reverse retrieval mechanism to match computing power benchmarks that meet computing power security requirements, accurately obtaining the maximum allowable processing load and dynamically adjusting the processing thread scale. Simultaneously, it uses real-time smoothed and noise-reduced CPU utilization as the final hardware security safety baseline, forming a dual-constraint control logic of predicted computing power compliance and actual hardware security. This ensures the stability of the uploading module while maximizing the utilization of idle computing resources and improving overall machine resource utilization.

[0084] This invention establishes a terminal-based adaptive traffic control mechanism to achieve peak shaving at the source. It differentiates the traffic control logic based on different load characteristics throughout the entire examination cycle: In the early stages of the examination, when the server is idle and the overall network load is stable, terminals use a standard reporting interval and upload granularity to ensure real-time synchronization of answer data; towards the end of the examination, when the server is under high load and the overall network concurrency pressure surges, terminals automatically lengthen the reporting interval, increase the granularity of a single upload, and raise the local cache reporting threshold. By staggering upload times and smoothing the traffic rhythm, it suppresses the simultaneous centralized submission behavior of massive terminals, achieving peak traffic shaving at the source and significantly reducing the instantaneous concurrent pressure on the server.

[0085] This invention establishes a full-cycle continuous iterative optimization mechanism, which updates the load sample and computing power mapping data in real time during the examination to ensure that the lookup benchmark keeps up with the changes in working conditions. After a single examination, the remaining data in the queue is processed, and the entire running data is summarized to iteratively update the question time parameters, time-series working condition data, and dynamic mapping table data, thereby achieving continuous optimization of the examination system.

[0086] To further illustrate the practical application effect of the technical solution of this invention, this specific embodiment selects an online computer-based examination scenario with a scale of 20,000 participants. The total examination time is set to 60 minutes, with N=20,000 active candidate terminals. The examination paper includes 20 multiple-choice questions QA1-QA20, 3 open-ended questions QB1, QB2, and QB3, and 2 practical questions QC1 and QC2. After iterative convergence of historical question submission data, the steady-state baseline time consumption parameters for each type of question are as follows: average reception time per multiple-choice question. All times are 3ms, and the average processing time per question is 6ms; the average reception time per question for QB1, QB2, and QB3 is... The average processing time per question was 7ms, 8ms, and 11ms, respectively. The average reception times for operation questions QC1 and QC2 are 15ms, 16ms, and 20ms respectively; the average processing times for each question are 60ms and 92ms respectively. The examination system has a preset load fluctuation tolerance threshold Kmax=2000, a maximum network request timeout of 30s, and a computing power baseline of 0.0100% corresponding to a load ratio coefficient of 1 / 2000 in the dynamic mapping table. The server-side hardware security CPU utilization threshold is set to 75%. Therefore, the maximum global load capacity S of the examination system is 300000.

[0087] The exam is running smoothly. Real-time data on the entire network's operational status is being collected. Currently, there are 18,620 active online terminals, and the number of multiple-choice questions received in real-time is [number missing]. There are 60 questions in total, with the number of short answer questions counted as... The questions are 25, 35, and 55 respectively, with the number of practical questions counted as follows: The numbers are 115 and 135 respectively; the number of multiple-choice questions processed in real time is counted as... There are 50 questions in total, with the number of short answer questions counted as... There are 5 questions each, and 5 questions in total. The number of practical questions is counted as follows: There are 6 questions each; the number of multiple-choice questions cached on the terminal is counted as... There are a total of 280 questions, with the number of short answer questions counted as... The numbers are 125, 180, and 210 respectively, with the number of practical questions counted as follows: The numbers are 326 and 252 respectively; the number of multiple-choice questions with backlog is counted as... There are a total of 1600 questions, with the number of short answer questions counted as... The numbers are 2950, ​​3110, and 4100 respectively, with the number of practical questions counted as... The questions were 4130 and 5050 respectively. After time-window smoothing and noise reduction to eliminate instantaneous sampling jitter, the average CPU utilization of the server was C=73.2%. Using a question-granular weighted quantization method, combining the number of questions of each type with the corresponding baseline time, a multi-dimensional refined load index was calculated, where the receiving load R, processing load P, terminal buffer load B, and queue backlog load Q were respectively:

[0088]

[0089] = 60*3 + 25*7 + 35*8 + 55*11 + 115*15 + 135*20 = 5665

[0090]

[0091] = 50*6 + 5*15 + 5*16 + 5*20 + 6*60 + 6*92 = 1467

[0092]

[0093] = 280*3 + 125*7 + 180*8 + 210*11 + 326*15 + 252*20 = 15395

[0094]

[0095] = 1600*6 + 2950*15 + 3110*16 + 4100*20 + 4130*60 + 5050*92 = 898010

[0096] Based on the real-time load data obtained through quantization, the current load ratio coefficient K = P / R ≈ 0.258959 is calculated. This value indicates that the current system is under normal balanced load conditions. Under this condition, the current computing power baseline CP is obtained. K =C / R≈0.0129%, and the mapping table is updated online by combining historical steady-state data to correct the computing power matching benchmark.

[0097] Further calculations were performed to predict and correct peak flood risk. Based on real-time load data, the global load congestion was calculated to be approximately 0.070 (R+B) / S. With 10 minutes remaining in the current exam, and considering the exam schedule and the number of active terminals across the network, a dimensionless progress correction coefficient of 1.5*18620 / 10000=2.793 was generated. The composite peak flood risk coefficient under the current conditions was calculated to be approximately 0.070*2.793≈0.196. Based on the global receiving load congestion and the composite risk coefficient, the predicted receiving capacity was Rq=R×1.196≈6775.

[0098] Based on the received capacity Rq and the processing load P, a standardized load ratio Kq = 1467 / 6775 = 0.216531 is constructed. The mapping table is searched to obtain CPq = 0.0123%. Based on CPq, the average CPU utilization Cq is deduced to be 83.1%. At this point, Cq > the hardware computing power safety threshold of 75%. The mapping table is searched in reverse order of decreasing load ratio to match the computing power benchmark CPnew < 75% / R that meets the computing power safety requirements, resulting in a suitable load ratio Knew = 0.100664. The maximum allowable processing load Pnew ≈ 682 is accurately obtained. The processing capacity and thread computing power usage are iteratively reduced to ensure that the processing load P does not exceed Pnew. The current average CPU utilization Ch = C = 73.2% is used as the hardware safety benchmark. Since Ch is within the hardware computing power safety threshold, no adjustment is made.

[0099] This invention achieves traffic peak shaving through a dual-state adaptive mechanism for busy and idle terminals. Under stable conditions in the early stages of the exam, terminals adopt a conventional reporting strategy with a data reporting interval of 5 seconds, uploading one question's answer data at a time to ensure real-time synchronization of examinee's answer data. Under high-load and high-risk conditions at the end of the exam, terminals automatically switch to peak shaving mode, proactively extending the data reporting interval to 15 seconds, raising the local cache threshold, and uploading three question data in batches at a time. By accumulating data locally and uploading in batches, the invention avoids the surge impact caused by a massive number of terminals submitting data at once.

[0100] This invention establishes a full-cycle continuous iterative optimization mechanism, which updates the load sample and computing power mapping data in real time during the examination to ensure that the lookup benchmark keeps up with the changes in working conditions. After a single examination, the remaining data in the queue is processed, and the entire running data is summarized to iteratively update the question time parameters, time-series working condition data, and dynamic mapping table data to achieve continuous system optimization.

[0101] Those skilled in the art will understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for intelligent control of online examinations based on question granularity, characterized in that, Includes the following steps: S1. Deploy an asynchronous decoupled infrastructure, splitting the server into independent receiving and processing modules. The receiving and processing modules are asynchronously isolated through message queues. The receiving module, as a high-priority module, is responsible for receiving, caching, and verifying the integrity of data, and reserves computing resources to cope with peak pressure based on load conditions. The processing module, as a flexible and adjustable module, is responsible for data parsing, logical verification, and database storage, and dynamically adjusts the number of threads and computing power ratio based on real-time load conditions and historical operating data. S2. Construct a unified load quantization system based on question granularity, and use the unified load quantization system to quantify and iteratively update multi-dimensional core performance indicators. The specific method is as follows: Obtain the actual reception time, the actual processing time, the real-time CPU utilization, and the number of currently online terminals; Using the average reception time and average processing time of each question as weighted benchmarks, a multi-dimensional core operational indicator is quantified. This multi-dimensional core operational indicator includes: average reception time T for each question. R Average processing time T for each question P The following metrics are measured: CPU average utilization C, concurrent terminal count N, receive load R, process load P, waiting backlog load Q, and terminal cache load B. Among them, the average reception time T for the question R The average historical reception time and the average of the current actual reception time are used. For the first run, the actual reception time of this run is used. Average processing time T for each problem P The average historical processing time and the average actual processing time are used, and the actual processing time is used for the first run. The average CPU utilization C is obtained by smoothing and reducing the noise of the real-time CPU utilization through a fixed time window. The receiving load R is the weighted sum of the number of questions received in real time and the average receiving time for the corresponding questions. The processing load P is the weighted sum of the number of questions currently being processed and the average processing time for the corresponding questions. The waiting backlog load Q is the weighted sum of the number of queued questions in the message queue and the average processing time of the corresponding questions. The terminal cache load B is the weighted sum of the number of cached questions that the terminal has not uploaded and the average reception time of the corresponding questions, and is used to predict the short-term uplink peak pressure. S3. Construct and iterate in real time a dynamic mapping table of problem granularity load and hardware computing power. The dynamic mapping table uses the load ratio coefficient as the index dimension and the computing power benchmark as the storage dimension to achieve normalization and unification of computing power for different sizes and the same ratio working conditions. Perform minimum distance matching lookup based on the load ratio coefficient, and combine the historical computing power benchmark with the current computing power mapping average to complete the real-time iterative update of the dynamic mapping table. S4. Based on the server's global carrying capacity, quantify the real-time load pressure and dynamically correct the risk of the final flood by combining the time sequence characteristics of the exam progress. Based on the quantitative computing power benchmark of the dynamic mapping table of question granular load and hardware computing power, construct a linkage control mechanism between the receiving and processing modules. Constrain the scheduling behavior of the processing module with the priority principle of the receiving module to realize the collaborative cooperation and dynamic adaptation of the two modules, and complete the risk prediction and dynamic resource matching in high-concurrency exam scenarios. S5. During the exam, the load sample and mapping data are updated in real time. After the exam ends, the unprocessed answer data in the processing module is completed, and the question time parameters, time sequence data and dynamic mapping table data are updated iteratively.

2. The online examination intelligent control method based on question granularity according to claim 1, characterized in that, The specific method for constructing the dynamic mapping table of problem granularity load and hardware computing power is as follows: The load fluctuation tolerance threshold Kmax is preset, the effective index range [0, Kmax] is defined, and the value range of the load ratio system is standardized within the effective index range according to the boundary calibration rules of the load ratio coefficient and the computing power benchmark. The boundary calibration rule is specifically as follows: When there is neither receiving load nor processing load, take K=0, CP K =C; When the processing load is greater than the receiving load or there is no receiving load, take K=Kmax, CP K =C / P; When the processing load is less than the receiving load or there is no processing load, take K = 1 / Kmax, CP K =C / R; When there is both receiving and processing loads, take K=P / R, CP K =C / R; In the formula, K is the load ratio coefficient, and CP K Here, C is the computing power benchmark, P is the average CPU utilization, and R is the processing load.

3. The online examination intelligent control method based on question granularity according to claim 1, characterized in that, The specific method for performing minimum distance matching lookup based on the load ratio coefficient K, and combining historical computing power benchmarks with the current average computing power mapping to complete the real-time iterative update of the dynamic mapping table is as follows: Using the load ratio coefficient K as the unique retrieval index, the historical index node with the highest fit to the current working condition is selected by the minimum distance matching principle, and the corresponding historical computing power benchmark data is read. The average value of the historical computing power benchmark value and the current computing power mapping field is used as the calibration of the computing power benchmark value, and the value is updated to the mapping table in real time. The minimum distance can be either Euclidean distance or Manhattan distance.

4. The online examination intelligent control method based on question granularity according to claim 1, characterized in that, The specific implementation method of step 4 is as follows: The global receiver load congestion is calculated by combining the receiver load R, which represents instantaneous pressure, and the terminal buffer load B, which represents short-term potential pressure. The specific formula is as follows: Global receiver load congestion = (R+B) / S; In the formula, S is the global maximum received load; The basic correction increment is generated based on the proportion of exam duration, and the progress correction coefficient is obtained by adjusting the basic correction increment in combination with the number of concurrent terminals N. The progress correction coefficient is multiplied by the global receiving load congestion to obtain the composite risk coefficient. Based on the quantified receiving load pressure and the corrected peak risk, a linkage control logic for the receiving and processing modules is constructed by combining a dynamic computing power mapping table: according to the real-time acquired receiving load and processing load, the hardware computing power characteristics calibration and dynamic mapping table iterative update are completed under the current load ratio according to the dynamic mapping table lookup matching and iterative update rules; the receiving guarantee volume is calculated based on the receiving load and composite risk coefficient. Based on the received capacity and processing load, a standardized load ratio is constructed. The dynamic mapping table is retrieved according to the load ratio to obtain the computing power benchmark mapping value. The CPU utilization is predicted based on the benchmark mapping value. The CPU utilization is compared with the hardware computing power safety threshold to determine the direction of adjustment. When the CPU utilization exceeds the hardware computing power safety threshold, the mapping table is retrieved in reverse along the direction of load ratio reduction to match the computing power benchmark that meets the computing power safety requirements, and the maximum allowable processing load capacity Pnew is obtained to adjust the reduction of the thread scale of the processing module. When the CPU utilization is within the hardware computing power safety threshold, the mapping table is searched in reverse along the direction of increasing load ratio to match the computing power benchmark that meets the computing power safety requirements, and the maximum allowable processing load Pnew is obtained to adjust the extent to which the processing module expands the thread scale. By adjusting the reporting interval, single-upload question granularity, and local cache reporting threshold by combining the busy / idle status and exam schedule of the candidate's terminal, source traffic control can be achieved to ensure real-time operation during idle periods and to stagger peak traffic during busy periods.

5. The online examination intelligent control method based on question granularity according to claim 4, characterized in that, The global maximum receive load S is specifically: S = Kmax * Tout / CP K1 ; In the formula, Tout is the maximum network request timeout allowed by the examination system, Kmax is the load fluctuation tolerance threshold, and CP K1 This is the computing power benchmark corresponding to extreme operating conditions in the dynamic mapping table.

6. The online examination intelligent control method based on question granularity according to claim 4, characterized in that, The progress correction factor is the ratio of the basic correction increment multiplied by the number of concurrent terminals N to the maximum number of concurrent terminals set on the server.

7. The online examination intelligent control method based on question granularity according to claim 4, characterized in that, Hardware computing power security threshold = upper limit of steady-state security utilization rate - reserved computing power margin.

8. The online examination intelligent control method based on question granularity according to claim 4, characterized in that, The busy / idle state is determined by the server based on a comparison between the ratio of the guaranteed reception volume to the maximum reception load and a given empirical threshold. When the ratio of the guaranteed reception volume to the maximum reception load is greater than the empirical threshold, the server is in a busy state. When the remaining time for the exam is less than a given threshold or the server is busy, the reporting interval is lengthened, the granularity of a single upload is increased, and the cache reporting threshold is raised to achieve peak traffic reduction.

9. An online examination intelligent control system based on question granularity, characterized in that, For performing the method according to any one of claims 1-8, comprising: The architecture is layered and modular, used to deploy an asynchronous decoupled architecture between the receiving module and the processing module. A message queue is used to achieve asynchronous isolation and priority scheduling between the receiving module and the processing module. The question-granular load quantification module is used to quantify and iteratively update multi-dimensional core operating indicators, including: obtaining the actual reception time, the actual processing time, the real-time CPU utilization, and the number of current online terminals; quantifying CPU utilization, the number of concurrent terminals, reception load, processing load, queue backlog load, and terminal cache load; and constructing a unified question-granular load evaluation system. The dynamic computing power mapping module is used to construct, calibrate the boundaries, and iterate the problem-granularity load-hardware computing power dynamic mapping table in real time. It achieves working condition lookup matching and computing power benchmark calibration update through minimum distance matching. The pressure prediction and risk correction module is used to calculate the global receiving load congestion, generate the progress correction coefficient and combine it to obtain the dynamic risk coefficient, so as to realize the timing risk correction and the early prediction of flood peak risk. The terminal traffic control module, deployed on the candidate's terminal, is used to adaptively adjust the reporting interval, upload granularity, and cached reporting threshold to achieve peak reduction at the source. The receiving and processing module linkage control module is used to complete the receiving module's operating condition calibration, mapping table iteration, and receiving guarantee volume locking, and to realize the processing module's computing power amplitude adjustment and hardware security control through quantitative reverse retrieval. The end-to-end iterative optimization module is used to iteratively update the load and mapping data in real time during the exam, and to iteratively optimize the question parameters and mapping table parameters across the entire domain after the exam.

Images