A server resource dynamic scheduling method for high concurrency access of video stream

Abstract

The application discloses a kind of server resource dynamic scheduling methods to cope with high concurrent access of video stream, specifically related to computer network and intelligent scheduling technical field;Through the construction user access behavior dataset;Based on time series characteristics, access hotspot prediction model is trained using deep learning model, for predicting future access hotspot area and peak trend;Collect the response delay of heterogeneous server group, CPU / GPU occupancy and bandwidth load information, generate resource state multidimensional parameter set;Model and parameter set are jointly modeled, resource scheduling priority model is constructed using graph neural network, and optimal scheduling path graph is generated based on A* improvement algorithm;According to model, task transfer, instance elastic expansion and cache warm-up and other scheduling operations are executed;The method can improve the resource utilization efficiency and service quality of video system in high concurrency scenario, with the advantages of strong real-time, intelligent scheduling and strong self-learning ability.

Description

Technical Field

[0001] This invention relates to the field of computer network and intelligent scheduling technology, specifically to a method for dynamic scheduling of server resources to cope with high-concurrency access to video streams. Background Technology

[0002] With the widespread deployment of video platforms, live streaming services, and online conferencing systems, server systems face significant challenges in maintaining stability and responsiveness when dealing with high concurrency access to video streams. Especially during trending events or emergencies, a large number of users accessing the same video resource within a short period can easily lead to an imbalance in server resource allocation, resulting in issues such as response delays, video stuttering, and service interruptions, severely impacting user experience.

[0003] Existing video server scheduling systems mostly employ static or preset load balancing strategies, which cannot detect and respond to dynamic changes in peak access areas in real time, and lack the ability to elastically adapt to sudden concurrent requests. In addition, their ability to model the distribution patterns of video streams and user behavior characteristics is limited, making it difficult to predict areas of concentrated access pressure from the source and optimize resource scheduling in advance.

[0004] Therefore, there is an urgent need to propose a dynamic scheduling method for server resources in high-concurrency video streaming scenarios. This method should be able to establish a predictive model based on real-time access data, user behavior patterns, and resource usage status, dynamically identify access hotspots and weak nodes, and thus achieve refined and predictive allocation of server resources to improve system response efficiency and stability. Summary of the Invention

[0005] The purpose of this invention is to provide a method for dynamic scheduling of server resources to cope with high-concurrency access to video streams, thereby addressing the shortcomings in the prior art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for dynamic scheduling of server resources to cope with high-concurrency access to video streams, comprising:

[0007] S100. Obtain video access behavior data, video type information, and server resource usage status within a fixed time period, and construct a user access behavior dataset D.

[0008] S200. Based on the time-series characteristics of user access behavior, a deep learning model is used to train the dataset D to establish an access hotspot prediction model S, which is used to predict the access hotspot areas and access peak trends of various video resources in the future time period.

[0009] S300: Obtain the current service response latency, CPU / GPU utilization, and network bandwidth load of the heterogeneous server group to form a multi-dimensional parameter set R of server resource status;

[0010] S400. Perform joint feature analysis on the access hotspot prediction model S and the resource status parameter set R to construct a resource scheduling priority model W, which is used to dynamically allocate resource weights and form a scheduling path graph.

[0011] S500. Execute a dynamic scheduling strategy for server resources according to the model W.

[0012] S600. Collect service response latency changes and user experience scores based on resource scheduling execution results to form a multidimensional feedback dataset F;

[0013] S700. Based on the feedback dataset F, adjust the access hotspot prediction model S and the scheduling priority model W in real time.

[0014] Preferably, S200 includes:

[0015] S201. Extract time series features of access frequency, number of active users and access bounce rate at each time granularity in the user access behavior dataset, and construct a unified sequence feature matrix.

[0016] S202. The sequence feature matrix is ​​trained using a bidirectional long short-term memory network with an attention mechanism to enhance the model's ability to identify historical access fluctuation trends and future high-concurrency nodes.

[0017] S203. Identify key time windows and influencing factors through the attention weight distribution output by the model, and generate a distribution map of video resource access hotspots and predicted peak access times based on the prediction results.

[0018] Preferably, S300 includes:

[0019] S301. Classify the configuration features of each node in the server group and divide them into CPU-type, GPU-type and hybrid heterogeneous server clusters according to the computing architecture.

[0020] S302: Utilize a lightweight edge monitoring agent to collect service response latency, CPU / GPU utilization, and network bandwidth load data of each node in real time, and send them to the central scheduling controller at fixed intervals.

[0021] S303. Through normalization and weighted aggregation, a unified server resource state vector is constructed by integrating service response latency, CPU / GPU utilization, and network bandwidth load data, and a multi-dimensional parameter set R is generated by combining node type labels.

[0022] Preferably, S400 includes:

[0023] S401. Jointly encode the hotspot distribution map output by the access hotspot prediction model S with the resource status parameter set R to construct a scheduling input tensor that integrates access intensity and server load status.

[0024] S402. Introduce a scheduling modeling method based on graph neural networks, with server nodes as graph vertices and inter-server links as edges, train a resource scheduling priority model W, and output the resource allocation weights of each node.

[0025] S403. Based on the output of model W, calculate the scheduling path cost between different servers, and dynamically construct the scheduling path graph between servers using an improved minimum cost path algorithm.

[0026] Preferably, the improved minimum cost path algorithm includes:

[0027] S431. Starting from the server node associated with the access hotspot area, initialize the path cost and priority indicators of all server nodes.

[0028] S432. In the path search process, a dynamic bandwidth prediction function and a node load penalty factor are introduced to improve the heuristic function in the traditional A* algorithm;

[0029] S433. During each round of path expansion, the path cost of candidate nodes is dynamically updated, and the path with the lowest overall cost is selected for search and recommendation.

[0030] S434. Finally, output one or more low-cost scheduling paths from the access pressure center node to the target resource node, which are used to guide task transfer or traffic redirection between servers.

[0031] Preferably, S600 includes:

[0032] S601. Record the average response latency of each server node before and after each resource scheduling. Set the latency change value ΔL as the difference between the average latency within the time window after scheduling and the latency before scheduling. The calculation expression is: ;in, This represents the average response time of requests collected within 5 minutes after scheduling. ΔL represents the average response time within the same window before scheduling, and is used to dynamically reflect the trend of the impact of the scheduling strategy on the node response performance.

[0033] Preferably, S600 further includes:

[0034] S602. Extract the initial video playback delay based on the user request processing log. lag rate and average bitrate The expression for calculating the user experience score U using three indicators is as follows: Where α, β, and γ are normalized weight parameters that satisfy α + β + γ = 1, and are set according to business requirements.

[0035] Preferably, S700 includes:

[0036] S701. Perform distribution analysis on the service response latency changes and user experience scores collected in the feedback dataset F, and construct a scheduling effect evaluation label set to identify the quality level of the model output results.

[0037] S702. The hotspot prediction model S is fine-tuned using an incremental training strategy, and only the time series samples corresponding to high error regions are locally retrained.

[0038] S703. Based on the deviation between the historical scheduling path and actual resource utilization of the scheduling priority model W, a reinforcement learning method is used to adjust the node weight update strategy in the model and optimize the resource allocation decision logic.

[0039] S704. Periodically evaluate the performance changes of the adjusted model on the validation set and determine whether to push the updated model into the production environment.

[0040] The technical effects and advantages provided by the present invention in the above technical solution are as follows:

[0041] 1. This invention constructs a multi-stage video server resource scheduling process centered on access behavior modeling, hotspot prediction, resource status awareness, graph neural network scheduling, dynamic path construction, and containerized resource elastic management. This enables accurate prediction, intelligent allocation, and rapid response of server resources in high-concurrency video access scenarios. Compared to traditional static load balancing or rule-based scheduling methods, this invention can dynamically generate scheduling paths and task migration schemes based on real-time changes in user access trends and server operating status, significantly reducing system latency and improving access success rate and video playback experience.

[0042] 2. This invention introduces a feedback-driven self-learning mechanism. By collecting service response performance and user experience scores, it continuously optimizes the hotspot prediction model and scheduling strategy, achieving closed-loop adaptive operation and iterative model updates within the scheduling system. This method possesses excellent generalization ability and deployment flexibility, and can be widely applied to various high-concurrency video service scenarios such as live streaming platforms, short video applications, online education, and large-scale conference systems, demonstrating significant engineering application value. Attached Figure Description

[0043] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0044] Figure 1 This is a mind map of the method of the present invention. Detailed Implementation

[0045] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0046] For examples, please refer to Figure 1 As shown in this embodiment, a method for dynamically scheduling server resources to cope with high-concurrency access to video streams includes:

[0047] S100. Obtain video access behavior data, video type information, and server resource usage status within a fixed time period, and construct a user access behavior dataset D.

[0048] S200. Based on the time-series characteristics of user access behavior, a deep learning model is used to train the dataset D to establish an access hotspot prediction model S, which is used to predict the access hotspot areas and access peak trends of various video resources in the future time period.

[0049] S300: Obtain the current service response latency, CPU / GPU utilization, and network bandwidth load of the heterogeneous server group to form a multi-dimensional parameter set R of server resource status;

[0050] S400. Perform joint feature analysis on the access hotspot prediction model S and the resource status parameter set R to construct a resource scheduling priority model W, which is used to dynamically allocate resource weights and form a scheduling path graph.

[0051] S500. Execute a dynamic scheduling strategy for server resources according to the model W.

[0052] S600. Collect service response latency changes and user experience scores based on resource scheduling execution results to form a multidimensional feedback dataset F;

[0053] S700. Based on the feedback dataset F, adjust the access hotspot prediction model S and the scheduling priority model W in real time.

[0054] In this invention, the specific implementation of step S100 includes the following:

[0055] First, the server system collects raw access behavior data from edge node and main server logs within a specified monitoring period (e.g., 7 consecutive days), including user ID, access timestamp, requested video ID, access duration, video type (on-demand or live), and other information.

[0056] S101: To extract multi-timescale features from access behavior, the system divides the raw data into time windows according to different granularities (such as days, hours, and minutes). For example, access records are grouped using 24-hour, 60-minute, and 1-minute sliding window units, respectively, generating three subsets of access behavior: a daily granularity subset, an hourly granularity subset, and a minute granularity subset. This multi-scale structure can reveal periodic trends in user access patterns (such as concentrated access every evening) and sudden event characteristics (such as a surge in concurrency at the start of a live stream).

[0057] S102: For each time-granularity subset of visits, the system uses unsupervised clustering algorithms such as K-Means++ or DBSCAN to cluster the visit behavior data. Clustering features may include visit frequency, average duration, and visit bounce rate. The clustering results can automatically identify highly active user groups, typical access paths, and potential hot video resources, providing structured access patterns for subsequent hotspot prediction models.

[0058] S103: The system further introduces video type (such as video-on-demand and live video) as a label feature, and performs hierarchical modeling of access behavior based on content type. Live videos emphasize concurrency peaks and latency sensitivity, while video-on-demand videos focus more on access time distribution and cache hit rate. Finally, the system integrates multi-granularity access features, access pattern clustering results, and video type information to construct a three-dimensional structured user access behavior dataset D, which includes time dimension, behavioral feature dimension, and content type dimension.

[0059] This dataset D serves as an important input for training subsequent access hotspot prediction models, significantly improving the accuracy and timeliness of the prediction models in identifying access hotspot areas and traffic trends, outperforming existing static methods that rely solely on access counts or historical average loads.

[0060] In this invention, the specific implementation of step S200 includes the following:

[0061] S201: The system first preprocesses the user access behavior dataset D constructed in step S100, extracting key time-series features at multiple time scales, including the number of video access requests per unit time (access frequency), the number of deduplicated active users, the bounce rate, and the average viewing time. Each feature is arranged into a multi-dimensional time series according to time granularity (day, hour, minute), and uniformly constructed into a sequence feature matrix. Each row of the matrix represents multiple access indicators for a specific time window, and each column represents the time series of a specific feature.

[0062] S202: To achieve the learning and prediction of complex time-series behaviors, the system employs a bidirectional long short-term memory network with an attention mechanism for modeling and training. The model structure includes an input layer, a bidirectional LSTM encoding layer, an attention weight calculation module, and an output layer. The bidirectional LSTM can simultaneously capture past and future temporal dependencies, while the attention mechanism assigns higher weights to key time windows or feature dimensions, thereby improving the accuracy of predicting hotspot peaks. Model training uses video access behavior data from the past several consecutive days, and the loss function is a weighted combination of mean squared error and access peak prediction bias.

[0063] S203: After model training is complete, the system uses its output attention weights to perform visual analysis of the prediction process, identifying time windows and behavioral characteristics that play a key role in predicting access peaks. For example, attention distribution can be used to identify potential concurrent nodes for live video access spikes 30 minutes before the event starts. Finally, the video resource access hotspot distribution map (categorized by video ID) and the predicted peak time periods output by the model will be used as input references for subsequent scheduling models (step S400) to prepare server resources or optimize load strategies in advance.

[0064] This method has better timeliness and prediction accuracy than traditional moving average prediction or static load threshold identification methods, and its scheduling response capability is significantly enhanced, especially in the case of sudden access scenarios.

[0065] In this invention, the specific implementation of step S300 includes the following:

[0066] S301: The server system first categorizes the computing resources deployed in data centers and edge nodes across various regions. Based on the hardware configuration characteristics of the nodes (such as whether they are equipped with GPUs, the number of CPU cores, and memory capacity), the server groups are divided into three categories: CPU-based servers (suitable for general request processing), GPU-based servers (suitable for high-concurrency image decoding and AI computing), and CPU+GPU hybrid servers (suitable for workloads requiring image preprocessing and deep model inference). This classification result will serve as a resource status label in subsequent scheduling decisions.

[0067] S302: Deploy a lightweight edge monitoring agent component on each server node to periodically collect key performance indicators, including: service response latency (average response time per request), CPU utilization, GPU utilization, and current inbound / outbound bandwidth load. The monitoring frequency is typically set to every 30 seconds to 1 minute, and the collected data is sent to the central scheduling controller through the internal network channel to ensure data synchronization and low latency.

[0068] S303: After receiving performance data from each server node, the central scheduling controller first normalizes different metrics to eliminate differences in numerical scale. Then, based on the type of each server node (e.g., GPU-based servers prioritize AI tasks), different metrics are assigned weighted values; for example, GPU utilization has a higher weight in GPU-based servers. Finally, the system generates a server resource status vector in a unified format. Each vector contains metrics such as response latency, CPU utilization, GPU utilization, and network bandwidth load, along with a server type label, forming a structured multidimensional parameter set R.

[0069] This parameter set R can fully describe the current operating status and scheduling capacity of each server, and serves as the input basis for subsequent resource scheduling strategy calculations. Through this method, the system can achieve real-time perception and accurate modeling of the resource status of heterogeneous servers, significantly improving the accuracy and efficiency of resource allocation.

[0070] In this invention, the specific implementation of step S400 includes the following:

[0071] S401: The system first fuses the output of the access hotspot prediction model S in step S200—that is, the predicted video access hotspot region and traffic peak map—with the server resource status multidimensional parameter set R generated in step S300. Specifically, the system maps each predicted hotspot region to its potentially associated server nodes and generates a set of joint features for each server node, including: the access intensity of associated videos (such as the expected number of concurrent connections), and the current load status of the node (such as CPU / GPU utilization, response latency, etc.). These joint features are encoded into a scheduling input tensor, which serves as the input basis for the subsequent scheduling model.

[0072] S402: Based on the above input tensor, the system employs a Graph Neural Network (GNN) for resource scheduling modeling. This model uses each server node in the server cluster as a vertex of the graph; edges are constructed where servers are connected by network connections, and edge weights can be set based on physical distance, network latency, or historical load balancing efficiency. The GNN model aggregates the state information of neighboring nodes through a graph structure propagation mechanism and combines it with local features to learn the scheduling priority weight for each server node. This priority value represents the server's scheduling suitability and distribution capability under current access pressure. Historical scheduling efficiency and system response metrics are used as supervision signals during the training process.

[0073] S403: Based on the scheduling priorities of each server node output by the GNN model, the system constructs a scheduling path graph for the entire server graph. Using an improved minimum cost path algorithm (such as Dijkstra's or an A* variant), starting from the access pressure center, it finds the optimal path to distribute tasks to other servers. Path costs can comprehensively consider scheduling latency, link bandwidth, and node load. The final generated scheduling path graph not only indicates how tasks should migrate or replicate between multiple nodes but also provides routing references for subsequent resource allocation and caching strategies.

[0074] This method breaks through the limitations of traditional static scheduling or single-node priority ranking methods. It utilizes graph neural networks and graph path optimization algorithms to achieve deep coupling and automated reasoning between scheduling priority modeling and dynamic generation of scheduling paths, which greatly improves scheduling intelligence and system response flexibility.

[0075] In this invention, after generating the scheduling priority model W in step S400, an improved minimum cost path algorithm (e.g., based on a variant of the A* algorithm) is further used to dynamically generate the scheduling path graph between servers, as specifically implemented below:

[0076] S431: The system first uses the core server nodes associated with the hotspot areas output by the access hotspot prediction model S as the path starting points, i.e., the access pressure centers. The scheduling target is set as a cluster of server nodes with available resources. The total path cost f(n) of all nodes is initialized to infinity, and the starting node's f(start) is initialized to 0. Simultaneously, a path heuristic evaluation value h(n) is defined for each node, initially estimated as the weighted prediction of the network latency from that node to the target server.

[0077] S432: The system introduces two improvements to the A* algorithm to enhance the adaptability of scheduling paths:

[0078] Dynamic bandwidth prediction function: This function uses historical bandwidth utilization sequences and a sliding window averaging method to predict the future available bandwidth of inter-server links, reflecting path transmission stability. This value is reflected in the edge weights and used to adjust path selection preferences.

[0079] Node load penalty factor: For server nodes with high current CPU / GPU load or response latency approaching the threshold, a penalty coefficient P(n) is introduced and used as a weighting term in the path cost of g(n), i.e., g(n) = cumulative path transmission cost + P(n). P(n) is dynamically determined by the node load to ensure that the scheduling path avoids overloaded nodes.

[0080] The heuristic function f(n) = g(n) + h(n) is thus updated to: f(n) = current cost + node load penalty + dynamic latency prediction + inverse bandwidth prediction.

[0081] S433: During path expansion, the system uses a priority queue to maintain a set of nodes to be expanded. Each time, the node with the smallest f(n) value is selected for search and recommendation. In each round of expansion, g(n) and h(n) are recalculated based on the current node's adjacent edges (i.e., reachable servers), and the path record is updated based on whether it is better than the existing cost value. This strategy supports real-time response to node state changes, ensuring that the path search process always tends towards expanding towards low-load, highly available nodes.

[0082] S434: The search termination condition is that the path successfully reaches any node in the target server set, or the set maximum search depth limit is reached. The system outputs one or more scheduling paths starting from the access hotspot area with the lowest overall cost, which can be used for subsequent video task migration, traffic redirection, or content preloading. For multiple feasible paths, the system can further select paths based on the scheduling purpose (such as latency-sensitive or throughput-optimized).

[0083] Compared to traditional static routing or simple shortest path methods, this path generation method has dynamic adaptation, load avoidance, and predictive capabilities, which can significantly improve the real-time performance and stability of video server cluster resource scheduling in high-concurrency access scenarios.

[0084] In this invention, the specific implementation of step S500 includes the following:

[0085] S501: After calculating the resource scheduling priority model W and generating the scheduling path graph, the system automatically generates a scheduling instruction set for task distribution and migration based on the resource weight allocation results of each server node and the optimal scheduling path. This instruction set includes: target nodes for hot video task distribution, cache deployment locations, service instance expansion strategies, etc., and is sent to the scheduling control center or Kubernetes controller in YAML or JSON format for parsing and execution.

[0086] S502: For accessing frequently accessed video resources, the system adopts a content segmentation and caching strategy. First, the frequently accessed video is segmented by time slices or bitrate levels (e.g., 10-second segments). Core segments are prioritized and deployed to high-priority nodes closer to the user, and pre-loading is completed before the user's first request, ensuring low-latency playback on the first access. Simultaneously, combined with a user access trajectory prediction model, the system anticipates potential user paths and pre-deploys potentially accessed video segments to nodes along those paths.

[0087] S503: To improve the ability to handle sudden concurrency spikes, the system dynamically adjusts the number of service instances based on a containerized deployment platform (such as Kubernetes or Docker Swarm). For server nodes that are about to be overloaded, replica service instances can be quickly started on adjacent low-load nodes, and traffic can be automatically routed to the new nodes through the service mesh. In addition, task instances can be migrated across nodes in seconds using a container image migration mechanism. During the process, the system selects the target node with the least response latency based on the scheduling path graph, achieving intelligent resource avoidance.

[0088] S504: During task execution, the system continuously monitors the service response status and instance processing efficiency of each node. If a significant increase in response latency or backlog in the processing queue is detected, the node status is immediately fed back to the central scheduling system, triggering a new round of model W retraining and path graph updates. This mechanism forms a real-time feedback loop, supporting continuous optimization and rapid convergence of resource scheduling strategies under high-concurrency access scenarios, and significantly reducing the risk of service bottlenecks caused by static scheduling lag.

[0089] This embodiment achieves efficient dynamic scheduling of server resources through four core methods: resource scheduling weight guidance, hotspot preheating caching, elastic container expansion, and feedback closed-loop scheduling. It has significant stability and response advantages in complex application scenarios such as sudden access peaks and concentrated requests for hot videos.

[0090] In this invention, step S600 is used to construct a scheduling feedback closed loop, and its specific implementation is as follows:

[0091] This step aims to establish a multidimensional feedback dataset F by collecting and analyzing the operational effects of resource scheduling strategies after their execution, providing a basis for subsequent model updates and scheduling optimization. This dataset covers two core feedback dimensions: changes in service response performance and changes in user experience, reflecting the actual impact of scheduling strategies at the system operation level and the end-user level, respectively.

[0092] After the server resources are scheduled and executed, the system continuously samples the service response time of each server node. The main monitoring indicator is the average response latency, which is the time taken from when a user request is sent to when the server returns the first byte of data per unit time.

[0093] The specific implementation is as follows:

[0094] Data acquisition period definition: A performance evaluation window is set 5 minutes before and after the scheduling action, and the average response time of each node within the window is recorded. Calculation expression: ;in: This indicates the average response latency within the first 5 minutes of scheduling execution; ΔL represents the average response latency within 5 minutes after the scheduling is executed; ΔL represents the performance change brought about by the scheduling. If ΔL < 0, it means that the response performance has been improved.

[0095] Server nodes periodically record processing logs and timestamp data via edge probes or local agents, which are then aggregated and the difference calculated by the central controller. The latency change value will serve as an important indicator for evaluating scheduling effectiveness and will be used in the next round of scheduling weight adjustments.

[0096] To measure the effectiveness of resource scheduling strategies from the end-user's perspective, this system introduces a User Experience Score (U), which integrates multiple user perception metrics for unified quantification.

[0097] Specifically, the following indicators are included:

[0098] Initial loading time for video playback, i.e., the delay from when the user clicks play to when the video starts playing;

[0099] : Stuttering rate, which is the ratio of the total duration of video playback interruptions to the total duration of the video;

[0100] Average playback bitrate, reflecting the video quality level.

[0101] The formula for calculating user experience scores is: Where: α, β, γ are empirically set weighting coefficients used to adjust the influence of each indicator on the overall score, generally set to α=0.3, β=0.4, γ=0.3;

[0102] All input metrics are normalized in the range [0,1] to ensure score comparability;

[0103] The higher the score U, the better the user's viewing experience during that time period.

[0104] The system continuously collects these experience parameters while users watch videos, combines them with metadata such as video type (on-demand / live), access device type, etc., to form a user experience score log, and aggregates them to form a subset of experience data.

[0105] After collecting data from the above two dimensions, the system stores the feedback results of each scheduling cycle in a structured format as a multidimensional feedback dataset F, with each record containing the following fields:

[0106] Scheduling timestamp; node ID and resource status snapshot; service response latency changes; user experience score; scheduling path and target node information;

[0107] The feedback dataset F will serve as input for the next step of model training and parameter correction, providing historical scheduling performance data to support the access hotspot prediction model S and the resource priority model W, thus forming a self-learning and self-optimization closed-loop mechanism for the scheduling system.

[0108] In this invention, step S700 is used to optimize and update the access hotspot prediction model S and the scheduling priority model W in real time after collecting feedback data, thereby constructing an adaptive and evolvable intelligent scheduling system.

[0109] This step involves building a feedback-driven model update mechanism, enabling the system to continuously learn and correct its judgments during operation, thereby responding to dynamic factors such as changes in access behavior, fluctuations in server status, and updates to trending content.

[0110] S701: Construct a label set for scheduling effect evaluation; The system first performs structured processing on the feedback dataset F formed in step S600, and extracts the service response latency change value (ΔL) and user experience score (U) as the main evaluation indicators.

[0111] Discretized label generation: After normalizing ΔL and U, they are divided into three levels: efficient (positive improvement), neutral (no significant change), and inefficient (deterioration) by using a threshold.

[0112] Example: If ΔL < -100ms and U increases by more than 10%, mark it as "Scheduling Excellent +"; if ΔL > +150ms or U decreases by more than 15%, mark it as "Scheduling Inferior -".

[0113] Label set structure: Each label record is bound to the comparison result of the model output and the actual feedback within a scheduling cycle, forming the "scoring label" of the training sample.

[0114] This step provides "supervision signals" in real-world scenarios for subsequent model training, guiding the model to learn corrective measures for scheduling deviations.

[0115] S702: Incremental fine-tuning of the hotspot prediction model S; the hotspot prediction model S employs a time-series deep neural network based on a Bi-LSTM with Attention architecture. To avoid the resource consumption caused by frequent full training, the system adopts an incremental learning strategy:

[0116] Error localization: High error samples are screened out by using the prediction deviation before and after (such as the difference between the predicted hot spots and the actual access distribution); combined with the label set, only video resources and time periods with significant error rates are selected as "high-value training samples".

[0117] Local retraining: High-error samples are constructed into short-cycle time series micro-sets; the original model structure is preserved, and only the weights of the later layers are fine-tuned to update the prediction preferences without destroying the global stability of the model; the training cycle is greatly shortened and can be quickly updated within the scheduling cycle (e.g., every 15 minutes).

[0118] The technological advantages are: enabling rapid iteration of the prediction model, improving the accuracy and real-time performance of hotspot identification, without affecting other stable prediction paths.

[0119] S703: Policy optimization (reinforcement learning) for scheduling priority model W; the scheduling priority model W uses a resource scheduling structure based on a graph neural network (GNN) to output the resource scheduling weights of server nodes. The system further introduces a reinforcement learning (RL) mechanism to optimize the scheduling policy:

[0120] State-action modeling:

[0121] Status: Consists of the current scheduling graph structure, node load status, predicted hotspot intensity, etc.

[0122] Action: Represents the decision on scheduling path selection and resource migration amount.

[0123] Reward function design: The reward function combines ΔL and U from the feedback dataset;

[0124] If the scheduling results improve (ΔL decreases, U increases), a positive reward is given; otherwise, a penalty is imposed.

[0125] Strategy Update: The scheduling path selection logic is optimized using strategies such as Q-learning or Actor-Critic; the model dynamically adjusts the weight update rate and priority determination rules of each node to avoid path congestion or resource misallocation. This step enables the scheduling model to have adaptive learning capabilities, improving its resource allocation efficiency under different system states.

[0126] S704: Model evaluation and deployment switching mechanism; To ensure the stable operation of the updated model and avoid system fluctuations, the system introduces a canary evaluation mechanism:

[0127] Model comparison test: The new model and the previous version were run simultaneously on the offline validation set and some production traffic; the comparison indicators included hotspot prediction accuracy, average path latency, changes in user experience score, etc.

[0128] Switching threshold setting: Set switching conditions, such as prediction accuracy improvement >5% and response latency reduction >10%, then the old model will be automatically replaced; if the new model does not meet the set improvement criteria, the current version will be retained and the switch will be delayed.

[0129] Deployment strategy: Use rolling updates and blue-green deployment mechanisms to minimize update risks and ensure system continuity.

[0130] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for dynamically scheduling server resources to cope with high-concurrency access to video streams, characterized in that: include: S100. Obtain video access behavior data, video type information, and server resource usage status within a fixed time period, and construct a user access behavior dataset D. S200. Based on the time-series characteristics of user access behavior, a deep learning model is used to train the dataset D to establish an access hotspot prediction model S, which is used to predict the access hotspot areas and access peak trends of various video resources in the future time period. S300: Obtain the current service response latency, CPU / GPU utilization, and network bandwidth load of the heterogeneous server group to form a multi-dimensional parameter set R of server resource status; S400. Perform joint feature analysis on the access hotspot prediction model S and the server resource status multidimensional parameter set R to construct a resource scheduling priority model W, which is used to dynamically allocate resource weights and form a scheduling path graph. The S400 includes: S401. Jointly encode the hotspot distribution map output by the access hotspot prediction model S with the multi-dimensional parameter set R of server resource status to construct a scheduling input tensor that integrates access intensity and server load status. S402. Introduce a scheduling modeling method based on graph neural networks, with server nodes as graph vertices and inter-server links as edges, train a resource scheduling priority model W, and output the resource allocation weights of each node. S403. Based on the output of model W, calculate the scheduling path cost between different servers, and dynamically construct the scheduling path graph between servers using the improved minimum cost path algorithm. S500. Execute a dynamic scheduling strategy for server resources according to the model W. S600. Collect service response latency changes and user experience scores based on resource scheduling execution results to form a multidimensional feedback dataset F; S700. Based on the multidimensional feedback dataset F, adjust the access hotspot prediction model S and the scheduling priority model W in real time.

2. The method for dynamic scheduling of server resources to cope with high concurrency access to video streams according to claim 1, characterized in that: S200 includes: S201. Extract time series features of access frequency, number of active users and access bounce rate at each time granularity in the user access behavior dataset, and construct a unified sequence feature matrix. S202. The sequence feature matrix is ​​trained using a bidirectional long short-term memory network with an attention mechanism to enhance the model's ability to identify historical access fluctuation trends and future high-concurrency nodes. S203. Identify key time windows and influencing factors through the attention weight distribution output by the model, and generate a distribution map of video resource access hotspots and predicted peak access times based on the prediction results.

3. The method for dynamic scheduling of server resources to cope with high concurrency access to video streams according to claim 1, wherein step S300 includes: S301. Classify the configuration features of each node in the server group and divide them into CPU-type, GPU-type and hybrid heterogeneous server clusters according to the computing architecture. S302: Utilize a lightweight edge monitoring agent to collect service response latency, CPU / GPU utilization, and network bandwidth load data of each node in real time, and send them to the central scheduling controller at fixed intervals. S303. Through normalization and weighted aggregation, a unified server resource state vector is constructed by integrating service response latency, CPU / GPU utilization and network bandwidth load data, and a multi-dimensional parameter set R of server resource state is generated by combining node type labels.

4. The method for dynamic scheduling of server resources to cope with high concurrency access to video streams according to claim 1, wherein the improved minimum cost path algorithm is adopted, comprising: S431. Starting from the server node associated with the access hotspot area, initialize the path cost and priority indicators of all server nodes. S432. In the path search process, a dynamic bandwidth prediction function and a node load penalty factor are introduced to improve the heuristic function in the traditional A* algorithm; S433. During each round of path expansion, the path cost of candidate nodes is dynamically updated, and the path with the lowest overall cost is selected for search and recommendation. S434. Finally, output one or more low-cost scheduling paths from the access pressure center node to the target resource node, which are used to guide task transfer or traffic redirection between servers.

5. The method for dynamic scheduling of server resources to cope with high concurrency access to video streams according to claim 1, wherein step S600 includes: S601. Record the average response latency of each server node before and after each resource scheduling. Set the latency change value ΔL as the difference between the average latency within the time window after scheduling and the latency before scheduling. The calculation expression is: ;in, This represents the average response time of requests collected within 5 minutes after scheduling. ΔL represents the average response time within the same window before scheduling, and is used to dynamically reflect the trend of the impact of the scheduling strategy on the node response performance.

6. The method for dynamic scheduling of server resources to cope with high concurrency access to video streams according to claim 5, wherein step S600 further includes: S602. Extract the initial video playback delay based on the user request processing log. lag rate and average bitrate The expression for calculating the user experience score U using three indicators is as follows: Where α, β, and γ are normalized weight parameters that satisfy α + β + γ = 1, and are set according to business requirements.

7. The method for dynamic scheduling of server resources to cope with high concurrency access to video streams according to claim 1, wherein step S700 includes: S701. Perform distribution analysis on the service response latency changes and user experience scores collected in the multidimensional feedback dataset F, and construct a scheduling effect evaluation label set to identify the quality level of the model output results. S702. The hotspot prediction model S is fine-tuned using an incremental training strategy, and only the time series samples corresponding to high error regions are locally retrained. S703. Based on the deviation between the historical scheduling path and actual resource utilization of the scheduling priority model W, a reinforcement learning method is used to adjust the node weight update strategy in the model and optimize the resource allocation decision logic. S704. Periodically evaluate the performance changes of the adjusted model on the validation set and determine whether to push the updated model into the production environment.

Images