Intelligent bandwidth scheduling method and system for computing power cloud platform

Abstract

The invention relates to the technical field of bandwidth scheduling, and particularly discloses an intelligent bandwidth scheduling method and system for a computing power cloud platform, and the method comprises the steps: distributing a task to a computing node through a task scheduler after receiving a computing task request, and storing task metadata; and the network performance monitoring module is synchronously started to monitor the network flow of the task. And further, based on the stored task metadata and the collected task network flow data, performing deep conjoint analysis on the task metadata and the collected task network flow data by using a deep learning algorithm to identify whether the task is in a network starting stage or not. If it is determined that the task is in the network starting stage, an application enhancement strategy request is sent to an intelligent bandwidth scheduler, and a high-priority QoS label is assigned to the task for bandwidth adjustment. Through the mode, the network demand characteristics of the task in different stages of the life cycle can be captured in a fine-grained manner, and the bandwidth resource allocation is dynamically adjusted, so that the bandwidth demand in the task starting stage is ensured, and the task execution is accelerated.

Description

Technical Field

[0001] The present application relates to the technical field of bandwidth scheduling, and more specifically, to a bandwidth intelligent scheduling method and system for a computing power cloud platform. Background Art

[0002] With the rapid development of cloud computing and computing cloud platforms, the number and complexity of computing tasks are constantly increasing. The rational allocation of network bandwidth resources has become critical to ensuring system performance and user experience. In computing cloud platforms, data transmission between computing nodes and between tasks and external systems is highly dependent on bandwidth. Without intelligent bandwidth scheduling, bandwidth resources may be wasted or bottlenecked, leading to task execution delays, reduced system efficiency, and even unstable service quality. Therefore, building an efficient and intelligent bandwidth scheduling solution is urgently needed to improve the overall performance of computing cloud platforms.

[0003] Currently, bandwidth scheduling methods in computing cloud platforms mostly use static allocation or simple priority-based strategies. Static allocation strategies typically set a fixed bandwidth cap for each task or virtual machine. While simple and easy to implement, they struggle to adapt to the dynamic changes in the actual network demands of tasks, potentially leaving some tasks with idle bandwidth resources while others are limited by insufficient bandwidth. While simple priority-based strategies can differentiate tasks of varying importance, they typically do so globally and fail to identify the network demand characteristics of a single task at different stages of its lifecycle. For example, many tasks require much more bandwidth during the startup phase (such as pulling images or loading initial datasets) than during their subsequent stable operation phase. Consequently, bandwidth priority allocation strategies based on global task cycles can result in insufficient bandwidth at task startup, impacting startup speed, or waste bandwidth during stable operation, reducing resource utilization. This limits the adaptability and intelligence of bandwidth scheduling strategies.

[0004] Therefore, an optimized bandwidth intelligent scheduling method and system for computing power cloud platforms is expected. Summary of the Invention

[0005] In order to solve the above technical problems, the present application is proposed. The embodiment of the present application provides a bandwidth intelligent scheduling method and system for a computing power cloud platform, which, after receiving a computing task request, assigns the task to the computing node through the task scheduler, stores the task metadata, and synchronously starts the network performance monitoring module to monitor the network traffic of the task. Furthermore, based on the stored task metadata and the collected task network traffic data, a deep learning algorithm is used to perform a deep joint analysis of the two to identify whether the task is in the network startup stage. If it is determined that the task is in the network startup stage, an application enhancement policy request is sent to the intelligent bandwidth scheduler to assign a high-priority QoS label to the task for bandwidth adjustment. In this way, the network demand characteristics of the task at different stages of the life cycle can be captured in a fine-grained manner, and the bandwidth resource allocation can be dynamically adjusted to ensure the bandwidth demand in the task startup stage and improve the utilization of bandwidth resources and the efficiency of task execution.

[0006] Accordingly, according to one aspect of the present application, a bandwidth intelligent scheduling method for a computing cloud platform is provided, which includes: After receiving the task request of the first computing task, the task scheduler assigns the first task to a computing node; The task scheduler sends the metadata of the first task to the task metadata storage module for storage, and notifies the network performance monitoring module to monitor the network traffic of the first task; Acquire task network traffic data of the first task collected by the network performance monitoring module and extract metadata of the first task from the task metadata storage module; determining, based on the task network traffic data of the first task and the metadata of the first task, whether the first task is in a network startup phase; In response to the first task being in the network startup phase, sending a request to the intelligent bandwidth scheduler to apply the enhancement policy; After receiving the request to apply the enhanced policy, the intelligent bandwidth scheduler assigns a high priority QoS label to the first computing task to generate a bandwidth adjustment policy instruction.

[0007] According to another aspect of the present application, a bandwidth intelligent scheduling system for a computing cloud platform is provided, comprising: A computing node allocation module, configured to allocate the first computing task to a computing node after receiving a task request of the first computing task; a network traffic monitoring module, configured to use a task scheduler to send metadata of the first task to a task metadata storage module for storage, and to notify a network performance monitoring module to monitor network traffic of the first task; a data extraction module, configured to obtain the task network traffic data of the first task collected by the network performance monitoring module and extract metadata of the first task from the task metadata storage module; a network stage identification module, configured to determine whether the first task is in a network startup stage based on task network traffic data of the first task and metadata of the first task; a bandwidth scheduling request module, configured to send a request for applying an enhancement policy to the intelligent bandwidth scheduler in response to the first task being in the network startup phase; The bandwidth adjustment module is configured to assign a high priority QoS label to the first computing task to generate a bandwidth adjustment policy instruction in response to the intelligent bandwidth scheduler receiving the request to apply the enhanced policy.

[0008] Compared with the prior art, the bandwidth intelligent scheduling method and system for a computing power cloud platform provided by the present application, after receiving a computing task request, assigns the task to the computing node through the task scheduler, stores the task metadata, and simultaneously starts the network performance monitoring module to monitor the network traffic of the task. Furthermore, based on the stored task metadata and the collected task network traffic data, a deep learning algorithm is used to perform a deep joint analysis of the two to identify whether the task is in the network startup stage. If it is determined that the task is in the network startup stage, an application enhancement policy request is sent to the intelligent bandwidth scheduler to assign a high-priority QoS label to the task for bandwidth adjustment. In this way, the network demand characteristics of the task at different stages of its life cycle can be captured in a fine-grained manner, and bandwidth resource allocation can be dynamically adjusted to ensure the bandwidth demand in the task startup stage and improve the utilization of bandwidth resources and task execution efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] The above and other purposes, features, and advantages of the present application will become more apparent through a more detailed description of the embodiments of the present application in conjunction with the accompanying drawings. The accompanying drawings are intended to provide a further understanding of the embodiments of the present application and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the present application and do not constitute a limitation of the present application. In the drawings, the same reference numerals generally represent the same components or steps.

[0010] Figure 1 This is a flowchart of a bandwidth intelligent scheduling method for a computing power cloud platform according to an embodiment of the present application.

[0011] Figure 2 Schematic diagram of data flow of a bandwidth intelligent scheduling method for a computing power cloud platform according to an embodiment of the present application.

[0012] Figure 3 This is a flowchart of step S4 in the bandwidth intelligent scheduling method for a computing power cloud platform according to an embodiment of the present application.

[0013] Figure 4 This is a flowchart of step S41 in the bandwidth intelligent scheduling method for a computing power cloud platform according to an embodiment of the present application.

[0014] Figure 5 This is a flowchart of step S43 in the bandwidth intelligent scheduling method for a computing power cloud platform according to an embodiment of the present application.

[0015] Figure 6 This is a block diagram of a bandwidth intelligent scheduling system for a computing cloud platform according to an embodiment of the present application. DETAILED DESCRIPTION

[0016] Below, an example embodiment according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiment is only a part of the embodiment of the present application, not all of the embodiments of the present application, and it should be understood that the present application is not limited to the example embodiment described herein. It is worth noting that in the present application, all actions to obtain data are carried out in compliance with the relevant data protection laws and policies of the country where they are located, and with the authorization of the corresponding device owner.

[0017] Figure 1 This is a flowchart of a bandwidth intelligent scheduling method for a computing power cloud platform according to an embodiment of the present application. Figure 2 Schematic diagram of data flow of the bandwidth intelligent scheduling method for computing cloud platform according to the embodiment of the present application. Figure 1 and Figure 2 As shown, according to an embodiment of the present application, a bandwidth intelligent scheduling method for a computing cloud platform includes the following steps: S1, after receiving a task request for a first computing task, the task scheduler assigns the first task to a computing node; S2, the task scheduler sends the metadata of the first task to the task metadata storage module for storage, and notifies the network performance monitoring module to monitor the network traffic of the first task; S3, obtains the task network traffic data of the first task collected by the network performance monitoring module and extracts the metadata of the first task from the task metadata storage module; S4, based on the task network traffic data of the first task and the metadata of the first task, determines whether the first task is in the network startup stage; S5, in response to the first task being in the network startup stage, sends a request to the intelligent bandwidth scheduler to apply an enhancement policy; S6, after receiving the request to apply the enhancement policy, the intelligent bandwidth scheduler assigns a high priority QoS label to the first computing task to generate a bandwidth adjustment policy instruction.

[0018] In the above-mentioned bandwidth intelligent scheduling method for the computing power cloud platform, in step S1, after receiving the task request of the first computing task, the task scheduler assigns the first task to a computing node. It should be understood that the computing power cloud platform is usually composed of a large number of computing nodes and runs many computing tasks at the same time, and the newly submitted task requires a specific physical or virtual operating environment. Therefore, in order to enable the computing task to be actually executed and to reasonably utilize the computing resources within the platform, this application is based on the resource scheduling and management principles of cloud computing, and converts the task request into an executable task instance through the task scheduler and deploys it to the appropriate computing node.

[0019] In one specific embodiment of the present application, a user can submit a computing task request through an API, command-line tool, or graphical user interface. This task request typically includes the task's image information (e.g., a Docker image), resource requirements (e.g., number of CPU cores, memory size, GPU type and quantity), startup commands, environment variables, data volume mounting requirements, and so on. Upon receiving this request, the computing cloud platform's central task scheduler (e.g., a Kubernetes-based scheduler) first performs a validity check and resource pre-check on the request, evaluating the required resources to ensure the platform has sufficient resources to fulfill the request. If the check and resource pre-check pass, the task scheduler then selects a computing node from the cluster's pool of available computing nodes that meets the task's resource requirements and complies with the scheduling policy based on a preset scheduling policy (e.g., least load first) and real-time information collected about the load status and available resources of each computing node. Once a computing node is selected, the task scheduler sends the complete description of the first task (including image, commands, configuration, etc.) and allocation instructions to the node agent (e.g., Kubelet) running on the selected computing node. After receiving the instructions, the node agent is responsible for preparing the runtime environment on the node. This includes pulling the container image required for the task (if not cached locally), configuring the network namespace, setting resource limits (cgroups), mounting the specified data volume, and finally starting the task process or container. This ensures that computing tasks are promptly and appropriately assigned to specific computing resources, laying the foundation for smooth task startup and execution. It also achieves efficient allocation and load balancing of computing resources, improving the platform's overall throughput and resource utilization.

[0020] In the above-mentioned intelligent bandwidth scheduling method for the computing cloud platform, in step S2, the task scheduler sends the metadata of the first task to the task metadata storage module for storage, and notifies the network performance monitoring module to monitor the network traffic of the first task. It should be understood that since the intelligent bandwidth scheduling decision is highly dependent on the comprehensive understanding of the inherent characteristics of the task and the real-time network behavior. Therefore, in order to achieve refined management of tasks and intelligent decision support, this application centrally stores the key metadata of the first task through the task scheduler, and simultaneously activates the monitoring mechanism of the network traffic of the task.

[0021] In one embodiment of the application, after the task scheduler assigns the first computing task to the computing node, the key metadata contained in the task request or generated during the scheduling process is first extracted. The metadata includes task type, image, mounted volume, priority, and user annotation. Specifically, the task type reveals the application scenario (such as big data analysis, deep learning training, web application service, etc.) and potential network behavior pattern to which the task belongs; the image information includes the name, tag, source repository, and image size of the image, wherein the image size is directly related to the network bandwidth and time required to pull the image when the task is started; the mounted volume information describes the persistent storage or shared file system that the task needs to access, and its type (such as distributed file system, block storage) and size can also affect the network demand in the data loading stage; the priority is set by the user or automatically assigned by the system according to the service level agreement (SLA), and is used for arbitration when resources are competed; and the user annotation is additional unstructured explanatory information provided by the user, which may contain special expectations or hints for network performance. The task scheduler encapsulates these structured or semi-structured metadata and sends them to the task metadata storage module through API calls or message passing (such as using gRPC or publishing messages to the message queue Kafka). The storage module is a highly available distributed database (such as etcd, Consul for storing key configurations, or MongoDB, PostgreSQL for storing more rich metadata), which ensures the persistence, consistency, and queryability of the metadata. At the same time, the task scheduler also sends a monitoring start notification to the network performance monitoring module. The monitoring start notification at least contains the unique identifier of the task (such as task ID, container ID, Pod IP address and port), the identifier of the computing node where the task is located, and the network interface information that needs to be monitored. After receiving the notification, the network performance monitoring module activates or configures its monitoring agent (such as eBPF-based probes, Packetbeat, or monitoring functions integrated in virtual switches such as Open vSwitch) deployed on the corresponding computing node to start collecting network traffic data of the task, such as real-time throughput (upload / download rate), packet quantity, connection number, TCP retransmission rate, etc., and periodically sends these data to the central data collection or analysis platform. In this way, not only the completeness and traceability of the task description information are ensured, providing a static context for subsequent intelligent analysis, but also the dynamic network behavior of the task can be captured in real time as soon as it starts running, providing key real-time data sources for accurately identifying the stage of the task (especially the bandwidth-sensitive startup stage).

[0022] In the above-mentioned intelligent bandwidth scheduling method for a computing cloud platform, step S3 obtains the task network traffic data of the first task collected by the network performance monitoring module and extracts the metadata of the first task from the task metadata storage module. It should be understood that accurately determining whether a task is in the network startup phase and then implementing a precise bandwidth enhancement strategy requires comprehensive consideration of the inherent attributes of the task (metadata) and its real-time operating status (network traffic data). Therefore, based on the data-driven decision-making principle, the present application actively collects the task network traffic data and metadata information of the first task from various data sources through a coordination component to support subsequent task operation phase analysis and bandwidth allocation decision-making process.

[0023] In one specific embodiment of the present application, within a predetermined time window after task initiation, or when periodically evaluating the network status of a first computing task, the analysis service engine first initiates a query request to the task metadata storage module. Using the task's unique identifier (e.g., the task ID generated and delivered by the task scheduler when the task is assigned), the query retrieves the previously stored complete metadata record for the first task. This involves performing an SQL query on a database, a document lookup in a NoSQL database, or a read operation from a key-value store. The retrieved metadata, such as task type, image size, mounted volume configuration, priority, and user annotations, is parsed and loaded into memory. The analysis service then requests network traffic data for the corresponding task ID from the network performance monitoring module. The network performance monitoring module extracts network traffic metrics collected within a specific time window (e.g., the past 1 minute or 5 minutes) from its internal data store (e.g., a time series database, Prometheus). These metrics include bandwidth utilization, number of TCP connections, packet loss rate, ingress bandwidth (bytes / sec), egress bandwidth (bytes / sec), and total bytes transferred. The data acquisition interface can be a REST API, gRPC service, or a direct database query. The acquired raw traffic data is then preprocessed, such as aligning timestamps, filling missing values ​​(if any), aggregating to a unified time granularity (e.g., one data point per second or every 5 seconds), and performing unit conversion (e.g., converting bytes to MBits) to ensure consistency in the data format and ease of subsequent processing. This approach successfully aggregates metadata describing the static properties of the task and traffic data reflecting the dynamic network behavior of the task, providing a complete and high-quality data foundation for the next stage of data analysis and processing.

[0024] In the above-mentioned intelligent bandwidth scheduling method for a computing cloud platform, step S4 determines whether the first task is in the startup network stage based on the task network traffic data of the first task and the metadata of the first task. It should be understood that the task in the startup stage (such as pulling images, loading initial data sets, and establishing initial connections) often exhibits a network traffic pattern that is significantly different from that in the stable operation stage, and this stage is extremely sensitive to bandwidth, which directly affects the startup efficiency of the task and the user experience. Therefore, in order to accurately identify whether the task is in this critical startup network stage, thereby providing accurate trigger conditions for subsequent bandwidth enhancement strategies, this application is based on machine learning and multi-source data fusion technology, and analyzes the metadata and real-time network traffic data of the task to make task stage judgments.

[0025] Figure 3 FIG. 4 is a flow chart of step S4 in the bandwidth intelligent scheduling method for a computing cloud platform according to an embodiment of the present application. Figure 3 As shown, the step S4 includes: S41, extracting traffic timing pattern characteristics from the task network traffic data to obtain network traffic timing pattern characteristics; S42, extracting the task type from the metadata of the first task; S43, inputting the network traffic timing pattern characteristics and the task type into the trained machine learning model to obtain a judgment result, and the judgment result is used to indicate whether the first task is in the network startup stage.

[0026] Specifically, the step S41 extracts traffic timing pattern features from the task network traffic data to obtain network traffic timing pattern features. It should be understood that since the network traffic of the task in different operation stages has significant timing dynamic characteristics, for example, the startup stage usually presents a sudden high bandwidth demand, while the stable stage tends to be flat, but traditional statistical methods (such as mean and peak detection) are difficult to capture the complex timing dependencies in the network traffic data. Therefore, in order to accurately characterize the network behavior pattern of the task, in a specific embodiment of the present application, based on multi-scale timing feature extraction technology, a hybrid architecture of LSTM (long short-term memory network) and RNN (recurrent neural network) is used to realize multi-level analysis of traffic patterns. Figure 4 As shown, the step S41 includes: S411, performing a first traffic time series pattern feature extraction based on the LSTM model on the task network traffic data to obtain a first-scale time series pattern feature of the network traffic; S412, performing a second traffic time series pattern feature extraction based on the RNN model on the task network traffic data to obtain a second-scale time series pattern feature of the network traffic; S413, fusing the first-scale time series pattern feature of the network traffic and the second-scale time series pattern feature of the network traffic to obtain the network traffic time series pattern feature.

[0027] Specifically, the collected task network traffic data (i.e., time series data containing metrics such as bandwidth utilization, number of TCP connections, packet loss rate, ingress bandwidth (bytes / sec), egress bandwidth (bytes / sec), and total bytes transmitted) was preprocessed. This included smoothing filtering and Z-score normalization of each metric to scale the data to a uniform range. This eliminated the impact of differences in the numerical distribution of different task traffic metrics. Subsequently, a long short-term memory (LSTM) model was constructed. This model consisted of 128 hidden units and control structures such as forget gates, input gates, and output gates to capture long-term dependencies in the task network traffic data. After the preprocessed task network traffic time series is input into the LSTM model time step by time, the LSTM model uses its internal gating structure to update the task network traffic time series and transfer memory time step by time to capture the long-term dependencies across time steps in the task network traffic time series, such as the overall trend of bandwidth continuously rising and reaching a peak at the initial task startup and then gradually declining or stabilizing. Finally, the output layer extracts the hidden state of each time step and obtains a 128-dimensional long-term time series pattern representation of the network traffic after maximum pooling, namely the first-scale time series pattern feature of the network traffic.

[0028] At the same time, since the LSTM model is good at capturing long-term dependencies in time series data, but is slightly insufficient in capturing short-term change details, in order to more comprehensively analyze the timing characteristics of task network traffic, in the embodiment of the present application, a recurrent neural network (RNN) model is also introduced as a supplement to extract short-term timing pattern features in task network traffic data. Specifically, the RNN model also contains 128 hidden units. After the preprocessed task network traffic time series is also input into the RNN model time step by time, the RNN model transmits information to the task network traffic time series time step by time through its internal cyclic connection structure to capture the short-term dependencies between adjacent time steps in the task network traffic time series, such as sudden bandwidth demand fluctuations, short-term TCP connection surges and other detailed features, and performs pooling and fusion processing on the hidden state of each time step through the output layer, thereby outputting the second-scale timing pattern features of the network traffic of a fixed dimension.

[0029] Subsequently, the first-scale temporal pattern features of network traffic output by the LSTM model and the second-scale temporal pattern features of network traffic output by the RNN model are fused. Through fusion methods such as feature splicing or positional summation, the advantages of both at different time scales are combined to obtain network traffic temporal pattern features that include both long-term trend information in task network traffic data and short-term change details. This provides comprehensive and accurate data support for subsequent task stage judgments, so as to more accurately distinguish the differences in network behavior between the startup stage and other stages.

[0030] Specifically, step S42 extracts the task type from the metadata of the first task. It should be understood that different types of computing tasks have inherent resource demand patterns, particularly during the startup phase, where their dependence on network bandwidth and behavior patterns may vary significantly. For example, a large AI model training task may require pulling a large model file and dataset from remote storage upon startup, resulting in startup network traffic far greater than the startup traffic of a lightweight web service. Without considering the task type, relying solely on traffic timing pattern characteristics may lead to misjudgment of the phase. Therefore, to provide key contextual information about the inherent attributes of the task and make judgments more targeted and accurate, the present application extracts task type information from the metadata of the first task to reveal the basic category and expected behavioral characteristics of the first task. In a specific embodiment of the present application, the metadata structure of the first task is first parsed, the predefined "task type" field is located and extracted, and the corresponding string label is obtained, such as "DeepLearning_Training_Vision", "Spark_Batch_ETL", "Redis_Cache_Init", "Nginx_WebServer_Start", etc. The task type can be specified by the user when submitting the task, or automatically inferred and classified by the platform based on information such as the task image and startup command. For example, if a task uses an official TensorFlow image and the startup command contains a typical training script call, the platform can automatically classify it as an AI training task.

[0031] Specifically, in step S43, the network traffic timing pattern characteristics and the task type are input into the trained machine learning model to obtain a judgment result, and the judgment result is used to indicate whether the first task is in the network startup phase. It should be understood that the network traffic timing pattern characteristics that rely solely on a single data source may mask the true traffic characteristics of the startup phase due to temporary network congestion, thereby affecting the accuracy of the phase judgment. The task type information, as a supplement to the inherent resource demand pattern of the task, helps the model to make more accurate judgments when faced with similar traffic characteristics but different task types. Therefore, in order to achieve highly robust stage identification, this application is based on the principle of multimodal feature collaborative reasoning, using a trained machine learning model to explore the potential correlation and complementary effects between network traffic timing pattern characteristics and task types, thereby making more accurate stage judgments.

[0032] Figure 5 FIG4 is a flow chart of step S43 in the bandwidth intelligent scheduling method for computing cloud platform according to an embodiment of the present application. Figure 5 As shown, the step S43 includes: S431, performing one-hot encoding on the task type to obtain a one-hot embedded coding vector of the task type; S432, splicing the one-hot embedded coding vector of the task type and the network traffic timing pattern feature, and inputting it into the trained machine learning model to obtain the judgment result.

[0033] Specifically, in step S431, the task type is one-hot encoded to obtain a one-hot embedding encoding vector of the task type. It should be understood that since the task type is a discrete categorical variable (such as "deep learning training", "real-time stream processing", etc.), its original label form cannot directly participate in the numerical calculation of the machine learning model. Therefore, in order to convert the semantic information of the task type into quantifiable features and retain the differences between categories, this application uses the one-hot encoding technology based on the discrete feature embedding principle to vectorize the task type to obtain a one-hot embedding encoding vector of the task type.

[0034] In a specific embodiment of the present application, first, task types with a frequency of occurrence exceeding 1% are counted based on historical task records to form a candidate set of task types containing several categories, and a unique index number is assigned to each task type. Then, for each task type in the candidate set, an all-zero vector with a length equal to the size of the candidate set is created (for example, if the candidate set contains 20 categories, the dimension of the vector is 20), and the element value at the position of the task type index number in the all-zero vector is set to 1, thereby generating a one-hot encoding vector representing the task type. In this way, each task type is represented as an independent feature that is orthogonal to other types in the vector space, so that the model can independently learn the association between each task type and the network behavior pattern in the startup phase without being affected by false numerical relationships between categories, thereby improving the effectiveness of model learning and the accuracy of judgment.

[0035] Specifically, in step S432, after splicing the task type one-hot embedding coding vector and the network traffic timing pattern characteristics, it is input into the trained machine learning model to obtain the judgment result. That is, in order to enable the machine learning model to simultaneously perceive and utilize the complementary information in the task type one-hot embedding coding vector and the network traffic timing pattern characteristics to make a more accurate judgment, the present application splices and fuses the task type one-hot embedding coding vector and the network traffic timing pattern characteristics, and inputs the fusion features of the two into the pre-trained machine learning model, so as to utilize the prior knowledge learned by the machine learning model training to dig out the potential correlation and synergy between the task type and the network traffic timing pattern characteristics, thereby outputting a judgment result on whether the first task is in the network startup stage.

[0036] In a specific embodiment of the present application, the step S432 includes: first, performing collaborative convergence constraint encoding based on dual space on the task type one-hot embedding coding vector and the network traffic time series pattern feature to obtain a low-dimensional coupling coding vector of the task type embedding feature and a low-dimensional coupling coding vector of the network traffic time series pattern feature. It should be understood that here, for the multi-scale local time domain correlation pattern feature representation of the network traffic time series pattern feature for the time series data and the one-hot embedding coding representation of the task type one-hot embedding coding vector, it is necessary to define a predetermined transformation in the fusion space to constrain the task type one-hot embedding coding vector and the network traffic time series pattern feature to be fused. Otherwise, the spliced ​​feature vector may be mapped to an infinite open interval based on the respective vectors in the trained machine learning model, thereby affecting the accuracy of the obtained judgment result.

[0037] Therefore, for the task type one-hot embedding encoding vector and the network traffic time series pattern feature vector, first, dual space mapping is performed on the task type one-hot embedding encoding vector and the network traffic time series pattern feature to obtain the task type embedding feature dual space mapping encoding vector and the network traffic time series pattern feature dual space mapping encoding vector, which can be expressed as follows: ;in, represents the one-hot embedding encoding vector of the task type, represents the network traffic time series pattern feature vector, that is, the network traffic time series pattern feature, Indicates point multiplication by position, is a natural constant, Represents the natural exponential function value of each eigenvalue in the calculation vector, Represents the task type embedding feature dual space mapping encoding vector, Represents the dual space mapping encoding vector of network traffic temporal pattern features.

[0038] That is, by utilizing the isometric embedding characteristics of the features of different modes into their high-order dual space, we can construct the generalized dual mapping coefficients through continuous natural distribution convergence, thereby completely mapping the dual space.

[0039] Then, local convexity extraction is performed on the dual space mapping encoding vector of the task type embedding feature and the dual space mapping encoding vector of the network traffic timing pattern feature to obtain the task type-network traffic timing pattern feature interaction attention weight vector and the network traffic timing pattern-task type feature interaction attention weight vector. In other words, the projected ideal subspace is constructed using the transformation kernel induced by the quotient mapping. That is, the projected local convexity of the quotient space under the dual space is utilized to achieve an ideal quasi-linear form, thereby ensuring convergence in the fusion space. This can be expressed as follows: ;in, Indicates the reciprocal of each eigenvalue in the calculation vector, Represents the attention weight vector of the interaction between task type and network traffic temporal pattern features, Represents the attention weight vector of the interaction between network traffic timing pattern and task type features.

[0040] Finally, based on the task type-network traffic time series pattern feature interaction attention weight vector and the network traffic time series pattern-task type feature interaction attention weight vector, the task type one-hot embedding coding vector and the network traffic time series pattern feature are subjected to dual space convergence constraints to obtain the task type embedding feature low-dimensional coupling coding vector and the network traffic time series pattern feature low-dimensional coupling coding vector. That is, the task type one-hot embedding coding vector is constrained by using the low-dimensional norm as the convergence neighborhood basis in the dual space. and the network traffic timing pattern feature vector Perform low-dimensional coupling encoding, which can be expressed as follows: ;in, represents the calculation of the vector's two norm, Represents the low-dimensional coupled encoding vector of the task type embedding feature, Represents the low-dimensional coupled encoding vector of network traffic temporal pattern characteristics.

[0041] In this way, the task type to be fused can be kept in a unique hot embedding encoding vector and the network traffic timing pattern characteristics The independence between the coupling principal components of the fusion space low-dimensional surface of the feature distribution is maintained while maintaining the overall distribution stability of the feature coupling, thereby improving the accuracy of the obtained judgment results.

[0042] Then, after the task type embedding feature low-dimensional coupling coding vector and the network traffic timing pattern feature low-dimensional coupling coding vector are spliced, they are input into the trained machine learning model to obtain the judgment result. In a specific example of the present application, the machine learning model is a support vector machine model. Specifically, the support vector machine (SVM) model uses a radial basis function (RBF) as a kernel function, and in the offline training stage, historical task data of different task types are used as training samples, wherein each training sample contains the task type embedding feature low-dimensional coupling coding vector of its task sample, the network traffic timing pattern feature low-dimensional coupling coding vector and a true label indicating whether the task is in the network startup stage, wherein the task type embedding feature low-dimensional coupling coding vector and the network traffic timing pattern feature low-dimensional coupling coding vector of each training sample are feature extracted and encoded using the above-mentioned method. The support vector machine model continuously optimizes its internal decision boundary by learning the nonlinear mapping relationship between the multimodal spliced ​​fusion features and the true labels in each training sample. This approach finds the optimal decision hyperplane in the high-dimensional feature space to maximize the separation between samples of different categories (i.e., tasks in the network startup phase and tasks not in the network startup phase). Furthermore, the support vector machine model is trained using 5-fold cross-validation and the SMOTE algorithm to balance positive and negative samples (initiated and non-initiation phases) to reduce training bias caused by sample imbalance. Model parameters, such as the penalty coefficient C and kernel function parameter γ, are continuously adjusted to optimize the model's classification accuracy on the test set. Ultimately, the optimal parameter configuration with the highest classification accuracy is selected to yield the trained support vector machine model. After the model is deployed, for new input samples (i.e., the concatenated fusion features of the low-dimensional coupled encoding vector of the task type embedding feature and the low-dimensional coupled encoding vector of the network traffic temporal pattern feature), the SVM calculates the distance d between the input sample and the hyperplane and uses a sigmoid function to convert the output of the SVM decision function into a probability value indicating that the input sample belongs to the "Network Startup Phase." Based on a preset threshold (e.g., 0.8), input samples with a probability greater than or equal to the preset threshold are considered to be in the "Network Startup Phase" and trigger the corresponding bandwidth scheduling policy. Otherwise, the input sample is considered not to be in the "Network Startup Phase" and the current bandwidth allocation status is maintained. This approach effectively combines the task's identity information (task type in metadata) with its real-time behavior (traffic pattern) and leverages the high-performance SVM classifier to make decisions, significantly improving the accuracy and generalization of judgments, providing a reliable signal for subsequent intelligent bandwidth scheduling.

[0043] In the aforementioned intelligent bandwidth scheduling method for a computing cloud platform, step S5, in response to the first task being in the network startup phase, sends a request to the intelligent bandwidth scheduler to apply an enhancement policy. Specifically, after determining that the first task is in the bandwidth-sensitive network startup phase, in order to trigger subsequent targeted bandwidth guarantee mechanisms and ensure that the task can quickly and efficiently complete the startup period, this application, based on the event-driven principle, proactively issues an explicit application enhancement instruction request to the intelligent bandwidth scheduler responsible for specific bandwidth resource allocation after determining that the task is in the network startup phase. In a specific embodiment of this application, this application enhancement policy request is a structured message or API call, which should at least include the unique identifier of the first task (such as the task ID, container ID, or Pod name) and a flag or command code that explicitly indicates the need to apply the bandwidth enhancement policy; optionally, it may also include some contextual information, such as the desired QoS (Quality of Service) improvement level. This request is sent to the intelligent bandwidth scheduler via internal system communication mechanisms, such as remote procedure calls (gRPC, Thrift), message queues (such as RabbitMQ and Kafka, which allow for asynchronous processing and decoupling), or direct function calls. The intelligent bandwidth scheduler is a component specifically responsible for dynamically adjusting network resource allocation. It maintains information such as network topology, link capacity, current bandwidth usage by each task, and QoS policy configuration. By transmitting intelligent perception results of the task initiation phase to the control unit that actually performs bandwidth adjustments (i.e., the intelligent bandwidth scheduler), the operational process of providing priority bandwidth guarantees for specific tasks is initiated.

[0044] In the aforementioned intelligent bandwidth scheduling method for a computing cloud platform, in step S6, upon receiving the request to apply the enhanced policy, the intelligent bandwidth scheduler assigns a high-priority QoS tag to the first computing task to generate bandwidth adjustment policy instructions. Specifically, in order to convert the request to apply the enhanced policy into actual preferential treatment for the network traffic of the first computing task, thereby significantly improving its network performance during the startup phase, this application, based on the principle of QoS (Quality of Service) assurance, assigns a high-priority tag to the traffic of the first computing task through the intelligent bandwidth scheduler to generate executable bandwidth adjustment instructions.

[0045] In a specific embodiment of the present application, after receiving the application enhancement policy request, the intelligent bandwidth scheduler first verifies the request to confirm the legitimacy of the request source and the validity of the task ID. After the verification is passed, the intelligent bandwidth scheduler queries the task information library and network status library it maintains to determine the current network connection information of the first computing task (such as IP address, port number, computing node and its network interface, etc.), and assigns a high-priority QoS label to the network traffic of the task. More specifically, assigning a high-priority QoS label to the first computing task can be achieved in a variety of ways: for example, setting a specific DSCP (Differentiated Services Code Point) value in the IP header of the data packet, which can be recognized by network devices that support DiffServ (such as switches and routers), and accordingly puts the data packet into a high-priority queue for forwarding. Alternatively, in a software-defined networking (SDN) environment, the intelligent bandwidth scheduler interacts with the SDN controller to issue flow table rules (such as OpenFlow rules) to data plane devices (such as OpenvSwitch or physical switches). This assigns higher queue priorities, larger guaranteed bandwidths, or lower packet loss rates to flow entries whose source / destination IP addresses or ports match the task's traffic. Finally, the high-priority QoS label assignments are translated into specific configuration commands or actions. For example, on the compute node where the task resides, traffic control tools (such as the tc command in Linux with queue disciplines like HTB and HFSC) can be used to modify the network interface configuration of the task's corresponding network namespace or cgroup, assigning it a higher bandwidth floor, weight, or lower rate ceiling (to ensure that other tasks are not starved of bandwidth when overall bandwidth is limited, while still maintaining their priority). In virtualized environments, such as OpenStack Neutron combined with OVS, the intelligent bandwidth scheduler can call the cloud platform's network service API to modify the QoS policy rules for the virtual port (vPort) connected to the task's VM, such as setting a specific minimum bandwidth guarantee (min_bw) and maximum bandwidth limit (max_bw) for the vPort and ensuring its priority. Furthermore, if the system supports bandwidth reservation, a portion of bandwidth resources will be temporarily reserved for the task on its communication path. Subsequently, the bandwidth adjustment policy instructions are distributed to the relevant network execution points (e.g., the network management agent on the compute node, the virtual switch controller, or the physical switch management interface) for execution. Furthermore, the intelligent bandwidth scheduler has the function of revoking the enhanced policy. That is, after the first computing task successfully launches and enters the stable operation phase, or after determining that the task launch has failed, the intelligent bandwidth scheduler will be notified again to restore the QoS tag to the normal level, releasing the prioritized resources.In this way, critical tasks in the network startup phase are guaranteed bandwidth, and their data pulling, dependency loading and other processes are accelerated, thereby significantly shortening the overall startup time of the task, improving user experience and platform operation efficiency; at the same time, once the task is running stably, priority resources will be recycled, avoiding continued unfair impact on other tasks and ensuring the overall optimized utilization of platform resources.

[0046] In summary, according to the embodiment of the present application, the bandwidth intelligent scheduling method for the computing power cloud platform is explained, which, after receiving the computing task request, assigns the task to the computing node through the task scheduler, stores the task metadata, and synchronously starts the network performance monitoring module to monitor the network traffic of the task. Furthermore, based on the stored task metadata and the collected task network traffic data, a deep learning algorithm is used to perform an in-depth joint analysis of the two to identify whether the task is in the network startup stage. If it is determined that the task is in the network startup stage, an application enhancement policy request is sent to the intelligent bandwidth scheduler to assign a high-priority QoS label to the task for bandwidth adjustment. In this way, the network demand characteristics of the task at different stages of the life cycle can be captured in a fine-grained manner, and the bandwidth resource allocation can be dynamically adjusted to ensure the bandwidth demand in the task startup stage and improve the utilization of bandwidth resources and the efficiency of task execution.

[0047] Furthermore, the present application also provides a bandwidth intelligent scheduling system for a computing power cloud platform.

[0048] Figure 6 FIG is a block diagram of a bandwidth intelligent scheduling system for a computing cloud platform according to an embodiment of the present application. Figure 6As shown, the bandwidth intelligent scheduling system 100 for the computing power cloud platform according to the embodiment of the present application comprises: a computing node allocation module 110, configured to allocate a first task to a certain computing node by a task scheduler after receiving a task request of the first computing task; a network flow monitoring module 120, configured to send metadata of the first task to a task metadata storage module for storage by the task scheduler, and notify a network performance monitoring module to monitor network flow of the first task; a data extraction module 130, configured to acquire task network flow data of the first task collected by the network performance monitoring module and extract metadata of the first task from the task metadata storage module; a network phase identification module 140, configured to determine whether the first task is in a start-up network phase based on the task network flow data of the first task and the metadata of the first task; a bandwidth scheduling request module 150, configured to send a request for applying an enhanced strategy to an intelligent bandwidth scheduler in response to the first task being in the start-up network phase; and a bandwidth adjustment module 160, configured to assign a high-priority QoS label to the first computing task to generate a bandwidth adjustment strategy instruction in response to the intelligent bandwidth scheduler receiving the request for applying the enhanced strategy.

[0049] Here, those skilled in the art can understand that the specific operations of each module in the above bandwidth intelligent scheduling system for the computing power cloud platform have been described in detail in the above description of the bandwidth intelligent scheduling method for the computing power cloud platform according to the embodiment of the present application, and therefore, the repeated description thereof will be omitted. Figures 1 to 5

[0050] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application and not to limit the present application. Although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present application.​

Claims

1. A bandwidth intelligent scheduling method for a computing power cloud platform, characterized in that: include: After receiving the task request of the first computing task, the task scheduler assigns the first task to a computing node; The task scheduler sends the metadata of the first task to the task metadata storage module for storage, and notifies the network performance monitoring module to monitor the network traffic of the first task; Acquire task network traffic data of the first task collected by the network performance monitoring module and extract metadata of the first task from the task metadata storage module; determining, based on the task network traffic data of the first task and the metadata of the first task, whether the first task is in a network startup phase; In response to the first task being in the network startup phase, sending a request to the intelligent bandwidth scheduler to apply the enhancement policy; After receiving the request to apply the enhanced policy, the intelligent bandwidth scheduler assigns a high priority QoS label to the first computing task to generate a bandwidth adjustment policy instruction.

2. The bandwidth intelligent scheduling method for a computing cloud platform according to claim 1 is characterized in that: The metadata includes task type, image, mounted volume, priority, and user annotations.

3. The bandwidth intelligent scheduling method for computing cloud platform according to claim 2 is characterized in that: Determining whether the first task is in a network startup phase based on task network traffic data of the first task and metadata of the first task includes: Extracting traffic time series pattern features from the task network traffic data to obtain network traffic time series pattern features; extracting a task type from metadata of the first task; The network traffic timing pattern characteristics and the task type are input into the trained machine learning model to obtain a judgment result, and the judgment result is used to indicate whether the first task is in the network startup stage.

4. The bandwidth intelligent scheduling method for a computing cloud platform according to claim 3 is characterized in that: Extracting traffic time series pattern features from the task network traffic data to obtain network traffic time series pattern features includes: Performing a first traffic time series pattern feature extraction based on an LSTM model on the task network traffic data to obtain a first-scale time series pattern feature of the network traffic; Performing RNN model-based second traffic time series pattern feature extraction on the task network traffic data to obtain a second-scale time series pattern feature of the network traffic; The first-scale time series pattern feature of the network traffic and the second-scale time series pattern feature of the network traffic are fused to obtain the network traffic time series pattern feature.

5. The bandwidth intelligent scheduling method for computing cloud platform according to claim 3 is characterized in that: Inputting the network traffic time series pattern feature and the task type into a trained machine learning model to obtain a judgment result, where the judgment result is used to indicate whether the first task is in the network startup phase, including: One-hot encoding the task type to obtain a one-hot embedding encoding vector of the task type; After the task type one-hot embedding encoding vector and the network traffic timing pattern feature are spliced, they are input into the trained machine learning model to obtain the judgment result.

6. The bandwidth intelligent scheduling method for computing cloud platform according to claim 5 is characterized in that: The machine learning model is a support vector machine model.

7. The bandwidth intelligent scheduling method for computing cloud platform according to claim 5 is characterized in that: After concatenating the task type one-hot embedding encoding vector and the network traffic time series pattern feature, the concatenated vectors are input into the trained machine learning model to obtain the judgment result, including: Performing collaborative convergence constraint encoding based on dual space on the task type one-hot embedding coding vector and the network traffic time series pattern feature to obtain a task type embedding feature low-dimensional coupling coding vector and a network traffic time series pattern feature low-dimensional coupling coding vector; After the task type embedding feature low-dimensional coupling coding vector and the network traffic timing pattern feature low-dimensional coupling coding vector are spliced, they are input into the trained machine learning model to obtain the judgment result.

8. The bandwidth intelligent scheduling method for a computing power cloud platform according to claim 7 is characterized in that: Performing collaborative convergence constraint encoding based on dual space on the task type one-hot embedding coding vector and the network traffic time series pattern feature to obtain a task type embedding feature low-dimensional coupling coding vector and a network traffic time series pattern feature low-dimensional coupling coding vector, including: Performing dual space mapping on the task type one-hot embedding encoding vector and the network traffic temporal pattern feature to obtain a task type embedding feature dual space mapping encoding vector and a network traffic temporal pattern feature dual space mapping encoding vector; Performing local convex feature extraction on the task type embedding feature dual space mapping encoding vector and the network traffic timing pattern feature dual space mapping encoding vector to obtain a task type-network traffic timing pattern feature interaction attention weight vector and a network traffic timing pattern-task type feature interaction attention weight vector; Based on the task type-network traffic timing pattern feature interaction attention weight vector and the network traffic timing pattern-task type feature interaction attention weight vector, the task type one-hot embedding coding vector and the network traffic timing pattern feature are subjected to dual space convergence constraints to obtain the task type embedding feature low-dimensional coupling coding vector and the network traffic timing pattern feature low-dimensional coupling coding vector.

9. A bandwidth intelligent scheduling system for a computing cloud platform, characterized in that: include: A computing node allocation module, configured to allocate the first computing task to a computing node after receiving a task request of the first computing task; a network traffic monitoring module, configured to use a task scheduler to send metadata of the first task to a task metadata storage module for storage, and to notify a network performance monitoring module to monitor network traffic of the first task; a data extraction module, configured to obtain the task network traffic data of the first task collected by the network performance monitoring module and extract metadata of the first task from the task metadata storage module; a network stage identification module, configured to determine whether the first task is in a network startup stage based on task network traffic data of the first task and metadata of the first task; a bandwidth scheduling request module, configured to send a request for applying an enhancement policy to the intelligent bandwidth scheduler in response to the first task being in the network startup phase; The bandwidth adjustment module is configured to assign a high priority QoS label to the first computing task to generate a bandwidth adjustment policy instruction in response to the intelligent bandwidth scheduler receiving the request to apply the enhanced policy.

Images