A cloud computing-based competition management system and method

Abstract

The application relates to a cloud computing-based competition management system and method, and belongs to the technical field of resource allocation. The number of virtual machines, physical resources and competition levels are determined by dynamically obtaining competition information, and the resource priority is automatically associated to guarantee stable allocation of high-level competition resource, a scheduling model is constructed by using a deep reinforcement learning algorithm, a virtual machine allocation probability matrix is generated, long-term benefits are evaluated by combining a reward function, and autonomous learning optimization of a scheduling strategy is realized; a multi-dimensional server state matrix is constructed, a matching task and hardware characteristics are quantified, and task interruption risk is reduced; a dynamic label system is generated by combining static code analysis and runtime data, and key event resource allocation is ensured; competition projects are phased and resource demand vectors are defined, and phased dynamic resource adjustment is realized. The application significantly improves resource utilization and load balancing capability, and improves the fairness, system stability and intelligent level of competition management.

Description

TECHNICAL FIELD

[0001] The application relates to a cloud computing-based competition management system and method, and belongs to the technical field of resource allocation. BACKGROUND

[0002] With the rapid development of big data, artificial intelligence and other technologies, the competition mode based on cloud computing has become the mainstream choice for education, scientific research and industry competitions due to its efficiency and flexibility. Such a system needs to allocate cloud computing resources (such as virtual machines, physical server computing power, storage, etc.) reasonably according to the competition task requirements, so as to guarantee the fairness, stability and resource utilization efficiency of the competition.

[0003] Traditional competition management systems mostly adopt static resource allocation strategies, such as deployment modes based on virtual machines or fixed servers. When multiple teams are online at the same time, there is a lack of dynamic perception ability for real-time load (such as CPU, memory, network bandwidth utilization), which easily leads to resource contention or waste. For example, when high-level and low-level competitions are concentrated in the same period, the traditional system cannot prioritize the resource requirements of high-level competitions, which may lead to task interruption due to insufficient resources; and in low-load periods, there are problems such as excessive activation of servers and idle resources, and only relying on basic resource utilization such as CPU and memory for scheduling ignores the influence of hardware heterogeneity (such as GPU computing power and disk performance differences) on task execution efficiency. For example, GPU-intensive AI competition tasks and memory-intensive big data processing tasks may lead to reduced computing efficiency or waste of hardware resources if they are allocated to mismatched servers, and there is a lack of standardized modeling and dynamic adaptation ability for phased resource requirements, which leads to traffic shock (such as network congestion caused by simultaneous migration of a large number of virtual machines) or resource allocation lag during phase switching. SUMMARY

[0004] The application provides a cloud computing-based competition management system and method to solve the problems of rigid resource allocation, inability to dynamically adapt to load changes, insufficient hardware heterogeneity matching, low task adaptation efficiency and insufficient intelligent scheduling ability in the prior art.

[0005] The application provides a cloud computing-based competition management system and method, which includes determining the required number of virtual machines, physical resources and competition levels according to competition information, automatically associating resource priorities according to competition levels, dynamically adjusting resource allocation, ensuring stable allocation of resources for high-level competitions, constructing a deep deterministic policy gradient algorithm model, generating a virtual machine allocation probability matrix, realizing autonomous learning and optimization of scheduling strategies, and guiding the model to learn the optimal scheduling strategy through a reward function to evaluate the long-term reward value of the current state-action pair.

[0006] A multi-dimensional server state matrix is constructed to deeply match tasks with server hardware characteristics, potential resource contention is predicted through memory capacity normalization values, the risk of task interruption caused by insufficient memory is reduced, tags are generated through static code scanning and runtime data, and the success of resource allocation for key event tasks is ensured through competition level associated tags.

[0007] The competition project is divided into multiple stages, and a resource demand vector is defined for each stage to form a standardized template. Resource allocation can be dynamically adjusted according to the needs of different stages.

[0008] Preferably, by real-time monitoring of student experiment progress and resource occupation, dynamic adjustment of resource allocation is realized, and the total amount of virtual resources required by different events in the same period is controlled within the set threshold range. Through competition time management, the total amount of virtual resources required in the same period can be controlled within the bearing range of the cloud computing center, and the competition can be staggered.

[0009] Preferably, a virtual machine allocation probability matrix is used to realize autonomous learning and optimization of scheduling strategies through end-to-end training. In high-load scenarios, particle swarm optimization algorithm is used for local optimization to realize the cooperation of global optimization and local optimization.

[0010] Preferably, the GPU computing power index, memory capacity normalization value, and disk performance level are converted into quantifiable parameters to construct a multi-dimensional server state matrix. Combined with static code scanning and runtime data, three levels of tags are generated, including basic tags, priority tags, and dynamic composite tags.

[0011] The server is monitored to generate a hardware basic portrait. The hardware basic portrait and real-time monitoring data generate a time-sequenced server portrait, reflecting the instantaneous state of the server.

[0012] Preferably, a two-dimensional calculation model of basic resource matching degree and priority matching degree is constructed to avoid long-term starvation of low-priority tasks through priority weighted matching.

[0013] Time-adjacent samples are generated through a sliding window, and high-priority task data is oversampled.

[0014] Preferably, a dynamic stage portrait matrix is constructed, and the resource allocation strategy is dynamically adjusted through a stage gating unit to reflect the state of the server in real time and provide accurate basis for resource allocation. The stage gating unit dynamically adjusts the resource allocation strategy according to the task stage tag.

[0015] Preferably, the phase switching event is triggered by timing scanning all competition tasks, the resource requirement template of the target phase is extracted, the resource capacity lower limit is extracted, the server suitable for the target phase is selected based on the dynamic phase image of the currently activated server, the total sum of the remaining resources of the candidate server is calculated and compared with the required resources in the resource requirement template of the target phase, the deviation between the actual resource consumption of each task in each phase and the expected deviation of the template is recorded to form a phase resource deviation matrix, the model is adjusted to guide the allocation weight, and the phase template library is updated.

[0016] Preferably, each competition task is divided, the team ID and member role label are bound to each team member task, the cross-task phase dependency relationship is established, and the manual coordination cost is reduced through the role priority and phase order constraint.

[0017] Preferably, the reward function of the reinforcement learning model is reconstructed, the team phase cooperation degree, the resource efficiency priority weight, the cooperative optimization weight and the delay penalty weight are introduced, the dual optimization of resource allocation and team cooperation is realized through multi-objective weight dynamic adjustment, the historical phase template library is loaded as prior knowledge to avoid model shock caused by dynamic changes of competition tasks, and the phase deployment deviation matrix is recorded to be retrained with annual data to adapt to long-term changes.

[0018] A competition management system based on cloud computing, comprising:

[0019] A supervisory control and data acquisition module is used to provide real-time resource state data, provide input for dynamic scheduling and model training, and support anomaly detection.

[0020] A data processing and feature engineering module is used to clean and structure the original data, and generate feature vectors suitable for model input.

[0021] A reinforcement learning model module is used to realize intelligent resource allocation, optimize the scheduling strategy through historical data training, and balance resource utilization rate and server activation number.

[0022] A load scenario scheduling module is used to dynamically adjust the strategy for different load scenarios, improve resource utilization efficiency, and guarantee system stability and resource demand of high-priority tasks.

[0023] A task label and matching module is used to solve the problem of matching hardware heterogeneity and real-time task demand, preferentially allocate GPU-intensive tasks to suitable servers, and reduce resource contention risk.

[0024] A phased resource management module is used to accurately match resource requirements of different phases, improve resource allocation flexibility, and reduce traffic impact during phase switching.

[0025] Data storage and management module: used for providing high-quality data set for model training, guaranteeing time sequence feature integrity and improving model generalization ability.

[0026] The beneficial effects of the present application are:

[0027] The application provides a competition management system and method based on cloud computing, which divides the competition project into multiple stages, defines a resource demand vector for each stage, and forms a standardized template. A dynamic stage portrait matrix is constructed, and the resource allocation strategy is dynamically adjusted through a stage gating unit to improve the flexibility and efficiency of resource management, ensure that tasks obtain sufficient resource support in different stages, bind role tags for team members, and establish cross-task stage dependency relationships to reduce manual coordination costs and avoid rework caused by task conflicts. The reward function of the reinforcement learning model is reconstructed, and team stage collaboration degree, resource efficiency priority weight, etc. are introduced to realize dual optimization of resource allocation and team collaboration. After randomly initializing the strategy network and the value network, the historical stage template library is loaded as prior knowledge, and soft updating is used to avoid model oscillation. The stage deployment deviation matrix is recorded, and the model fine-tuning is triggered by using high-frequency conflict scene data to adapt to long-term changes and ensure the continuous optimization of the model. A sliding window is used to generate a multi-task stage sequence, and the key stage data of high-level competitions is oversampled to avoid model bias caused by insufficient data in high-difficulty stages. The adaptability of the model to stage switching is enhanced, the generalization ability of the model is improved, and the stable operation of the system in different scenarios is ensured. BRIEF DESCRIPTION OF DRAWINGS

[0028] Figure 1 The application provides a competition management system and method based on cloud computing. DETAILED DESCRIPTION

[0029] The technical solutions in the embodiments of the present application will be described clearly and completely in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0030] In the description of the present application, the terms "first", "second" are used only for the purpose of description, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of indicated technical features. Therefore, the features defined as "first", "second" can explicitly or implicitly include one or more of the features. In the description of the present application, the meaning of "multiple" is two or more, unless otherwise specifically limited.

[0031] In the description of the present application, the term "for example" is used to mean "serving as an example, instance, or illustration." Any embodiment described as "for example" in this application is not necessarily to be construed as preferred or advantageous over other embodiments. The following description is presented to enable any person skilled in the art to make and use the application. In the following description, for purposes of explanation, specific details are set forth. It will be apparent to those skilled in the art that the present application can be practiced without the specific details. In other instances, well-known structures and processes are not shown in detail to avoid obscuring the application. Thus, the present application is not intended to be limited to the embodiments shown, but is to be accorded the widest scope consistent with the principles and features presented herein.

[0032] Embodiment 1

[0033] The present application provides a cloud computing-based competition management method, comprising:

[0034] The competition project, the time of the competition item, the number of participants of each competition item are obtained, the number of virtual machines required by students, physical resources and competition levels are determined according to the competition project, wherein the physical resources are composed of four types of resources, specifically including cpu, memory, disk and network resources, a virtual resource demand table based on the time of the competition item is established, the total amount of physical resources of the cloud computing center is obtained, it is judged whether the total amount of virtual resources required by different competition items in the same period exceeds the set threshold, if yes, the competition item time is adjusted, and the total amount of virtual resources required by different competition items in the same period is controlled within the set threshold range. Through competition time management, the total amount of virtual resources required in the same period can be controlled within the bearable range of the cloud computing center, the competition is staggered, a reliable competition environment is provided for various competitions, and the historical scheduling scheme is recorded;

[0035] The creation, destruction, suspension, recovery and one-key recycling of various experimental clusters required for the competition are provided, and diversified cluster experimental simulation environments are ensured;

[0036] The experiment progress and resource occupation of each student, the resource occupation of each physical server and each virtual machine are listened to, the corresponding resource occupation curves are drawn, various curves are graphically output, and the listening results are fed back

[0037] The CPU utilization rate, memory utilization rate, disk IO utilization rate, network bandwidth utilization rate, server activation state physical server ID are sampled, and the virtual machine belongs to the competition item ID, virtual machine CPU utilization rate, virtual machine memory utilization rate, virtual machine disk IO utilization rate, virtual machine network bandwidth utilization rate, competition level are associated, the competition item start and end time, the number of participating students, the number of virtual machines applied by each student, the actual allocated physical resources, the scheduling results are collected;

[0038] The collected data is cleaned, wherein a single record with >30% missing fields is directly discarded, a single field missing is filled with the average value of the data 5 minutes before and after, feature engineering and state vector construction are performed, specifically: time window aggregation is performed in units of 1 hour, 4 types of resource utilization rate and activation state are extracted to splice into a complete state vector , according to the 4 types of resource utilization rate and activation state output matrix, the dimension is , wherein is the total number of physical servers in the data center, and the 5 columns are 4 types of resource utilization rate + activation state;

[0039] According to the type of the event, the total number of virtual machines required in the window, the virtual machine proportion of each type of cluster, and the average competition level are counted, and the output vector dimension is , wherein is the number of experimental cluster types, and the 2 columns are the total number of virtual machines + the average priority;

[0040] The final input dimension is: ;

[0041] According to the historical scheduling scheme as the initial action label, a virtual machine allocation matrix is established The dimension is , wherein is the total number of virtual machines in the window, and the element indicates that the virtual machine is allocated to the physical server , otherwise 0;

[0042] The data set is divided and enhanced, and the training set and the test set are divided according to the ratio of 8:2, ensuring that the time sequence does not change, wherein the training set contains historical scheduling data for 12 consecutive months, each sample corresponds to a 1-hour window, and the adjacent samples are separated by 30 minutes. The test set selects the burst load scene (such as the peak period of competition) in the last 3 months to verify the generalization ability of the model;

[0043] The numerical features (such as resource utilization rate and the number of active servers) are subjected to Min-Max normalization, specifically , wherein is the minimum and maximum value of the corresponding feature in the training set;

[0044] A deep reinforcement learning (DRL) model is constructed using a deep deterministic policy gradient algorithm, which includes: input layer, hidden layer, output layer;

[0045] The input layer dimension is consistent with the output of the feature engineering, that is , wherein Min-Max normalization is performed before input to accelerate the convergence of gradient descent;

[0046] The hidden layer performs feature extraction and strategy mapping, which is a 2-layer fully connected layer with 256 neurons in each layer. Through experimental verification, the model complexity and generalization ability are balanced. Nonlinear mapping is introduced to avoid gradient disappearance. A Dropout layer (dropout rate 0.1) is added to each layer to prevent overfitting and enhance generalization ability.

[0047] The output layer performs action generation , including policy network output and value network output, where the policy network output is a continuous value vector, normalized by Softmax, with a value range of [0, 1] and a dimension of , which outputs continuous probabilities during training and allocates according to the highest probability principle during actual scheduling (e.g., if the probability is >0.5, it is allocated, otherwise it is not allocated).

[0048] The value network output is a 1-dimensional scalar (Q value) that evaluates the long-term reward value of the current state-action pair.

[0049] Where the reward function is: , where is the current number of active servers, is the resource utilization rate, , is the weight coefficient , the weight coefficient is determined by experiment, where the initial value , and then optimized by grid search. Specifically, test different combinations on the validation set, select the combination that makes the average reward highest, and automatically fine-tune every 24 hours based on real-time scheduling results.

[0050] Get the current 1-hour window state vector , the policy network outputs the allocation probability based on , generates a discrete action , implements the allocation scheme for cluster management, and records the actual active servers and resource utilization rate.

[0051] By interval listening to the experiment progress and resource occupation of each student, the resource occupation of each physical server and each virtual machine, real-time pulling of physical server state matrix, virtual machine demand vector, and global state indicators, the format and feature engineering output are completely consistent.

[0052] By detecting that the average resource utilization of the physical server is <40% in the last 3 time windows (15 minutes each, a total of 45 minutes), and no new event reservation triggers the light load scenario scheduling, the policy network outputs the virtual machine allocation probability matrix, filters the allocation scheme with a screening probability >0.6, and allocates all running virtual machines to the top 50% servers with the highest utilization in the current active servers. The remaining servers are marked as "shut down", and if the number of active servers M decreases by ≥10% after shutdown, trigger a reward calculation once;

[0053] In the last 3 time windows, the average resource utilization of the physical server is ≥60%, or the total number of virtual machine demands Trigger high-load scenario scheduling, where the upper limit of single-server load is: CPU and memory utilization ≤90% (to avoid overload), disk and network utilization ≤80% (IO and bandwidth buffer reserved), according to the priority of competition level, initial allocation matrix , clear the candidate server list of each virtual machine, perform PSO local optimization, each particle represents a distribution scheme through particle coding, only adjust the virtual machines on the top 20% servers in the initial solution, specifically , where , balance the number of servers and utilization, in each iteration, part of the virtual machines on the high-load server are migrated to the low-load server, if the utilization of a server is <20% and there is no high-priority virtual machine after migration, it is marked as "to be closed", high-priority event virtual machines are allocated preferentially, if there is still a conflict (such as insufficient resources) after allocation, trigger the event management module to adjust the event time, record the scheduling delay (target <50ms), if it is overdue, keep the initial solution of DRL and give up PSO optimization,

[0054] For the scheduling of sudden task scenarios, the trigger condition is to receive high-priority event applications (such as national emergency additional events), or the priority of existing event tasks suddenly increases, the virtual machine resource occupancy rate increases by ≥150% within 10 minutes, and causes the utilization rate of the server to be >95% to trigger an overload warning. Based on the priority rules, low-priority tasks are forcibly migrated to ensure high-priority resource requirements, wherein the migration rule engine includes: condition 1: the target server idle resource ≥ the demand of the virtual machine to be migrated by 120%, direct migration; condition 2: if condition 1 is not met, select the server with the lowest current load, and forcibly migrate the task with the lowest priority on the server to the standby resource pool (preset 5% physical servers as hot standby, long-term active state); mark the emergency event virtual machine as "high priority", suspend the DRL / PSO process, directly start the rule engine, and feed back the standby server state to the DRL model in real time to avoid subsequent scheduling of standby resources. Light load / high load detection: triggered every 15 minutes (consistent with the sampling frequency), sudden task detection, real-time triggering, when both light load and sudden task conditions are met, the sudden task is processed first (the rule engine occupies the scheduling right).

[0055] In use, receive user input competition details, the number of students participating in the competition and the number of virtual machines applied by each person, automatically associate resource priority according to the competition level (the higher the level, the greater the resource allocation weight), for example, the default allocation of high configuration physical server for national competition, determine the number of virtual machines required for single event, physical resource type (CPU / memory / disk / network) and configuration specification according to competition project type (such as programming competition, big data analysis competition), generate "virtual resource demand table based on event time", including the resource peak expectation of each event in different period (such as CPU demand surge in programming debugging stage), synchronize the total amount of physical resources in cloud computing center (real-time updated CPU / memory / disk / network available amount), calculate the total amount of virtual resources in the same period according to 15 minute granularity, if the total amount of virtual resources in a period exceeds 80% of the total amount of physical resources (set threshold), start the event time adjustment mechanism, adjust the time window of low level event first to ensure the stability of high level event resources, record the historical scheduling scheme (including resource comparison before and after time adjustment, event level weight), form reusable time coordination strategy library, create exclusive experiment cluster according to event demand (such as GPU cluster for AI competition, isolated network environment for network security competition), support "one key creation", "batch destruction", "suspend / resume" operation, bind event ID, competition level and student permission for each cluster, perform resource pre allocation test on the created cluster, simulate resource occupation when participating students start virtual machine at the same time, ensure the success rate of cluster initialization, monitor and dynamically schedule in real time during the competition, listen to the experiment progress and virtual machine resource occupation of each student, identify abnormal resource surge (such as CPU utilization of single virtual machine exceeding 90% in 10 minutes), collect the activation state and four types of resource utilization of each physical server, generate dynamic resource occupation curve, associate virtual machine data with event ID and competition level, bind physical server data with resource type and activation state, discard records with missing fields ≥30%; fill in the average value of 5 minutes before and after the missing field (such as disk IO data missing of a server, take the average value of adjacent period), aggregate the average value of four types of resource utilization (CPU / memory / disk / network) and server activation state according to 1 hour window, form a matrix with dimension N x 5 (N is the total number of physical servers), count the total number of virtual machines required by each event, the number of cluster types and the average competition level in the window, generate a vector with dimension K x 2 (K is the number of cluster types), finally input the dimension (N x 5 + K x 2) as the state input of DRL model, output the virtual machine allocation probability matrix, generate discrete action according to the principle, evaluate the long-term reward value of the current allocation scheme, allocate virtual machine resources to high level events first, select allocation probability > 0.Scheme 6 of 6, the virtual machine is concentrated to the current activation server with the highest utilization of the first 50% of the server, resource convergence, if the number of closed servers ≥ 10% of the current activation, trigger reward calculation, while marking the "closed" server as standby state, high load scenario single server CPU / memory utilization ≤ 90%, disk / network utilization ≤ 80% (IO and bandwidth buffer reserved), when exceeding, start PSO local optimization, allocate resources according to the descending order of competition level (such as national competition priority to occupy idle resources), migrate virtual machines on the top 20% of servers, if the resource is insufficient, trigger the competition time adjustment module (such as delaying the start time of the next stage of low-level competition), record the scheduling delay, receive the national emergency competition application, the priority of the existing competition is improved or the utilization rate of a server > 95%, force to migrate low priority competition virtual machine to standby resource pool, mark the emergency competition virtual machine as "high priority", suspend the DRL / PSO process, directly allocate resources through the rule engine, the standby server state is synchronized to the DRL model in real time to avoid repeated occupation of standby resources in subsequent scheduling.

[0056] The competition time, the number of participating students, the actual allocated physical resources (CPU core number / memory capacity / disk space / network bandwidth), the scheduling results (successful migration times / time adjustment times) are data archived, the training set (continuous 12 months of historical scheduling data) and the test set (the last 3 months of peak period data) are divided according to the ratio of 8:2, the sliding window (adjacent sample interval 30 minutes) with time sequence unchanged is adopted, the numerical value type features (such as resource utilization, the number of activated servers) are Min-Max normalized to improve the training efficiency of DRL model, through the post-competition review, the scheduling delay cases in high load scenarios are analyzed (such as a certain competition causes resource allocation delay due to PSO optimization timeout), the probability threshold of DRL output layer is adjusted, the DRL model retraining is automatically triggered, and the latest high priority scheduling strategy is injected.

[0057] Compared with the existing design, by real-time monitoring of the student experiment progress and resource occupation, the system can dynamically adjust the resource allocation, ensure that the total amount of virtual resources required by different events in the same period does not exceed the set threshold, effectively solve the problem of rigid resource allocation in traditional competition management system, realize efficient use of resources through real-time data feedback, reduce resource waste, improve the overall operation efficiency of the system, through the event time adjustment mechanism, the high load event is dispersed to different time periods, avoiding resource overload, breaking through the limitation of traditional resource allocation only focusing on spatial dimension, significantly improving the carrying capacity of the cloud computing center, according to the average resource utilization of the physical server, and according to different load scenarios, resource convergence or PSO local optimization strategy is adopted, through the scene scheduling strategy, the self-adaptive ability to complex load environment is realized, the stability and fairness of resource allocation are ensured, according to the competition project, event time, number of participants and other factors, the resource demand table is generated, which provides quantitative basis for resource allocation, through the structured data model, the visualization of resource demand is realized, which provides scientific basis for resource allocation decision, avoiding the blindness of traditional experience allocation, through the matrix model, the allocation position of each virtual machine is determined, realizing the accurate allocation of resources, the matrix allocation method improves the transparency and traceability of resource allocation, providing data basis for subsequent scheduling strategy optimization, according to the competition level, the resource priority is automatically associated, ensuring the stable allocation of resources for high-level events, through the priority rule engine, the differentiation of resource allocation is realized, ensuring the smooth operation of key events, improving the fairness and quality of the competition, using deep deterministic policy gradient algorithm (DRL) to build the model, generating virtual machine allocation probability matrix, through end-to-end training, realizing the autonomous learning and optimization of scheduling strategy, significantly improving the intelligence and efficiency of resource allocation, through the reward function to evaluate the long-term reward value of the current state-action pair, guiding the model to learn the optimal scheduling strategy, avoiding the local optimal problem of traditional heuristic algorithm, in the high load scenario, particle swarm optimization algorithm (PSO) is used for local optimization, further adjusting the resource allocation scheme, the combination of PSO algorithm and DRL model realizes the cooperation of global optimization and local optimization, significantly improving the robustness of the system in high load scenario.

[0058] Embodiment 2

[0059] In embodiment 1, according to different load scenarios, resource convergence or PSO local optimization strategy is adopted, through the scene scheduling strategy, the self-adaptive ability to complex load environment is realized, ensuring the stability and fairness of resource allocation, this embodiment is further improved on the basis of the above embodiment.

[0060] The embodiment also includes three new hardware features on the basis of the original four types of resource utilization and activation state: GPU computing power index, memory capacity normalization value, and disk performance level. The GPU computing power index is mapped to a quantitative value according to the model, the memory capacity normalization value is the current available memory / total memory, and the disk performance level is NVMe=3, SSD=2, and HDD=1; a server state matrix is constructed, with dimensions of server number x 8 (4 utilization rates + 1 activation state + 3 hardware features); data is aggregated in a 1-hour time window, the task feature vector is spliced with the server state matrix to form a model input sample, and the final input dimension is: The training set and the test set are divided in a ratio of 8:2, the training set contains continuous 12-month data, the test set selects near-3-month peak period data, and the time sequence is kept unchanged;

[0061] Task resource demand labels are generated in batches through a static code scanning tool, and server hardware characteristic data is labeled synchronously. Static code features are extracted, resource sensitive operations are identified, and memory-intensive operations (such as matrix operations and large data set loading) are marked as “high memory”; high-frequency network I / O operations (such as distributed communication and API calls) are marked as “high bandwidth”;

[0062] Based on the historical running data of the task, a K-means clustering algorithm is used for resource demand classification:

[0063] If the memory utilization rate exceeds 80% for 30 minutes and the peak value reaches 70% of the physical memory, it is marked as “memory bottleneck type”; if the network bandwidth occupancy exceeds 500 Mbps and the fluctuation coefficient (standard deviation / mean) > 0.7, it is marked as “network sensitive type”, and a three-level label structure is established, including basic labels according to single resource dimensions, composite labels of multiple resource combinations, and priority labels associated with competition levels;

[0064] When submitting a competition task, the system automatically parses the code package and triggers static code scanning; at the same time, historical data of the same type of task is queried to generate initial labels (basic labels + priority labels), and during the task running process, if the resource occupancy mode changes significantly, real-time dynamic composite labels are updated to ensure that the latest demand is obtained during scheduling;

[0065] The basic labels are converted into numerical vectors, and the priority labels are mapped to integers from 1 to 3 according to the competition level;

[0066] The server is monitored to obtain server configuration in real time, generate a basic hardware profile, and automatically classify it based on hardware configuration. Specifically, servers with GPU and ≥24GB of video memory are marked as "AI computing power type", servers with ≥1TB of memory and NVMe disks are marked as "big data processing type", and servers with network bandwidth ≥100Gbps are marked as "low latency communication type". Resource utilization and activation status are updated every 5 minutes and merged with the basic hardware profile to form a real-time server profile that includes hardware characteristics, resource utilization, activation status, and adaptation tags.

[0067] A multi-dimensional matching degree calculation model is established, with the specific basic resource matching degree as follows:

[0068] ;in For resource types (such as CPU, memory, bandwidth);

[0069] Priority matching degree is:

[0070] Reinforcement learning model reconstruction, where the input layer receives the task vector and the server state matrix, with dimension 1. ,in Number of servers The number of cluster types is specified; the hidden layer uses two fully connected layers (256 neurons each), and a Dropout layer (0.1% inactivation rate) is introduced to prevent overfitting; a matching degree sub-network is added after the hidden layers, and the output task-server matching degree matrix is ​​output and fused with the output of the DRL policy network; a sliding window method is used to generate adjacent samples (30-minute interval) to increase sample diversity, and oversampling is performed on high-priority competition data to balance the sample distribution and avoid the model being biased towards low-priority tasks;

[0071] The task vectors and server profiles are embedded into the existing competition management system. Real-time labels and profile data are cached through middleware (such as Redis). By designing a scheduling engine interface, the system receives the allocation matrix output by the reinforcement learning model and calls the underlying resource management. After each large-scale competition, the scheduling logs are analyzed to identify cases of high latency or resource waste, and abnormal data is manually labeled and injected into the training set.

[0072] In use, 3 hardware features are added to the server and quantified: GPU computing power index is mapped to a quantized value according to the model, memory capacity normalized value = current available memory / total memory, disk performance level, a server state matrix is constructed, the dimension is "server number x 8", historical data is aggregated according to a 1-hour time window, and the training set and test set are divided according to an 8:2 ratio, the training set contains continuous 12 months of historical data, the test set selects the data of the recent 3 months of the competition peak period, and the time sequence is kept unchanged to retain the time sequence characteristics, when the user submits the competition task code package, the system automatically triggers the static code scanning tool, parses the code package and extracts the static code features, identifies resource sensitive operations, generates basic labels, synchronously queries historical data of the same type of competition task, obtains the historical resource demand mode, generates the initial priority label, based on the historical running data of the task, the K-means clustering algorithm is used for resource demand classification, the basic label + priority label forms the initial vector of the task, the server configuration is obtained in real time, and the preset rule is automatically analyzed and classified according to the preset rule, and the hardware basic portrait is generated. Every 5 minutes, the server resource utilization rate, the active state and the hardware basic portrait are merged to form a real-time server portrait, based on the task basic label and the real-time resource utilization rate of the server, the matching degree of CPU, memory, bandwidth and other single resources is calculated, according to the task priority label and the current load priority adaptation degree of the server, it is directly mapped to a numerical value, the task-server matching degree matrix is output through the matching degree sub-network, the reinforcement learning (DRL) strategy network is fused to output the final allocation strategy, the task feature vector and the server state matrix are spliced, the input dimension is "server number x 8", and the sliding window method is used to generate adjacent samples to increase the diversity of samples; oversample high-priority competition task data to balance sample distribution and avoid model bias towards low-priority tasks, the scheduling engine receives the allocation matrix output by the reinforcement learning model, calls the underlying resource management interface (such as Kubernetes API), and assigns the task to the server cluster with the highest matching degree. Through the middleware (such as Redis), the real-time task label and server portrait data are cached to ensure low-latency data interaction, which is suitable for real-time scheduling requirements of competitions.

[0073] Compared with the prior art, by converting GPU computing power index, memory capacity normalization value, disk performance level into quantifiable parameters, a multi-dimensional server state matrix is constructed, the limitations of the traditional scheduling system relying only on CPU / memory utilization are overcome, the task adaptation problem caused by hardware heterogeneity is solved, the task and the hardware characteristics of the server are deeply matched (such as GPU-intensive tasks being preferentially assigned to high-computing nodes), the potential resource contention is predicted by the memory capacity normalization value, the risk of task interruption caused by insufficient memory is reduced, the basic tags (high memory / network sensitive), priority tags (competition level association) and dynamic composite tags are generated by combining static code scanning and runtime data, the problem that the traditional tag system cannot reflect the real-time resource demand fluctuation of the task is solved, the resource allocation of key event tasks is ensured through the competition level association tag, the time-sequenced server portrait is generated based on the hardware basic portrait and real-time monitoring data (resource utilization, load fluctuation), the problem that the traditional static portrait cannot reflect the instantaneous state of the server is solved, a two-dimensional calculation model of basic resource matching degree (CPU / memory / disk) and priority matching degree (competition level) is constructed, the problem of resource waste or priority inversion caused by traditional single-dimensional matching is solved, the long-term starvation of low-priority tasks is avoided through priority weighted matching, the matching degree sub-network is embedded in the DRL framework, the allocation strategy of the fusion matching degree score and Q value is output, the problems of slow convergence speed and one-sided decision-making of the traditional reinforcement learning model in complex scheduling scenarios are solved, in the case of sudden traffic, the scheduling strategy adjustment time is shortened, the system resilience is enhanced, time-adjacent samples are generated through a sliding window, and high-priority task data is oversampled, solving the problems of insufficient time and space correlation of training data and class imbalance.

[0074] Embodiment 3

[0075] In embodiment 2, further adjustments are made to task allocation, competition types are divided, and resource demand tags are marked, realizing heterogeneous tasks. This embodiment further improves on the basis of the above-mentioned embodiments;

[0076] In this embodiment, the competition project is divided into multiple stages, specifically including registration preparation, training debugging, and formal competition; each stage defines a resource demand vector, and combines historical competition data to verify the typicality of resource demand in each stage through a time series clustering algorithm to form a standardized template. When a user submits a competition task, the system automatically matches the stage template according to the competition type;

[0077] On the basis of the three-level tags, a stage tag is added, and the task vector dimension is expanded to four types of tags: basic tags, composite tags, priority tags, and stage tags. On the basis of the server state matrix, a stage adaptation degree field is added, wherein the stage adaptation degree is specifically:

[0078] , wherein Resource weight, hardware resource and stage template demand are calculated by normalized value;

[0079] When collecting server status every 5 minutes, the adaptation degree of each stage is calculated synchronously, the server stage image is updated in real time, the dynamic stage image matrix is formed, and the new stage matching degree is calculated, which is:

[0080] , wherein is an independent encoding vector, and only the dimension corresponding to the current stage is activated;

[0081] , wherein The training is determined by a multi-objective optimization algorithm, and the target is to maximize resource utilization and stage task completion rate;

[0082] The input and output layers of the reinforcement learning model are modified, the original input dimension is replaced by server number x 13 + cluster type number x 2 + task stage vector 5, wherein the task stage vector is one-hot encoding (5 dimensions); A stage gating unit is added after the hidden layer to dynamically adjust the resource allocation strategy according to the task stage label; during the training and debugging stage, the server with GPU is preferentially allocated, and the server utilization rate is allowed to fluctuate; during the formal competition stage, the server with high adaptation degree is forced to be allocated, and the server with the active state of "stable running" is activated;

[0083] Through a timing task, all competition tasks are scanned every 5 minutes, whether the current time reaches the preset switching time in the task stage time axis or the task state change interface is listened to, the task output is detected to meet the stage promotion characteristics through log analysis, the stage switching event is triggered, the resource demand template of the target stage and the resource capacity lower limit are extracted, based on the dynamic stage image of the current active server, the server with the adaptation degree to the target stage is selected, the total remaining resources of the candidate server are calculated and compared with the required resources in the resource demand template of the target stage, if the gap is ≥ 20%, the standby resource wake-up process is triggered, the priority of the migrated task is sorted according to the "stage priority + competition level + current resource occupation rate" three rules, the comprehensive matching degree of the candidate server is calculated, the "batch migration + load balancing" strategy is adopted, and each time migration does not exceed 10% of the current active server number, avoiding instantaneous traffic impact, through the Pod scheduling strategy of Kubernetes, the task container is migrated across nodes, the service is not interrupted during the migration process, the server resource fluctuation during the migration process is monitored in real time through Prometheus + Grafana, if the CPU utilization of a server after migration is > 90% or the network bandwidth occupation is > 80%, the subsequent migration task of the server is immediately suspended, and local rebalancing is triggered;

[0084] The deviation of the actual resource consumption of each task at each stage from the template expectation is recorded to form a stage resource deviation matrix, which has dimensions of: number of tasks × number of stages × resource dimension, where the number of tasks is all task IDs in the current competition period, the number of stages is the stage label experienced by the task, and the resource dimension includes CPU core number, memory capacity, GPU computing power, disk performance, and network bandwidth, five types of core resources, each resource dimension contains two sub-items of absolute deviation and relative deviation. By analyzing the deviation distribution, unreasonable items of the stage resource demand template are identified, and the stage-resource combination with high resource allocation conflict is located to guide the DRL model to adjust the allocation weight. The stage template library is updated every quarter, the stage resource demand vector is recalculated based on the latest competition data, and the DRL model parameters are fine-tuned through online reinforcement learning.

[0085] In use, according to the typical process of the competition, three core stages of “registration preparation, training debugging, and formal competition” are preset, and a resource demand vector including CPU core number, memory capacity, GPU computing power, disk performance, and network bandwidth, five types of core resources, is defined for each stage. The typicality of resource demand in each stage is verified based on historical competition data through a time series clustering algorithm to form a standardized stage template, which is stored in a stage template library. The task vector dimension is defined to include four types of labels, including basic labels, composite labels, priority labels, and stage labels. A stage adaptation degree field is added to the server state matrix. When a user submits a competition task, the system automatically extracts information such as competition type and expected period, matches the corresponding stage template based on the preset stage timeline, generates a task stage vector, and allocates basic resources for the registration preparation stage task based on the real-time server state matrix through a multi-objective optimization algorithm. The stage gating unit of the reinforcement learning model automatically activates the “training debugging stage strategy”: preferentially allocates servers with GPUs, collects server states every 5 minutes, updates the dynamic stage portrait matrix, and switches the stage gating unit to the “formal competition strategy”: forces the allocation of servers with a stage adaptation degree ≥0.9 and a state of “stable running”, prohibits dynamic resource migration, extracts the resource demand template and resource capacity lower limit of the target stage, and calculates the total remaining resources of the currently activated servers. If the gap is ≥20%, the standby resource wake-up process is triggered, sorted according to the “stage priority + competition level + current resource occupancy rate” three rules, and the number of tasks migrated each time does not exceed 10% of the number of currently activated servers. Cross-node migration of containers is realized through the Pod scheduling strategy of Kubernetes. The service is ensured not to be interrupted during the migration process through session preservation technology. The deviation of the actual resource consumption of each task at each stage from the template expectation is recorded to form a stage resource deviation matrix, which is analyzed every month to identify the “stage-resource combination” with high resource allocation conflict, guide the DRL model to adjust the allocation weight, and recalculates the stage resource demand vector based on the latest competition data every quarter. The model parameters are fine-tuned through online reinforcement learning, and the stage template library is updated.

[0086] Compared with existing designs, by dividing the competition project into multiple stages and defining resource requirement vectors for each stage, a standardized template is formed, making resource allocation more accurate and enabling dynamic adjustment according to the needs of different stages, improving the flexibility and efficiency of resource management. A dynamic stage portrait matrix is constructed, and the resource allocation strategy is dynamically adjusted through the stage gating unit. The dynamic stage portrait can reflect the state of the server in real time, providing accurate basis for resource allocation. The stage gating unit dynamically adjusts the resource allocation strategy according to the task stage label, ensuring that the task obtains sufficient resource support in different stages, thereby improving the resource utilization rate and task completion rate. The batch migration + load balancing strategy is adopted to ensure the system stability and resource utilization rate during task migration, avoiding instantaneous traffic impact, ensuring stable operation of the system, improving resource utilization, reducing resource waste, and continuously optimizing resource allocation strategy through online reinforcement learning model parameter fine-tuning. Online reinforcement learning enables the system to adapt to changes in competition data and maintain high efficiency in resource allocation. Even in the face of changing competition demands, the system can continuously optimize through self-learning to maintain excellent performance. Through phased resource management and dynamic stage portrait, precise allocation of resources is achieved, avoiding waste of resources, significantly improving resource utilization, dynamically adjusting resource allocation strategy according to task stage label to ensure that tasks obtain sufficient resource support in different stages, thereby improving task completion rate, adopting batch migration + load balancing strategy to avoid instantaneous traffic impact, ensuring stable operation of the system, reducing the risk of system crash or performance degradation.

[0087] Embodiment 4

[0088] In embodiment 3, according to the task stage driven resource demand prediction, the dynamic resource quota adjustment strategy is set, and the stage-aware reinforcement learning scheduling is realized. This embodiment further improves the above-mentioned embodiments;

[0089] In this embodiment, each competition task is divided into demand analysis, environment building, code development, debugging optimization, and formal submission. Each stage defines a dynamic resource requirement vector containing CPU core number, memory capacity, GPU computing power, disk performance, and network bandwidth.

[0090] For each team member task, a team ID and member role label are bound, where the role affects the resource priority. A cross-task stage dependency relationship is established, and five resource queues are established according to the stage type. Each queue is configured with a dynamic capacity upper limit. The priority scheduling is sorted according to the competition level + team progress deviation. Based on the stage adaptation degree formula, a team task aggregation factor is introduced, specifically:

[0091] where is the single task priority, For team coordination weight, specifically, through grid search dynamic tuning, avoid the same team task dispersion in different servers leading to communication overhead, server dynamic portrait expansion adds "current bearing stage distribution" field, used for predicting resource hotspots;

[0092] The input and output layers of the multi-objective reinforcement learning model are expanded. The state input is the fusion of multi-task stage vector + server dynamic stage portrait, and the final input dimension is generated. The action output is a multi-task allocation matrix, where the element represents the task allocation of team to server N, supports the same server to bear multiple stage tasks, and reconstructs the reward function. Specifically:

[0093]

[0094] Among them: the team stage coordination degree is the number of tasks in the same team in the same stage / total number of tasks, the higher it is, the lower the communication overhead, is the resource efficiency priority weight (initial value is 0.6), is the coordination optimization weight (initial value is 0.3), is the delay penalty weight (initial value is 0.1), and the weight is dynamically tuned through grid search;

[0095] A sliding window is used to generate a multi-task stage sequence. The debugging optimization / formal submission stage data of high-level competition is oversampled. The debugging optimization / formal submission stage data (when the proportion is insufficient) is duplicated twice to balance the sample distribution. The time window for generating samples is 1 hour, including 30 minutes of context before and after, to enhance the adaptability of the model to stage switching;

[0096] Randomly initialize the number of policy networks and value networks, load the stage template library in embodiment 3 as prior knowledge; output the allocation policy according to the current state (task stage + server state), obtain the real-time reward (resource utilization, coordination degree, migration delay) after execution, record the experience tuple (state s, action a, reward r, next state s'), and store it in the experience replay buffer. Sample batch data from the buffer, calculate the TD error to update the value network, use the deep deterministic policy gradient (DDPG) algorithm to update the policy network, and stabilize the training process through soft update (Polyak average);

[0097] After each cross-stage migration, record the stage allocation bias matrix (task number x stage number x resource dimension), identify resource allocation conflicts, inject high-frequency conflict scenario data into the training set every hour, trigger model fine-tuning, dynamically update the stage template library based on example 3, recalculate the stage adaptation weight, adjust the stage gating unit parameters, retrain the model using the full year data to adapt to the long-term changes in competition types and resource requirements.

[0098] In use, by strictly dividing each competition task into five core stages of requirement analysis, environment building, code development, debugging optimization and formal submission, defining a dynamic resource requirement vector including CPU core number, memory capacity, GPU computing power, disk performance and network bandwidth for each stage, establishing stage dependency relationship between different tasks (such as the “code development” stage of a team needs to wait for the “environment building” stage to complete before starting), ensuring that the scheduling conforms to the logical order of the competition task, creating five resource queues according to the type of stage (each stage corresponds to a queue), setting a dynamic capacity upper limit for each queue (such as the “debugging optimization” queue has a lower capacity upper limit than the “requirement analysis” queue due to the shortage of GPU resources), and prioritizing tasks in the queue according to the competition level + team progress deviation (the difference between actual progress and planned progress) (high-level competition tasks, progress lag team tasks are prioritized), adding a “current bearing stage distribution” field in the server dynamic portrait to predict resource hotspots, combining the multi-task stage vector (resource requirements of the current stage of each task) with the server dynamic stage portrait (server bearing stage distribution) to form the final input dimension (capture the matching degree of task requirements and server load), generating a multi-task allocation matrix (elements represent the allocation of a team task to a server), supporting the same server to bear multiple stage tasks, introducing team stage coordination degree, building a reward function, adjusting the weight for different competition types through grid search dynamic optimization, doubling the sampling of debugging optimization and formal submission data (debugging optimization and formal submission, resource requirements are complex and data volume is small) for high-level competition, balancing sample distribution, solving the problem of insufficient data in high difficulty stages, generating training samples with a 1-hour time window, including 30 minutes of context before and after, capturing the change of resource requirements before and after stage switching, randomly initializing the strategy network and value network, loading the stage template library in embodiment 3 as prior knowledge to accelerate training convergence, outputting the task allocation strategy through the strategy network according to the current state, recording the experience tuple and storing it in the experience replay buffer after executing the allocation strategy, sampling batch data from the buffer, updating the strategy network using the deep deterministic policy gradient (DDPG) algorithm, stabilizing the training process through soft update (Polyak average) to avoid model oscillation caused by dynamic changes of competition tasks, triggering grid search once an hour, dynamically adjusting the reward function weight according to the current competition type, recording the stage allocation deviation matrix after each cross-stage migration, identifying resource allocation conflicts, injecting high-frequency conflict scenario data into the training set to trigger model fine-tuning, thereby realizing precise scheduling of competition tasks in all stages, multiple teams and dynamic resource requirements, and significantly improving resource utilization efficiency and team collaboration efficiency.

[0099] Compared with the prior art, the resource quota can be dynamically adjusted through phased prediction, independent resource queues are established according to the types of phases, and dynamic sorting is performed in combination with competition levels and team progress deviations, so that the problem of 'high-priority tasks being blocked by inefficient tasks' in traditional scheduling is avoided, role tags are bound for team members, and cross-phase dependency relationships are established, through role priority and phase order constraints, manual coordination costs are reduced, rework caused by task conflicts is avoided, the reward function of the reinforcement learning model is reconstructed, team phase collaboration degree, resource efficiency priority weight, collaboration optimization weight and delay penalty weight are introduced, through multi-objective weight dynamic adjustment, dual optimization of resource allocation and team collaboration is realized, after random initialization of the strategy network and the value network, the historical phase template library is loaded as prior knowledge, soft updating is realized, and model shock caused by dynamic changes of competition tasks is avoided, a phase allocation deviation matrix is recorded, high-frequency conflict scene data is injected every hour to trigger model fine-tuning, and the data is retrained using the annual data to adapt to long-term changes, a sliding window is used to generate a multi-task phase sequence, and the key phase data of high-level competitions is oversampled to avoid model bias caused by insufficient data of high-difficulty phases.

[0100] Embodiment 5

[0101] The application provides a competition management system based on cloud computing, which comprises a resource demand management module, an experimental cluster management module, a monitoring and data acquisition module, a data processing and feature engineering module, a reinforcement learning model module, a load scene scheduling module, a task label and matching module, a phased resource management module and a data storage and management module.

[0102] The experimental cluster management module is used for quickly responding to the environmental requirements of different competitions, realizing automatic management of the life cycle of the cluster, and guaranteeing the isolation and reliability of the experimental environment.

[0103] The monitoring and data acquisition module is used to provide real-time resource status data, provide input for dynamic scheduling and model training, and support anomaly detection. It real-time monitors student experiment progress, resource occupation of virtual machines and physical servers, draws resource occupation curves and outputs them in graphical form, collects competition metadata and virtual machine association data;

[0104] The data processing and feature engineering module is used to clean and structure raw data, generate feature vectors suitable for model input, and improve the training efficiency of subsequent reinforcement learning models. It aggregates resource data in 1-hour windows, constructs state vectors, and calculates competition features. Numerical features are normalized using Min-Max normalization.

[0105] The reinforcement learning model module is used to implement intelligent resource allocation. It trains and optimizes scheduling strategies through historical data, balances resource utilization and server activation number, and avoids the local optimization problem of traditional heuristic algorithms. It uses the Deep Deterministic Policy Gradient (DDPG) algorithm to build the model, which includes input layer, hidden layer, and output layer. The reward function combines the number of activated servers, resource utilization, and priority weight to support end-to-end autonomous learning of scheduling strategies.

[0106] The load scenario scheduling module is used to dynamically adjust strategies for different load scenarios, improve resource utilization efficiency, and ensure system stability and resource needs of high-priority tasks. It schedules light and heavy loads, and when high-priority competitions or overload warnings occur, it starts the rule engine to forcibly migrate low-priority tasks to the backup resource pool.

[0107] The task label and matching module is used to solve the problem of matching hardware heterogeneity and real-time task requirements, so that GPU-intensive tasks are preferentially allocated to compatible servers, reducing resource contention risks. It uses static code scanning to identify resource-sensitive operations, combines K-means clustering to generate task labels, adds server hardware features, constructs a multi-dimensional matching degree model, outputs a task-server matching degree matrix, and integrates it into the DRL strategy.

[0108] The phased resource management module is used to accurately match resource needs at different stages, improve resource allocation flexibility, reduce traffic impact during stage switching, and optimize resource scheduling in team collaboration scenarios. It divides the competition into multiple stages, defines stage resource demand templates, dynamically generates server stage adaptation degrees, adjusts strategies through stage gating units, supports batch migration and load balancing during stage switching, records stage resource deviation matrices to optimize the template library, and introduces team collaboration factors and multi-task queue scheduling.

[0109] The data storage and management module is used for providing a high-quality data set for model training, guaranteeing time sequence feature integrity, and improving model generalization capability, wherein historical scheduling data is archived, a training set and a test set are divided according to 8:2, a sliding window with time sequence invariance is adopted, high-priority data is oversampled, and sample distribution is balanced.

[0110] The above describes the present application and its embodiments, which are not limited, and the drawings only show one of the embodiments of the present application, and the actual structure is not limited thereto. In general, if a person skilled in the art is inspired, without departing from the purpose of the present application, without creating a similar structure and embodiment of the technical solution, which is not creative, should belong to the protection scope of the present application.

[0111] It should be noted that in the above embodiments, the description of each embodiment has its own emphasis, and the parts not described in detail in a certain embodiment can be referred to the related description of other embodiments.

[0112] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can adopt a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer usable storage media containing computer usable program code (including but not limited to disk memory, CD-ROM, optical memory, etc.).

[0113] The present application is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and the combination of flows and / or blocks in the flowcharts and / or block diagrams can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded computer, or other programmable data processing devices to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing devices produce a means for implementing the functions specified in the flowcharts and / or block diagrams. Figure 1 The functions specified in one or more flows and / or blocks. Figure 1 The means for implementing the functions specified in one or more flows and / or blocks.

[0114] These computer program instructions can also be stored in a computer readable storage medium that can guide the computer or other programmable data processing devices to work in a specific way, so that the instructions stored in the computer readable storage medium produce a product including instruction means, which implements the functions specified in the flowcharts and / or block diagrams. Figure 1 The functions specified in one or more flows and / or blocks. Figure 1 The means for implementing the functions specified in one or more flows and / or blocks.

[0115] These computer program instructions can also be loaded into computer or other programmable data processing devices, so that a series of operations steps are performed on the computer or other programmable data processing devices to generate computer-implemented processes, thus the instructions executed on the computer or other programmable data processing devices provide the function of realizing the processes specified in the flowchart Figure 1 one flowchart or multiple flowcharts and / or blocks Figure 1 one block or multiple blocks.

[0116] Although the preferred embodiments of the application have been described, those skilled in the art will be able to make additional modifications and variations to these embodiments without departing from the spirit and scope of the application. Accordingly, it is intended that the appended claims be construed to include all such modifications and variations as fall within the scope of the application.

[0117] Obviously, various modifications and changes are possible in the present application without departing from the spirit and scope of the application. It is to be understood that the application includes any such modifications and changes only insofar as they come within the scope of the appended claims and their equivalents.

Claims

1. A cloud computing-based contest management method, characterized by, The method comprises the following steps: determining the required number of virtual machines, physical resources and competition levels according to competition information, automatically associating resource priorities according to competition levels, dynamically adjusting resource allocation, ensuring stable allocation of resources for high-level competitions, constructing a deep deterministic policy gradient algorithm model, generating a virtual machine allocation probability matrix, and realizing autonomous learning and optimization of the scheduling strategy; and evaluating the long-term reward value of the current state-action pair through a reward function to guide the model to learn the optimal scheduling strategy. A multi-dimensional server state matrix is constructed to deeply match tasks with server hardware characteristics, potential resource contention is predicted through memory capacity normalization values to reduce the risk of task interruption caused by insufficient memory, labels are generated through static code scanning and runtime data, and the labels are associated with competition levels to ensure successful resource allocation for key competition tasks. The competition project is divided into multiple stages, and a resource demand vector is defined for each stage to form a standardized template, and the resource allocation can be dynamically adjusted according to the demand of different stages. Each competition task is divided, and a team ID and a member role label are bound to each team member task to establish a cross-task stage dependency relationship, and the role priority and stage order are constrained. The reward function of the reinforcement learning model is reconstructed, and team stage coordination degree, resource efficiency priority weight, coordination optimization weight and delay penalty weight are introduced, and through multi-objective weight dynamic adjustment, the dual optimization of resource allocation and team cooperation is realized, the historical stage template library is loaded as prior knowledge to avoid model shock caused by dynamic changes of competition tasks, and a stage deployment deviation matrix is recorded.

2. The cloud computing-based competition management method of claim 1, wherein: By real-time monitoring of student experiment progress and resource occupation, dynamic adjustment of resource allocation is realized, and the total amount of virtual resources required by different competitions in the same period is controlled within the set threshold range, and through competition time management, the total amount of virtual resources required in the same period is controlled within the bearable range of the cloud computing center, and the competition is staggered. 3.The cloud computing-based competition management method of claim 1, wherein: The virtual machine allocation probability matrix realizes autonomous learning and optimization of the scheduling strategy through end-to-end training, and in the high-load scenario, the particle swarm optimization algorithm is used for local optimization to realize the cooperation of global optimization and local optimization.

4. The cloud computing-based competition management method of claim 1, wherein: The GPU computing index, memory capacity normalization value and disk performance level are converted into quantifiable parameters to construct a multi-dimensional server state matrix, and the static code scanning and runtime data are combined to generate three-level labels of basic labels, priority labels and dynamic composite labels. The server is monitored to generate a hardware basic portrait, and the hardware basic portrait and real-time monitoring data generate a time-sequenced server portrait to reflect the instantaneous state of the server.

5. The cloud computing-based competition management method of claim 1, wherein: A two-dimensional calculation model of basic resource matching degree and priority matching degree is constructed to avoid long-term starvation of low-priority tasks through priority weighted matching. Time-adjacent samples are generated through a sliding window, and high-priority task data is oversampled.

6. The cloud computing-based competition management method of claim 1, wherein: A dynamic stage portrait matrix is constructed, and a stage gating unit is used to dynamically adjust the resource allocation strategy to reflect the state of the server in real time and provide accurate basis for resource allocation, and the stage gating unit dynamically adjusts the resource allocation strategy according to the task stage label.

7. The cloud computing-based competition management method of claim 1, wherein: By triggering the phase switching event through a timed scan of all competition tasks, the resource requirement template of the target phase is extracted, the resource capacity lower limit is extracted, and the server suitable for the target phase is selected based on the dynamic phase image of the current active server. The sum of the remaining resources of the candidate server is calculated and compared with the required resources in the resource requirement template of the target phase. The deviation between the actual resource consumption of each task in each phase and the expected deviation of the template is recorded to form a phase resource deviation matrix, which guides the model to adjust the allocation weight and update the phase template library.

8. A cloud computing-based competition management system applied to a cloud computing-based competition management method according to any one of claims 1 to 7, characterized in that, Comprise: A monitoring and data acquisition module for providing real-time resource state data, providing input for dynamic scheduling and model training, and supporting anomaly detection; A data processing and feature engineering module for cleaning and structuring raw data to generate feature vectors suitable for model input; A reinforcement learning model module for implementing intelligent resource allocation, optimizing scheduling strategies through historical data training, and balancing resource utilization and server activation number; A load scenario scheduling module for dynamically adjusting strategies for different load scenarios to improve resource utilization efficiency and ensure system stability and high-priority task resource requirements; A task label and matching module for solving the problem of matching hardware heterogeneity and real-time task requirements, prioritizing GPU-intensive tasks to suitable servers to reduce resource contention risk; wherein each competition task is divided, each team member task is bound with a team ID and member role label, a cross-task phase dependency relationship is established, and the role priority and phase order constraints are constrained; the reward function of the reinforcement learning model is reconstructed, and the team phase coordination degree, resource efficiency priority weight, coordination optimization weight, and delay penalty weight are introduced; through multi-objective weight dynamic adjustment, dual optimization of resource allocation and team cooperation is realized, the historical phase template library is loaded as prior knowledge to avoid model shock caused by dynamic changes in competition tasks, and the phase deployment deviation matrix is recorded; A phased resource management module for accurately matching resource requirements of different phases, improving resource allocation flexibility, and reducing traffic impact during phase switching; A data storage and management module for providing high-quality data sets for model training, ensuring the integrity of time series features, and improving model generalization ability.

Images