Smart campus-oriented multi-hyper fusion platform collaborative scheduling system and method

Abstract

The invention provides a smart campus-oriented multi-hyper fusion platform collaborative scheduling system and method, and is applied to the technical field of data processing. Resource sensing and dynamic modeling processing is performed on multi-hyper fusion platform resource pool data to generate target resource model data, and the target resource model data is composed of a resource real-time monitoring index, load prediction model output, a resource isomerism adaptation result and a cross-platform protocol conversion adaptation parameter; the target resource model data is processed, platform collaborative scheduling strategy parameters are generated based on reinforcement learning, and a campus business scene reward and punishment mechanism is introduced in the reinforcement learning process; processing the platform collaborative scheduling strategy parameters to generate a dynamic resource allocation scheme; processing the dynamic resource allocation scheme and the campus service demand data, and generating a service and resource matching agent model based on an intelligent optimization algorithm; and processing the target campus information based on the service and resource matching agent model to generate campus resource scheduling information.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, in particular to a multi-super-converged platform collaborative scheduling system and method for a smart campus. BACKGROUND

[0002] In today's era of rapid development of information technology, the construction of a smart campus has become an important trend in the field of education. With the continuous advancement of informationization construction in colleges and universities, various businesses in the campus, such as teaching, scientific research, and management, have generated a large amount of data, which has put forward very high requirements for computing, storage, and network resources. However, there are many problems in the management and scheduling of resources in the campus that need to be solved.

[0003] On the one hand, existing campus resources are often scattered in different systems and devices, forming a "information island". Computing resources, storage resources, and network resources lack effective integration and collaboration, resulting in low resource utilization. Some servers are idle in certain time periods, while other businesses cannot efficiently carry out due to insufficient computing power. Storage systems also have unreasonable capacity allocation, with some storage space being tight while others being idle for a long time. Different businesses have different demands for network bandwidth, and traditional network architecture cannot dynamically adjust bandwidth according to real-time business needs, resulting in waste of network resources or business congestion.

[0004] On the other hand, with the diversification and complexity of campus businesses, such as online teaching, large-scale scientific research data processing, real-time response of smart campus management systems, the scheduling and allocation of resources need to have higher intelligence and flexibility. However, existing scheduling strategies are often based on static rules or simple threshold judgments, which cannot accurately predict the dynamic changes of business load and are difficult to achieve precise and efficient allocation of resources. For example, during the concentrated development of online courses, traditional scheduling methods may not be able to provide enough computing and network resources for teaching platforms in a timely manner, affecting the quality of teaching. In addition, there are a large number of heterogeneous resources in the campus, such as servers, storage devices, and network devices of different brands and models, which are different in interface, protocol, and other aspects, which brings great difficulty to the unified management and collaborative scheduling of resources.

[0005] It should be noted that the information disclosed in the above background section is only used to strengthen the understanding of the background of the present disclosure, and therefore can include information that does not constitute prior art known to those of ordinary skill in the art. SUMMARY

[0006] The purpose of the present application is to provide a multi-super convergence platform collaborative scheduling system and method for smart campus, at least to some extent, overcome the problems existing in the prior art, by constructing a multi-super convergence platform containing three types of resource pools, integrating resource and business information. Through the edge node collection index, the load is predicted by fusing LSTM and business cycle, and the target model is generated by adapting heterogeneous resources and converting protocols. Based on reinforcement learning and reward and punishment mechanism, generate strategy, generate optimized allocation scheme by algorithm, combine intelligent algorithm to build model to realize elastic expansion, generate scheduling information, and improve utilization and adaptability.

[0007] According to one aspect of the present application, a multi-super convergence platform collaborative scheduling method for smart campus is provided, comprising: constructing a multi-super convergence platform resource pool and target campus information for smart campus; performing resource sensing and dynamic modeling processing on the multi-super convergence platform resource pool data to generate target resource model data; processing the target resource model data, generating platform collaborative scheduling strategy parameters based on reinforcement learning; processing the platform collaborative scheduling strategy parameters to generate a dynamic resource allocation scheme; processing the dynamic resource allocation scheme and campus business demand data, generating a business and resource matching agent model based on intelligent optimization algorithm; processing the target campus information based on the business and resource matching agent model to generate campus resource scheduling information.

[0008] Another aspect of the present application, a multi-super convergence platform collaborative scheduling device for smart campus, characterized in that, comprising: an acquisition module for constructing a multi-super convergence platform resource pool and target campus information for smart campus; a processing module for performing resource sensing and dynamic modeling processing on the multi-super convergence platform resource pool data to generate target resource model data; processing the target resource model data, generating platform collaborative scheduling strategy parameters based on reinforcement learning; processing the platform collaborative scheduling strategy parameters to generate a dynamic resource allocation scheme; processing the dynamic resource allocation scheme and campus business demand data, generating a business and resource matching agent model based on intelligent optimization algorithm; processing the target campus information based on the business and resource matching agent model to generate campus resource scheduling information.

[0009] According to another aspect of the present application, an electronic device comprises: a first processor; and a memory for storing executable instructions of the first processor; wherein the first processor is configured to execute the executable instructions to implement the above-mentioned motion training three-dimensional reconstruction and biomechanical analysis method based on sparse multi-view video.

[0010] According to another aspect of the present application, a computer readable storage medium is provided, which stores a computer program, the computer program is executed by a second processor to implement the above-mentioned motion training three-dimensional reconstruction and biomechanical analysis method based on sparse multi-view video.

[0011] The multi-super-converged platform collaborative scheduling system and method for a smart campus provided by the present application realize efficient collaboration of campus computing, storage and network resources through resource perception, intelligent decision-making and dynamic adaptation. First, a multi-super-converged platform containing three types of resource pools is constructed, and resource parameters and business information are integrated; through edge node collection of indicators, LSTM and business cycle prediction load are fused, heterogeneous resources are adapted, protocols are converted, and target resource model data are generated. Based on reinforcement learning and business reward and punishment mechanism, a scheduling strategy is generated, a dynamic allocation scheme is generated and optimized by using computing power formula, storage algorithm, etc., a matching agent model is constructed by combining intelligent algorithms, resource elasticity is realized, scheduling information containing allocation details is generated, and utilization and adaptability are improved.

[0012] It should be understood that the foregoing general description and the following detailed description are only exemplary and explanatory, and are not limiting to the present disclosure. BRIEF DESCRIPTION OF DRAWINGS

[0013] Figure 1 A flowchart of a multi-super-converged platform collaborative scheduling method for a smart campus provided by an embodiment of the present application is shown; Figure 2 A structural schematic diagram of a multi-super-converged platform collaborative scheduling device for a smart campus provided by an embodiment of the present application is shown. DETAILED DESCRIPTION

[0014] The preferred embodiments of the present application are described below with reference to the accompanying drawings, and it should be understood that the preferred embodiments described herein are only used to illustrate and explain the present application, and are not used to limit the present application.

[0015] The motion training three-dimensional reconstruction and biomechanical analysis method based on sparse multi-view video according to the exemplary embodiments of the present application will be described below with reference to Figure 1 It should be noted that the following application scenarios are only shown for the purpose of facilitating understanding of the spirit and principles of the present application, and the embodiments of the present application are not limited in this respect. On the contrary, the embodiments of the present application are applicable to any applicable scenario.

[0016] In an embodiment, the present application also provides a multi-super-converged platform collaborative scheduling system and method for a smart campus. Figure 1 A flowchart of a multi-super-converged platform collaborative scheduling method for a smart campus according to an embodiment of the present application is shown.

[0017] S101, a multi-super-converged platform resource pool for a smart campus is constructed, and target campus information is constructed.

[0018] In one implementation, the multi-super-fusion platform resource pool integrates computing, storage, and network three types of core resources to form a standardized resource scheduling foundation. The computing resource pool is constructed to include server computing power, virtual machine load, and container deployment parameters. The server computing power covers the number of physical server CPU cores (such as 32 cores), memory capacity (such as 256 GB), and GPU computing power (such as NVIDIA A100 single card computing power); the virtual machine load records the CPU usage rate (real-time monitoring threshold ≤ 70%) and memory occupancy (such as 8 GB of memory allocated to a single virtual machine) of the teaching cloud desktop virtual machine; and the container deployment parameters include the image version (such as Docker image v2.3.1) and the number of instances (dynamically adjusted according to the course concurrency demand, with 1 instance deployed for every 50 students) of the teaching application container.

[0019] The storage resource pool is constructed to include distributed storage capacity, data read / write rate, redundancy backup strategy, and data cold / hot tiered storage threshold. The total size of the distributed storage capacity is 100 TB, divided into teaching resources (60 TB), scientific research data (30 TB), and management data (10 TB); the data read / write rate requires that the peak read rate of teaching video on-demand be ≥ 1000 MB / s and the scientific research data write rate be ≥ 500 MB / s; the redundancy backup strategy adopts a 3-copy mechanism (main copy + local copy + remote copy), and key examination data additionally enables a timed snapshot (snapshot generated at 2 a.m. every day); and the data cold / hot tiered storage threshold is set to automatically migrate cold data (access frequency < 1 time / month) to low-cost storage (such as a tape library) and retain hot data (access frequency ≥ 5 times / week) in the SSD storage layer.

[0020] The network resource pool is constructed to include bandwidth allocation, link delay, and topology connection relationship. The bandwidth allocation is divided according to business types, with teaching live broadcast guaranteed bandwidth ≥ 10 Gbps, scientific research data transmission reserved bandwidth 5 Gbps, and office network basic bandwidth 2 Gbps; the link delay requires that the delay between core nodes be ≤ 5 ms and the delay from the access layer to the core layer be ≤ 20 ms; and the topology connection relationship adopts a tree structure, with core switches (2 redundant) connecting various building aggregation switches, which are connected to classroom and laboratory access switches through fiber link to form a redundant link topology.

[0021] The target campus information covers business characteristics and demand data supporting resource scheduling, including teaching business demand data (such as course schedule, number of students for each course, online teaching platform concurrent access peak period), scientific research business demand data (such as peak computing power demand of scientific research computing tasks, data storage period), administrative management demand data (such as frequency of financial system data backup, personnel system access permission configuration), and campus business priority weight matrix (teaching business priority 0.8, scientific research business 0.7, management business 0.5).

[0022] S102, the multi-super-converged platform resource pool data is subjected to resource perception and dynamic modeling processing to generate target resource model data.

[0023] In an implementation, the multi-super-converged platform resource pool data is subjected to distributed real-time monitoring and collection processing based on edge computing nodes. Lightweight collection agents (such as exporter components based on Prometheus) are deployed on different super-converged nodes such as FusionCompute and VMware vSphere to accurately collect key resource indicators at a frequency of 1 second per time.

[0024] The server CPU utilization rate (such as the real-time CPU usage rate fluctuation of the teaching cloud server cluster being 30%-65%), the memory occupancy rate (the peak value of the memory occupancy of the scientific research virtual machine being 80%), the storage IOPS (the peak value of the IOPS of the library digital resource storage being 1200 times per second), and the network bandwidth usage rate (the core link bandwidth usage rate during the teaching live broadcast period being ≥70%) are generated to generate resource real-time monitoring indicators containing real-time numerical values, collection timestamps, and node identifiers.

[0025] The multi-super-converged platform resource pool data is subjected to load trend prediction processing based on fusion LSTM and campus business cycle characteristics. Historical load data (such as the average CPU load of the server per hour in the past 3 days) in the past 72 hours are input, and business characteristic parameters such as teaching schedules (such as theory courses from 1-4 am on Monday) and examination periods (such as a surge in access to review materials two weeks before the final examination week) are fused. The LSTM model is iteratively trained using a sliding window (with a window size of 24 hours).

[0026] A resource load prediction curve for the next 24 hours is generated through model training, wherein the prediction result shows that the CPU load of the teaching cloud desktop from 9:00 to 11:30 the next day will reach 75% (20% higher than the daily average), and the peak memory demand of the scientific research computing node from 14:00 to 17:00 will be 200 GB. The load prediction model output containing prediction values, confidence intervals, and characteristic influence weights is generated.

[0027] The multi-super-converged platform resource pool data is subjected to heterogeneous resource adaptation analysis processing based on device fingerprints and capability matrices. A hardware specification database covering super-converged devices of manufacturers such as Huawei, Dell, and VMware (recording parameters such as CPU model, memory type, and storage medium) is established, and a performance benchmark model (such as an algorithm conversion coefficient based on the FusionServer) is constructed. The resource capability matching degree score is calculated (the formula is , , ), the GPU computing power of Dell PowerEdge server is normalized to 92% of Huawei A100, realizing the performance measurement of heterogeneous devices, and generating resource heterogeneity adaptation results containing device fingerprint ID, normalized performance value, and adaptation priority.

[0028] The multi-hyper-converged platform resource pool data is processed based on dynamic protocol conversion gateway protocol compatibility conversion. A conversion gateway supporting multiple manufacturer private protocols is developed, with built-in VMware vSphere API, Huawei FusionCompute SDK, and other adaptation interfaces.

[0029] The FusionCompute "host sleep" private instruction is converted into the standardized scheduling instruction "power-saving-mode:enable", and the VMware "virtual machine migration" instruction is converted into "vm-migrate:{src:host1,dst:host2}", realizing real-time mutual recognition of different platform instructions, and generating cross-platform protocol conversion adaptation parameters containing original instructions, converted instructions, and conversion state.

[0030] Based on resource real-time monitoring indicators, load prediction model output, resource heterogeneity adaptation results, and cross-platform protocol conversion adaptation parameters, multi-dimensional weighted fusion processing is performed. A resource state evaluation matrix (rows represent resource types and columns represent evaluation dimensions) is established, and noise is eliminated through a data consistency verification mechanism (such as outlier detection and timestamp alignment).

[0031] The CPU utilization real-time value (weight 0.3), 24-hour prediction value (weight 0.2), heterogeneous normalization score (weight 0.3), and protocol conversion success rate (weight 0.2) are weighted calculated, and a high-quality resource node list with a comprehensive score ≥85 points is generated, and finally a target resource model data containing resource comprehensive state, data reliability, and abnormal warning identifier is generated.

[0032] S103, processing the target resource model data, generating platform collaborative scheduling strategy parameters based on reinforcement learning.

[0033] In one implementation, the resource state characteristic value and the business demand parameter in the target resource model data are processed by state space mapping to generate a resource state vector and a scheduling action space parameter. Two types of core parameters are extracted from the target resource model data, as follows: the resource state characteristic value includes server CPU utilization (real-time collection range 0%-100%), memory occupancy rate (comprehensive occupancy of physical memory and virtual memory), storage IOPS (input / output operations per second, including read operation and write operation), network bandwidth usage rate (real-time occupancy proportion of bidirectional transmission bandwidth), storage capacity remaining amount (available space proportion of distributed storage), virtual machine startup success rate (virtual machine deployment success number proportion in nearly 1 hour), container health state (ratio of surviving container instances to total instance number), and link connectivity (real-time on-off state of core network link), totaling 8 resource characteristic indexes.

[0034] The business demand parameter includes teaching business computing power demand (such as GPU computing power demand of online classroom), scientific research business storage capacity demand (such as storage occupancy space of experimental data), management business bandwidth demand (such as real-time bandwidth consumption of video conference), and mixed business priority weight (resource competition weight when multiple businesses are concurrent), totaling 4 business demand indexes.

[0035] The Min-Max normalization method is used to standardize the above parameters. The CPU utilization (actual range 30%-80%) is mapped to a continuous value in the range [0.3, 0.8] by formula; the memory occupancy rate (actual range 20%-90%) is mapped to a continuous value in the range [0.2, 0.9]; and the teaching business computing power demand (actual range 500-2000 GHz) is mapped to a continuous value in the range [0.25, 1.0]. After standardization, the 8 resource characteristics and 4 business demand parameters are integrated to generate a resource state vector with a length of 12, and each element in the vector is in the interval [0, 1].

[0036] The scheduling action space parameter is defined based on the scheduling capability of the campus hyper-converged platform. The action types include: computing power scheduling, such as “migrate virtual machine A to server B”, “add 2 instances to container cluster C”, and “release the computing power resources of idle server D”; storage scheduling, such as “migrate teaching video data from cold storage to hot storage” and “expand the capacity of scientific research data set E by 10 TB”; and network scheduling, such as “reserve 5 Gbps bandwidth for teaching live broadcast business” and “adjust the bandwidth allocation proportion of inter-building links”. Twenty discrete executable actions are defined, each corresponding to a unique action code (such as 001 for virtual machine migration and 002 for container expansion), forming the scheduling action space parameter and realizing the standardized description of resource scheduling actions.

[0037] The campus business priority weight and the resource utilization threshold are fused and processed based on the resource state vector and the scheduling action space parameter to generate a scenario-based reward and punishment coefficient. Based on the business demand parameter (such as the real-time demand proportion of teaching / scientific research / management business) in the resource state vector and the scheduling action space parameter (such as the execution range of various scheduling actions), two types of key constraint parameters are integrated: the campus business priority weight, which is set according to the business importance, wherein the teaching business priority weight is 0.8 (core guarantee business), the scientific research business is 0.7 (key support business), and the management business is 0.5 (regular guarantee business); the weight value is used for differentiated calculation of the reward and punishment strength. The resource utilization threshold is set in combination with the optimal interval of the super-converged platform performance, including CPU utilization ≤ 70% (to avoid overload and lag), memory utilization ≤ 80% (to reserve buffer space), storage IOPS load rate ≤ 75% (to guarantee read / write speed), and network bandwidth usage rate ≤ 80% (to prevent link congestion); the threshold is used as a judgment standard for whether the resource state is healthy.

[0038] The reward and punishment rule system covering business guarantee effect and resource utilization efficiency is constructed based on the above parameters, and the specific rules are as follows: business demand satisfaction reward, when the scheduling action makes the business resource satisfaction rate reach the target threshold, a positive reward is given according to the business priority. When the teaching business resource satisfaction rate is ≥ 95% (core demand), the reward value = business priority × (satisfaction rate - 0.9) × 20 (extra incentive for over-protection); when the satisfaction rate is between 90% and 95%, the reward value = business priority × (1 - resource utilization rate) × 10 (basic incentive). When the scientific research business resource satisfaction rate is ≥ 90%, the reward value = business priority × (satisfaction rate - 0.85) × 15; when the satisfaction rate is between 85% and 90%, the reward value = business priority × (1 - resource utilization rate) × 8. When the management business resource satisfaction rate is ≥ 85%, the reward value = business priority × (satisfaction rate - 0.8) × 10. Resource overload punishment, when the scheduling action causes the resource utilization rate to exceed the threshold, a punishment is given according to the exceeding range.

[0039] If the CPU utilization rate > 70% or the memory utilization rate > 80%, the punishment value = -business priority × (resource utilization rate - threshold) × 20 (the more serious the overload, the heavier the punishment); if the storage IOPS load rate > 75% or the bandwidth usage rate > 80%, the punishment value = -business priority × (load rate - threshold) × 15. Business under-protection punishment: when the scheduling action does not reach the minimum standard of business demand, the punishment mechanism is triggered. When the teaching business satisfaction rate < 90%, the punishment value = -business priority × (0.9 - satisfaction rate) × 25; when the scientific research business satisfaction rate < 80%, the punishment value = -business priority × (0.8 - satisfaction rate) × 20; when the management business satisfaction rate < 75%, the punishment value = -business priority × (0.75 - satisfaction rate) × 15.

[0040] To adapt to the time characteristics of campus business, the basic reward and punishment values are dynamically adjusted according to the scene to generate a scene-based reward and punishment coefficient: teaching peak period (such as weekdays 8:00-12:00, 14:00-18:00): the reward coefficient of teaching business is increased to 1.2 times to ensure that resources are preferentially guaranteed in the core period; research computing peak period (such as weekdays 20:00-2:00 the next morning): the reward coefficient of research business is increased to 1.1 times to adapt to the working habits of researchers at night; non-working period (such as weekends, statutory holidays): the punishment coefficient of all businesses is reduced to 0.8 times to allow resource scheduling to remain flexible; examination / major event period: the reward coefficient of teaching business is temporarily increased to 1.5 times, and the punishment coefficient is increased to 1.3 times to strengthen the guarantee. Through the above fusion processing, the business priority, resource threshold and scene characteristics are converted into quantitative reward and punishment signals to provide a clear optimization direction for the reinforcement learning model, and the generated scene-based reward and punishment coefficient is directly used for feedback value calculation in subsequent iterative training.

[0041] Based on the reinforcement learning algorithm framework (such as DeepQ-Network) and the scene-based reward and punishment coefficient, the resource allocation trial-and-error data and scheduling effect feedback values are iteratively trained. With the resource state vector as input and the scheduling action space as output, trial-and-error data (such as the change in resource utilization after "migrating virtual machine A to server B") is generated through interaction with the environment, and feedback rewards are calculated in combination with the scene-based reward and punishment coefficient.

[0042] The number of training iterations is set to 10,000 rounds, each round contains 500 time steps, and the ε-greedy strategy (ε linearly decays from 0.9 to 0.1) is used to explore the action space. When a scheduling action improves the resource balance rate by 15%, the feedback reward value is 12.8; when the action causes business interruption, the feedback penalty value is -25. The neural network parameters are optimized through gradient descent to generate policy optimization gradient parameters (such as Q value update gradient of each action, loss function convergence value).

[0043] The scene-based reward and punishment coefficient and the policy optimization gradient parameter are analyzed and processed. Combined with the distribution characteristics of the reward and punishment coefficient (such as the proportion of positive rewards for teaching business is 60%) and the convergence of the policy optimization gradient (such as the gradient norm is less than or equal to 0.01), stable and efficient scheduling strategy parameters are selected. Through analysis, core rules such as "teaching business is preferentially allocated GPU resources" and "research computing is expanded in the early morning" are determined to generate platform collaborative scheduling strategy parameters including resource allocation weight matrix (teaching business accounts for 40% in CPU resources), scheduling priority sequence (virtual machine migration > container expansion > storage adjustment), and threshold trigger condition (triggering re-scheduling when the resource satisfaction rate of a certain business is less than 90% for 5 consecutive minutes) to guide resource allocation and scheduling decisions of multiple hyper-converged platforms.

[0044] S104, the platform collaborative scheduling strategy parameters are processed to generate a dynamic resource allocation scheme.

[0045] In an embodiment, the resource quota parameter in the platform collaborative scheduling strategy parameter is processed by power allocation to generate a server power allocation table, wherein the server power allocation table is generated by the resource quota parameter through a power demand matching formula The calculation is composed of The allocation proportion of the i-th server is The service priority weight is The current service power demand value is The current CPU utilization rate of the i-th server is n, and Q is the total service power demand quota. The current power demand of a certain teaching service =0.8) is 1000GHz, the total power quota Q=1200GHz, and there are 3 available servers (n=3), and their CPU utilization rates are =50%、 =60%、 =40%.

[0046] The numerator is calculated, specifically, the server 1 is 0.8 1000 (1-0.5)=400; the server 2 is 0.8 1000 (1-0.6)=320; the server 3 is 0.8 1000 (1-0.4)=480; the denominator sum is 400+320+480=1200.

[0047] The allocation proportion is =400 / 1200 1200=400GHz; =320 / 1200 1200=320GHz; =480 / 1200 1200=480GHz; The generated server power allocation table clearly indicates the specific power allocation of each server.

[0048] ​The storage allocation parameter in the platform collaborative scheduling strategy parameter is subjected to storage resource mapping processing, and a distributed storage allocation scheme is generated, wherein the distributed storage allocation scheme is composed of the capacity allocation ratio in the storage allocation parameter and calculated by a storage load balancing algorithm. A scientific research business needs to allocate 50TB of storage resources, and the capacity allocation ratio in the storage allocation parameter is “hot storage layer: cold storage layer = 7:3”. Through the storage load balancing algorithm, the current load rate of each storage node is calculated (for example, the load rate of node A is 60%, and the load rate of node B is 40%); the hot storage capacity of 50 70% is allocated to nodes B (20TB) and A (15TB) with lower load; the cold storage capacity of 50 30% is allocated to low-cost tape library nodes; the generated distributed storage allocation scheme includes the storage quota of each node and the data migration path.

[0049] The server computing power allocation table and the distributed storage allocation scheme are subjected to network link adaptation processing, and a network resource scheduling plan is generated, wherein the network resource scheduling plan is composed of the server computing power allocation table, the distributed storage allocation scheme and the network topology structure, and is fused by a bandwidth demand interpolation algorithm. It is known that the computing power allocation of server A to storage node X is 400GHz, the storage read-write demand is 200MB / s, and the link bandwidth upper limit in the network topology is 1Gbps. Based on the computing power allocation table, the computing power transmission bandwidth demand is calculated: 400GHz corresponds to about 300Mbps; based on the storage allocation scheme, the storage bandwidth demand is calculated: 200MB / s≈1600Mbps (exceeding the link upper limit); through the bandwidth demand interpolation algorithm adjustment: according to the link upper limit 1Gbps, 600Mbps is allocated to the standby link between server A and storage node Y. The generated network resource scheduling plan includes the bandwidth allocation of the main and standby links and the time period occupation table.

[0050] The network resource scheduling plan is subjected to business priority filtering processing, and an initial dynamic resource allocation scheme is generated, wherein the business priority filtering processing converts the network resource scheduling plan into a preliminary result by setting a screening threshold based on the campus business priority weight. Based on the campus business priority weight system (teaching 0.8, scientific research 0.7, management 0.5), combined with the resource shortage degree, a dynamic screening threshold is set. When the overall utilization rate of network resources is ≤70%, the threshold is set to 0.5 (all businesses are retained); when the utilization rate is >70% and ≤90%, the threshold is set to 0.6 (low-priority management businesses are filtered); when the utilization rate is >90%, the threshold is set to 0.75 (only core teaching businesses are retained). The threshold parameter is stored in the strategy configuration library, and automatic adjustment is supported according to scenes such as semesters and holidays.

[0051] Extract the resource demand records of various businesses (including bandwidth, link occupation time, etc.) from the network resource scheduling plan, and compare and filter them according to priority weight and screening threshold: when the business priority ≥ screening threshold, keep its complete resource demand parameters; when the business priority < screening threshold, mark it as "delayable business" and only keep the basic guarantee demand (such as management business only keeps 20% basic bandwidth). Specifically, the network resource utilization rate during the examination period is 85%, and the screening threshold is automatically set to 0.6. The network resource scheduling plan originally contains: teaching live broadcast business (priority 0.8, demand bandwidth 5Gbps); scientific research data transmission (priority 0.7, demand bandwidth 3Gbps); administrative office network (priority 0.5, demand bandwidth 2Gbps). After filtering, the complete demand of teaching and scientific research business is kept, and the administrative office network only keeps 0.4Gbps basic bandwidth, and 1.6Gbps unnecessary demand is filtered out.

[0052] For the resource demand of the filtered business, allocate the available resource quota according to the priority weight ratio, and the calculation formula is , wherein is the allocated bandwidth of the bth business, is its priority weight, is the weight sum of the retained businesses, is the total amount of available bandwidth after filtering. For example, the remaining available bandwidth after filtering is 10Gbps, and the weight sum of the retained businesses = 0.8+0.7=1.5. The teaching business is allocated 10 0.8 / 1.5≈5.33Gbps; the scientific research business is allocated 10 0.7 / 1.5≈4.67Gbps.

[0053] Integrate the filtered business resource demand, weighted allocation result and link occupation time sequence to generate an initial dynamic resource allocation scheme. The scheme includes: business priority label (such as "P0 core guarantee" and "P1 key support"); resource allocation details (bandwidth value, link identification, effective period); filtering explanation (list of delayed businesses and reasons). The initial dynamic resource allocation scheme clearly marks "teaching live broadcast business (P0) occupies 5.33Gbps bandwidth of link A from 8:00 to 12:00, and scientific research data transmission (P1) occupies 4.67Gbps bandwidth of link B from 14:00 to 20:00", and attaches detailed explanation of the delayed allocation of management business.

[0054] Perform resource conflict resolution processing on the initial dynamic resource allocation scheme to generate a final dynamic resource allocation scheme. The resource conflict resolution processing checks and corrects the initial result through resource capacity constraint equation and business demand satisfaction evaluation. Based on the resource capacity constraint equation, the initial scheme is checked for compliance, and the core equation is , The application quantity of the kth type of resource for the bth type of service, The total available capacity of the kth type of resource. In the initial scheme, the teaching service applies for 80 units of GPU resources, and the scientific research service applies for 30 units, while the actual available GPU is only 90 units (80+30=110>90), triggering a capacity conflict.

[0055] The satisfaction of the conflicting service is quantitatively evaluated, and the formula is Set the minimum acceptable satisfaction threshold (teaching service ≥80%, scientific research service ≥70%, and management service ≥60%). If the teaching service GPU is reduced to 60 units, the satisfaction is 60 / 80=75% (lower than the threshold of 80%), and further adjustment is needed; if it is reduced to 65 units, the satisfaction is 81.25% (meets the requirements).

[0056] According to the check result and the satisfaction threshold, a hierarchical adjustment strategy is used to eliminate the conflict: adjustment in descending order of priority, priority protection of high-priority services, and resource reduction of low-priority services. In the above GPU conflict, the teaching service demand is prioritized, and the scientific research service GPU is reduced from 30 units to 25 units (satisfaction 25 / 30≈83.3%), and the total allocation amount is 65+25=90 units (meets the capacity constraint). Resource replacement compensation: for the services that are reduced in resources, compensation is made through other resource types.

[0057] If there is redundancy in the storage resources of the scientific research service, part of the computing tasks can be migrated to the CPU cluster to make up for the loss of computing power due to the reduction of GPU. Time shift scheduling: for conflicts that cannot be solved by resource substitution, adjust the execution time of the service. Part of the scientific research computing tasks are shifted from daytime to early morning (GPU resource idle period) to avoid competing for resources with teaching services.

[0058] The adjusted scheme is re-executed for capacity constraint verification and satisfaction evaluation until all services meet the following conditions: the total amount of resource allocation ≤ available capacity; the satisfaction of the service ≥ the minimum threshold. The final dynamic resource allocation scheme generated includes: conflict resolution details (adjusted service, resource type and amplitude); the final resource allocation and time arrangement of each service; satisfaction evaluation report (to ensure that the core service satisfaction ≥85%). The final scheme determines that the teaching service is allocated 65 units of GPU (satisfaction 81.25%), the scientific research service is allocated 25 units of GPU (satisfaction 83.3%), and the computing demand of 10 units of GPU of the scientific research service is shifted to the early morning for execution, which not only meets the capacity constraint, but also guarantees the core demand of the service.

[0059] S105, process the dynamic resource allocation scheme and the campus service demand data, and generate a service and resource matching agent model based on an intelligent optimization algorithm.

[0060] In an embodiment, the computing power / storage / network resource quota parameters in the dynamic resource allocation scheme are subjected to service load mapping processing to generate a service and resource basic matching matrix, wherein the service and resource basic matching matrix is composed of resource quota parameters calculated by a service load characteristic matching formula. The matrix is calculated by the service load characteristic matching formula, and the formula is: wherein is the matching degree of the bth service and the rth resource, is the quota parameter of service b to resource r, is the priority weight of service b, is the maximum quota value of the same type of resource.

[0061] For example, teaching service =0.8 obtains computing power quota 800GHz, storage quota 60TB, bandwidth quota 5Gbps, and the maximum quota of the same type of resource is 1000GHz, 100TB, and 10Gbps respectively. The matching degree is calculated as follows: computing power =(800 0.8) / 1000=0.64; storage =(60 0.8) / 100=0.48; bandwidth =(5 0.8) / 10=0.4 A 3x3 service and resource basic matching matrix (rows represent services and columns represent resources) is generated, and the matrix element value range is [0, 1].

[0062] The load fluctuation characteristics and peak period parameters in the campus service demand data are subjected to fluctuation rule extraction processing to generate a service load fluctuation prediction model, wherein the service load fluctuation prediction model is generated by fitting the load fluctuation characteristics with time series decomposition algorithm and periodic factors. The load data is decomposed into trend item, periodic item and residual item by time series decomposition algorithm (STL decomposition), and a prediction model is generated by fitting with periodic factors wherein is the predicted load, is the trend item, is the periodic item (such as weekday / weekend period), is the residual item.

[0063] The CPU load data of teaching service in the past 3 months is processed to decompose the trend item (2% growth per week) and the periodic item (peak value appears at 9:00-11:30 on weekdays), and a load prediction model for the next week is fitted and generated. The prediction shows that the load peak on Monday morning will reach 75%, which is 20% higher than the average.

[0064] The business and resource basis matching matrix and the business load fluctuation prediction model are associated with rule mining processing to generate a resource elastic scaling trigger condition, wherein the resource elastic scaling trigger condition is composed of a matching matrix correlation degree score and a fluctuation threshold value through rule inference engine fusion. Based on the Apriori algorithm, the business and resource basis matching matrix is associated with rule mining. First, the minimum support (the frequency of business and resource association ≥ 10%) and the minimum confidence (the probability of rule establishment ≥ 60%) are set. The correlation degree score of the business characteristics and the resource state in the matching matrix is calculated, and the formula is: correlation degree score = support 0.4 + confidence 0.6, wherein the support reflects the frequency of rule occurrence, and the confidence reflects the reliability of the rule.

[0065] In the association analysis of teaching business and computing power resources, the support of the simultaneous occurrence of “computing power matching degree > 0.6” and “load fluctuation > 15%” is 12% (satisfying ≥ 10%), and the confidence is 75% (satisfying ≥ 60%). Therefore, the correlation degree score = 12% × 0.4 + 75% × 0.6 = 0.048 + 0.45 = 0.498 (initially not reaching the threshold value). Further combined with the business priority weight correction, the score is improved to 0.85. The correction formula is: corrected score = correlation degree score × (1 + / 2), The priority of the teaching business is 0.8.

[0066] Combined with the fluctuation amplitude, duration and other parameters output by the business load fluctuation prediction model, the differentiated fluctuation threshold value is set: for teaching business, load fluctuation > 15% (core business sensitivity is high), duration ≥ 5 minutes; for scientific research business, load fluctuation > 20% (allowing certain elasticity), duration ≥ 10 minutes; for management business, load fluctuation > 25% (non-core business threshold relaxation), duration ≥ 15 minutes. The threshold value is stored in the dynamic configuration library, and the sensitivity can be automatically adjusted according to the semester period (such as the opening season, the examination week) (such as the teaching business fluctuation threshold being reduced to 10% in the examination week).

[0067] The correlation degree score (≥ 0.7 threshold value) and the fluctuation threshold value are input into the rule inference engine to generate the resource elastic scaling trigger condition through the “condition-action” logic. The inference engine includes three levels of rule library: basic trigger rule, single business load and resource matching degree association (such as “teaching business storage matching degree < 0.3 and load fluctuation > 15%”); combined trigger rule, association in multiple business competition scenarios (such as “teaching and scientific research business bandwidth matching degree > 0.6 and total load fluctuation > 20%”); time period correction rule, dynamic adjustment combined with the campus time period characteristics (such as “non-working time period trigger threshold value is automatically increased by 20%”).

[0068] The engine matches the combination condition of "teaching business computing power matching degree > 0.6 (correlation degree score 0.85), predicted load fluctuation 18% (> 15% threshold), duration prediction 8 minutes (≥ 5 minutes)", generates the trigger condition: "when the teaching business CPU utilization rate is ≥ 70% for 5 consecutive minutes and the predicted increase within 1 hour is ≥ 10%, trigger computing power expansion, the expansion range is 1.2 times the current load", and adds the constraint: "the CPU utilization rate of a single server after expansion should not exceed 85%".

[0069] The generated trigger condition is simulated and checked, and the false trigger rate (target ≤ 5%) and the missed trigger rate (target ≤ 3%) of the rule are verified through historical data playback. If the false trigger rate of a certain trigger condition reaches 8% in the test, the constraint condition (such as "simultaneously satisfying memory utilization rate ≥ 60%") is added to optimize it until the index requirements are met. The initial "scientific research business computing power expansion trigger condition" has a missed trigger rate of 6% in the test, and after adding the auxiliary condition of "GPU load rate ≥ 75%", the missed trigger rate is reduced to 2.5%, and finally it is included in the resource elasticity expansion rule library.

[0070] The resource elasticity expansion trigger condition is parameter optimized to generate an initial business and resource matching agent model, wherein the parameter optimization process optimizes the expansion threshold and response delay parameters through a particle swarm optimization algorithm. The core parameters to be optimized and their value ranges are as follows: the expansion threshold, which is the load critical value that triggers resource elasticity expansion, is set to search in the interval of 60%-80% for teaching business, 65%-85% for scientific research business, and 70%-90% for management business; the response delay, which is the interval time from the trigger condition being met to the resource adjustment being executed, is set to search in the interval of 30-180 seconds (teaching business prioritizes real-time performance, and the interval is set to 30-120 seconds). Each parameter combination constitutes a "particle" in the PSO algorithm, and the initial population size is set to 30 particles (i.e. 30 parameter combinations).

[0071] The weighted sum of resource waste rate and business delay rate is taken as the optimization target, and the objective function is , where = 0.6 (resource utilization efficiency weight is higher), = 0.4. The resource waste rate refers to the ratio of the amount of excess allocated resources to the actual demand, and the formula is ; the business delay rate refers to the proportion of the time length of business response delay caused by the non-timely adjustment of resources, and the formula is .

[0072] The fitness value of the particle directly adopts the objective function value J, and the smaller the value, the better the parameter combination. The PSO iterative optimization process is as follows: 30 sets of initial parameters (such as teaching business initial threshold 70%, delay 120 seconds) are randomly generated, the J value corresponding to each set of parameters is calculated, and the global optimal particle (initial optimal J = 0.6*15% + 0.4*8% = 12.2%) is recorded.

[0073] In each iteration, the particle adjusts the flight speed and direction according to the individual optimal position and the global optimal position, and the speed update formula is: wherein = 0.8 (inertia weight), = = 2 (learning factor), (r1, r2) are random numbers in [0, 1], is the individual optimal position, is the global optimal position. The position update formula is . If the updated parameter exceeds the search interval (such as the threshold being lower than 60%), it is forced to be truncated to the interval boundary value. For teaching business, after 50 iterations, the particle converges to the optimal parameters: scaling threshold 68%, response delay 90 seconds. At this time, the resource waste rate is reduced to 10% (due to the reduction of excess allocation caused by the reduction of threshold), the business delay rate is reduced to 5% (due to the acceleration of response speed), the objective function value (J = 0.6*10% + 0.4*5% = 8%) is reduced by 40% compared with the initial value (12.2%).

[0074] Initial agent model generation, the optimal parameters obtained by PSO optimization (such as the scaling threshold of each business type, response delay) are embedded into the resource elasticity scaling trigger condition to form an initial business and resource matching agent model. The model includes: parameter configuration table of each business (such as teaching business: threshold 68%, delay 90 seconds; scientific research business: threshold 72%, delay 120 seconds); mapping relationship between parameters and objective function value (used for subsequent model iteration); constraint condition description (such as “response delay should not be lower than 30 seconds to avoid frequent scheduling”). The initial model can automatically determine whether to trigger computing power expansion, storage migration or bandwidth adjustment according to real-time business data and resource state, and provide a basis for decision-making for campus resource dynamic scheduling.

[0075] Adaptability verification processing is performed on the initial business and resource matching agent model to generate a final business and resource matching agent model, wherein the adaptability verification processing iteratively corrects the initial model through campus typical business scenario simulation test and resource utilization rate improvement rate evaluation.

[0076] Adaptability verification processing is performed on the initial business and resource matching agent model to generate a final business and resource matching agent model, wherein the adaptability verification processing iteratively corrects the initial model through campus typical business scenario simulation test and resource utilization rate improvement rate evaluation.

[0077] ​S106, processing the target campus information based on the service and resource matching agent model to generate campus resource scheduling information.

[0078] In an implementation, the service priority parameter and real-time load data in the target campus information are processed by feature extraction based on the service and resource matching agent model, and the resource demand clustering algorithm Identify the core scheduling object from the service feature data. Extract three types of core service feature parameters from the target campus information, including the following, CPU demand feature : CPU average occupancy rate (range 0%-100%) during service operation, such as the CPU occupancy rate peak of 60% for live teaching service; memory demand feature : average memory resource occupancy ratio (range 0%-100%) of the service, such as 70% memory occupancy rate for scientific research computing tasks; bandwidth demand feature : average bandwidth usage rate (range 0%-100%) during service data transmission, such as 80% bandwidth usage rate for online classroom. Standardize the extracted original feature values (map to [0,1] interval through Min-Max normalization) to eliminate the influence of dimension difference on clustering results.

[0079] Group the standardized service feature data using the K-means clustering algorithm, the formula is , where is the resource demand clustering center of the jth service, representing the comprehensive resource demand intensity of the service; is the sample number of the jth service (such as 12 samples for teaching service, corresponding to 12 online courses); =0.4, =0.3, =0.3 are feature weights (CPU demand weight is the highest, as it has the greatest impact on service smoothness).

[0080] The clustering process is as follows, initialize 3 clustering centers (corresponding to initial classification of teaching, scientific research, and management services); calculate the Euclidean distance between each service sample and the clustering center, and divide the sample into categories according to the nearest distance principle; update the clustering center according to the formula until the center change is ≤0.01 (convergence condition). Determine the core scheduling object by comparing the numerical values of the clustering centers, and the service category with the highest clustering center is determined as the core scheduling object. The teaching service clustering center is calculated as . The scientific research service clustering center =55% (the sample contains 8 scientific research projects); the management service clustering center =40% (the sample contains 5 management systems); because > , determine the teaching business as the core scheduling object, and give the highest priority in resource allocation (such as reserving 30% redundant resources to respond to sudden demand).

[0081] To adapt to the dynamic changes of campus business, the clustering algorithm is re-executed every 24 hours to update the core scheduling object: during the examination period, the teaching business cluster center may rise to 75%, maintaining the core position; in the research completion stage, the research business cluster center may rise to 68%, temporarily becoming the core scheduling object. Through the above processing, accurate identification of core business is realized, ensuring that resources are tilted to high-demand and high-priority businesses, and improving the resource scheduling efficiency of the campus super-converged platform.

[0082] Dynamic threshold adjustment algorithm Optimize the resource scheduling trigger condition , the iteration termination condition is resource utilization fluctuation ≤5%, and the scheduling trigger threshold parameter is generated. The dynamic threshold adjustment algorithm is used to optimize the resource scheduling trigger condition, and the formula is , where is the threshold after iteration, is the current threshold, =0.1 is the sensitivity parameter, is the current resource utilization fluctuation standard deviation, =5% is the target fluctuation threshold, is the historical maximum fluctuation standard deviation, and the iteration termination condition is 5%.

[0083] The initial scheduling trigger threshold =70%, in the first iteration =8%, =10%, then: =70%+0.1 =70.3%, after 5 iterations, =4.8% (≤5%), the final threshold =70.5%, and the scheduling trigger threshold parameter is generated.

[0084] ​The campus resource scheduling execution matrix is constructed, combined with the multi-super fusion platform topology structure and the business and resource adaptation rules, to calculate the computing power allocation weight, the storage mapping relationship and the bandwidth reservation ratio. Among them, the computing power allocation adopts the load balancing coefficient correction method, and the bandwidth reservation adopts the business priority weighting algorithm. The construction of the campus resource scheduling execution matrix is the core link to realize the accurate allocation of multi-super fusion platform resources, which needs to be combined with the platform topology structure (such as server cluster level, storage node distribution, network link connection relationship) and the business and resource adaptation rules (such as teaching business priority adaptation to GPU server, scientific data priority storage in high IOPS node), through the load balancing coefficient correction method to calculate the computing power allocation weight, through the business priority weighting algorithm to calculate the bandwidth reservation ratio, and finally form a matrix containing the allocation relationship of computing power, storage and network resources. The specific processing process is as follows: The computing power allocation weight is generated by combining the load state of the fusion server and the business adaptation characteristics, and the formula is Wherein, is the real-time load rate of the i-th server (such as server A load rate 50%); is the server computing power benchmark value (taking Huawei FusionServer as the standard, unit: relative computing power value, such as server A is 100, server B is 80); is the business priority (teaching 0.8, scientific research 0.7); is the adaptation degree score of business and server (0-1 points, such as teaching business and GPU server adaptation degree 0.9); =0.6 (the load balancing weight is higher than the adaptation degree weight, and the server overload is avoided in priority).

[0085] A certain teaching business =0.8 needs to be allocated to 3 servers, and the parameters are as follows: server A: , , ; server B: , , ; server C: , , . Calculate the load balancing term, server A: ; server B: ≈0.23; server C: ≈0.12. Calculate the adaptation degree term, server A: ≈0.16; server B: ≈0.14; server C: ≈0.12. Total weight, , , (after rounding, it is consistent with example one).

[0086] The bandwidth reservation ratio is used to guarantee the sudden demand of services. The formula is: ;in, Peak bandwidth requirements for services (e.g. 5Gbps for live teaching); The currently allocated bandwidth (for example, 4 Gbps has been allocated for teaching services); =1.2 is the redundancy coefficient (reserving 20% ​​to cope with fluctuations); m is the total number of business types (such as teaching and scientific research).

[0087] In addition to teaching business, new management business =0.5, =2Gbps, =1.5Gbps: =[0.8×(5-4)+0.7×(3-2)+0.5×(2-1.5)] 1.2=(0.8+0.7+0.25)×1.2=2.1Gbps.

[0088] The scheduling execution matrix is ​​generated. The execution matrix is ​​a 3×3 two-dimensional table (rows: teaching, scientific research, management business; columns: computing power, storage, bandwidth), and the element values ​​are specific allocation ratios or quotas: Based on the scheduling execution matrix and real-time resource monitoring data, combined with the characteristics of campus business periods, campus resource scheduling information is generated, including resource allocation details, scheduling execution sequence, and load warning thresholds. Based on the scheduling execution matrix and real-time resource monitoring data, combined with the characteristics of campus business periods, campus resource scheduling information is generated as follows: Resource allocation details: Teaching business is allocated 42% computing power, 60% hot storage, and 5.33Gbps bandwidth; scientific research business is allocated 38% computing power, 30% hot storage, and 4.67Gbps bandwidth. Scheduling execution sequence: GPU acceleration is enabled for teaching business from 8:00 AM to 12:00 PM, and additional computing power is used for scientific research business from 8:00 PM to 2:00 AM the next day. Load warning threshold: CPU utilization ≥ 70.5% triggers capacity expansion, and memory utilization ≥ 80% triggers a warning. The scheduling information clearly states that "the live teaching service will prioritize occupying 42% of the computing power of server A and 5.33Gbps bandwidth of link A from 8:00 to 12:00. When the CPU utilization rate is ≥70.5% for three consecutive minutes, it will automatically expand to 20% redundant computing power of server B."

[0089] The application realizes efficient cooperation of campus computing, storage and network resources through resource perception, intelligent decision and dynamic adaptation. The system first constructs a multi-super-converged platform containing computing, storage and network resource pools, integrates server computing power, storage capacity, bandwidth and other parameters and campus business information. Then, the edge node collects resource indicators in real time, predicts the load by combining LSTM and business cycle characteristics, adapts heterogeneous resources and converts protocols to generate target resource model data.

[0090] Based on reinforcement learning, the scheduling strategy parameters are generated in combination with the campus business reward and punishment mechanism, the dynamic resource allocation scheme is generated through the computing power allocation formula and storage balancing algorithm, and the conflict is eliminated and optimized. Finally, combined with the intelligent optimization algorithm, the business and resource matching agent model is constructed to realize resource elastic scaling, generate scheduling information including allocation details, timing and early warning thresholds, and improve resource utilization and business adaptability.

[0091] In an embodiment, as shown in Figure 2 The application also provides a multi-super-converged platform cooperative scheduling device for a smart campus, which comprises: The acquisition module 201 is configured to construct a smart campus multi-super-converged platform resource pool and target campus information. The processing module 202 is configured to perform resource perception and dynamic modeling processing on the multi-super-converged platform resource pool data to generate target resource model data, process the target resource model data, generate platform cooperative scheduling strategy parameters based on reinforcement learning, process the platform cooperative scheduling strategy parameters, generate a dynamic resource allocation scheme, process the dynamic resource allocation scheme and campus business demand data, generate a business and resource matching agent model based on an intelligent optimization algorithm, process the target campus information based on the business and resource matching agent model, and generate campus resource scheduling information.

[0092] The computer readable storage medium provided by the above embodiments of the application and the method of three-dimensional reconstruction and biomechanical analysis of sports training based on sparse multi-view video provided by the embodiments of the application have the same beneficial effects as the method adopted, run or implemented by the application program stored therein.

[0093] The various embodiments in the present application are described in a related manner, and the same or similar parts among the various embodiments can be referred to each other. Each embodiment focuses on the difference from other embodiments. In particular, for the evaluation method, the electronic device, the electronic equipment, and the readable storage medium embodiments of the three-dimensional reconstruction and biomechanical analysis of sports training based on sparse multi-view video, since they are basically similar to the above-mentioned embodiments of the three-dimensional reconstruction and biomechanical analysis of sports training based on sparse multi-view video, the description is relatively simple, and the relevant parts can be referred to the above-mentioned embodiments of the three-dimensional reconstruction and biomechanical analysis of sports training based on sparse multi-view video.

Claims

1. A multi-hyper-convergence platform collaborative scheduling method for smart campuses, characterized in that: include: Construct a multi-hyper-convergence platform resource pool for a smart campus and target campus information. The multi-hyper-convergence platform resource pool includes a computing resource pool, a storage resource pool, and a network resource pool. The computing resource pool includes server computing power, virtual machine load, and container deployment parameters. The storage resource pool includes distributed storage capacity, data read and write rates, redundant backup strategies, and data hot and cold tiered storage thresholds. The network resource pool includes bandwidth allocation, link latency, and topological connection relationships. Perform resource perception and dynamic modeling on resource pool data from multiple hyper-converged platforms to generate target resource model data. The target resource model data consists of real-time resource monitoring indicators, load prediction model output, resource heterogeneity adaptation results, and cross-platform protocol conversion adaptation parameters. Process the target resource model data and generate platform collaborative scheduling strategy parameters based on reinforcement learning. The reinforcement learning process introduces a reward and punishment mechanism for campus business scenarios. Process the platform collaborative scheduling strategy parameters to generate a dynamic resource allocation plan; Dynamic resource allocation plans and campus business demand data are processed to generate a business and resource matching agent model based on an intelligent optimization algorithm. The business and resource matching agent model is used to characterize the adaptation relationship between various campus businesses and hyper-converged resources, and new business load fluctuation prediction and resource elastic scaling rules are added. The target campus information is processed based on the business and resource matching agent model to generate campus resource scheduling information.

2. The method according to claim 1, wherein Perform resource awareness and dynamic modeling on multiple hyper-converged platform resource pool data to generate target resource model data, including: Distributed real-time monitoring, collection, and processing of data from multiple hyper-converged platform resource pools based on edge computing nodes. Lightweight collection agents deployed on each hyper-converged node collect server CPU utilization, memory occupancy, storage IOPS, and network bandwidth usage at a frequency of 1 second to generate real-time resource monitoring indicators. Load trend prediction is performed on resource pool data from multiple hyper-converged platforms by integrating LSTM with campus business cycle characteristics. Input includes 72 hours of historical load data and business characteristic parameters of the teaching schedule and exam cycle. Through sliding window iterative training, a resource load forecast curve for the next 24 hours is generated, generating the load prediction model output. Perform heterogeneous resource adaptation analysis and processing based on device fingerprints and capability matrices on resource pool data from multiple hyper-converged platforms, establish a hardware specification database and performance benchmark model for hyper-converged devices from different manufacturers, normalize the performance of heterogeneous devices by calculating resource capability matching scores, and generate resource heterogeneity adaptation results; Perform protocol compatibility conversion processing on multiple hyper-converged platform resource pool data based on a dynamic protocol conversion gateway, achieve real-time conversion of private instruction sets to standardized scheduling instructions, and generate cross-platform protocol conversion adaptation parameters; Multi-dimensional weighted fusion processing is performed based on real-time resource monitoring indicators, load prediction model output, resource heterogeneity adaptation results and cross-platform protocol conversion adaptation parameters. By establishing a resource status assessment matrix and a data consistency verification mechanism, data noise and heterogeneous deviation are eliminated to generate target resource model data.

3. The method according to claim 2, wherein Process the target resource model data and generate platform collaborative scheduling strategy parameters based on reinforcement learning, including: Perform state space mapping on the resource state feature values ​​and business demand parameters in the target resource model data to generate resource state vectors and scheduling action space parameters; Based on the resource state vector and scheduling action space parameters, the reward and punishment rules are integrated with the campus business priority weight and resource utilization threshold to generate scenario-based reward and punishment coefficients. Based on the reinforcement learning algorithm framework and scenario-based reward and punishment coefficients, iterative training is performed on resource allocation trial and error data and scheduling effect feedback values ​​to generate policy optimization gradient parameters; The scenario-based reward and punishment coefficients and policy optimization gradient parameters are analyzed and processed to generate platform collaborative scheduling policy parameters. The platform collaborative scheduling policy parameters are used to characterize the resource allocation rules and scheduling priorities of multiple hyper-converged platforms.

4. The method according to claim 1, wherein Process the platform collaborative scheduling policy parameters to generate a dynamic resource allocation plan, including: The resource quota parameters in the platform collaborative scheduling strategy parameters are processed for computing power allocation to generate a server computing power allocation table, where the server computing power allocation table is composed of resource quota parameters through the computing power demand matching formula Calculation composition, The ratio of the calculation examples allocated to the i-th server, is the business priority weight, is the current computing power demand value of the business, is the current CPU utilization of the i-th server, n is the total number of available servers, and Q is the total computing power required for the business; Performing storage resource mapping processing on the storage allocation parameters in the platform collaborative scheduling policy parameters to generate a distributed storage allocation plan, wherein the distributed storage allocation plan is composed of the capacity allocation ratio in the storage allocation parameters calculated by the storage load balancing algorithm; Perform network link adaptation on the server computing power allocation table and the distributed storage allocation plan to generate a network resource scheduling plan. The network resource scheduling plan is formed by integrating the server computing power allocation table, the distributed storage allocation plan, and the network topology structure through a bandwidth demand interpolation algorithm. Performing service priority filtering on the network resource scheduling plan to generate an initial dynamic resource allocation plan, wherein the service priority filtering converts the network resource scheduling plan into a preliminary result by setting a screening threshold based on the campus service priority weight; The initial dynamic resource allocation plan is processed for resource conflict resolution to generate a final dynamic resource allocation plan. The resource conflict resolution process verifies and corrects the initial result through the resource capacity constraint equation and business demand satisfaction evaluation.

5. The method according to claim 1, wherein Process dynamic resource allocation plans and campus business demand data, and generate business and resource matching agent models based on intelligent optimization algorithms, including: Perform business load mapping on the computing power / storage / network resource quota parameters in the dynamic resource allocation plan to generate a business and resource basic matching matrix. The business and resource basic matching matrix is ​​composed of resource quota parameters calculated using the business load feature matching formula. The load fluctuation characteristics and peak period parameters in campus business demand data are extracted and processed to generate a business load fluctuation prediction model. The business load fluctuation prediction model is generated by fitting the load fluctuation characteristics through a time series decomposition algorithm and a period factor. Association rule mining is performed on the business and resource basic matching matrix and the business load fluctuation prediction model to generate resource elastic scaling trigger conditions. The resource elastic scaling trigger conditions are formed by fusing the matching matrix association score and the fluctuation threshold through the rule inference engine. Optimize the parameters of resource elastic scaling trigger conditions to generate an initial service and resource matching proxy model. The parameter optimization process uses a particle swarm optimization algorithm to optimize scaling thresholds and response delay parameters. The adaptability verification process is performed on the initial business and resource matching agent model to generate the final business and resource matching agent model. The adaptability verification process iteratively corrects the initial model through simulation tests of typical campus business scenarios and evaluation of resource utilization improvement rate.

6. The method according to claim 5, wherein The target campus information is processed based on the business and resource matching agent model to generate campus resource scheduling information, including: Based on the business and resource matching agent model, the business priority parameters and real-time load data in the target campus information are extracted and processed, and the resource demand clustering algorithm is used to Identify core scheduling objects from business feature data; Adopt dynamic threshold adjustment algorithm Optimize resource scheduling trigger conditions to is the sensitivity parameter, and the iteration termination condition is that the resource utilization fluctuation is ≤5%, which generates the scheduling trigger threshold parameter; Build a campus resource scheduling execution matrix, combining the multi-hyper-converged platform topology and business and resource adaptation rules to calculate computing power allocation weights, storage mapping relationships, and bandwidth reservation ratios. Computing power allocation uses a load balancing coefficient correction method, and bandwidth reservation uses a business priority weighted algorithm. Based on the scheduling execution matrix and real-time resource monitoring data, combined with the characteristics of campus business periods, campus resource scheduling information is generated, including resource allocation details, scheduling execution timing, and load warning thresholds.

7. A multi-hyper-convergence platform collaborative scheduling device for smart campuses, characterized in that: The device comprises: Acquisition module, used to build a smart campus multi-hyper-convergence platform resource pool and target campus information; The processing module is used to perform resource perception and dynamic modeling processing on the resource pool data of multiple hyper-converged platforms to generate target resource model data; process the target resource model data to generate platform collaborative scheduling strategy parameters based on reinforcement learning; process the platform collaborative scheduling strategy parameters to generate a dynamic resource allocation plan; process the dynamic resource allocation plan and campus business demand data to generate a business and resource matching agent model based on an intelligent optimization algorithm; process the target campus information based on the business and resource matching agent model to generate campus resource scheduling information.

8. An electronic device, characterized in that: include: a first processor; and a memory for storing executable instructions of the first processor; The first processor is configured to execute the method for three-dimensional reconstruction and biomechanical analysis of sports training based on sparse multi-view videos according to any one of claims 1 to 6 by executing the executable instructions.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the second processor, the method for three-dimensional reconstruction and biomechanical analysis of sports training based on sparse multi-view videos according to any one of claims 1 to 6 is implemented.

Images