A network cluster computing power scheduling system and method
By introducing a time-series prediction model and weight calculation, the allocation of computing resources in the network cluster is dynamically adjusted, solving the problems of resource allocation imbalance and regional management in traditional load balancing schemes, and achieving more efficient resource utilization and stability.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HANGZHOU ELECTRONIC INFORMATION VOCATIONAL SCHOOL
- Filing Date
- 2026-03-23
- Publication Date
- 2026-07-21
AI Technical Summary
Traditional load balancing solutions cannot accurately identify peak and off-peak traffic periods, leading to unbalanced resource allocation, wasted hardware resources, and delayed user response. Furthermore, they are difficult to manage regionally, impacting user experience and system stability.
Temporal prediction models (such as Temporal Fusion Transformer) are used to analyze traffic change patterns, predict future traffic data, and dynamically adjust resource allocation based on the computing power and latency weights of the network cluster to achieve precise load scheduling.
It improves resource utilization, reduces hardware resource waste and energy costs, and ensures system stability and user experience under different load conditions.
Smart Images
Figure CN122437845A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of network computing power allocation technology, and in particular to a network cluster computing power scheduling system and method. Background Technology
[0002] In the existing fields of computer networks and distributed computing, load balancing technology is widely used to rationally distribute user requests to various computing nodes in order to improve overall system performance and resource utilization. However, traditional load balancing solutions have two significant problems in practical applications, which restrict the efficient operation of the system and urgently need to be improved. First, traditional load balancing solutions suffer from severe resource imbalance. This stems from their lack of accurate identification of peak and off-peak traffic periods and their inability to adjust in real-time based on dynamic traffic changes. During off-peak periods, the inability to detect load reduction leads to many computing nodes remaining operational, resulting in multiple idle and redundant nodes. This not only wastes hardware resources but also increases system energy costs and reduces overall resource utilization efficiency. Conversely, during peak periods, the failure to anticipate and effectively distribute the load causes some computing nodes to become overloaded due to excessive user requests. At this point, the computing nodes reach their processing limits, leading to delayed responses to user requests and, in severe cases, service interruptions, significantly impacting user experience and system stability. Secondly, traditional load balancing solutions struggle to achieve effective regional management. In these solutions, user requests are typically sent directly to the computing cluster for processing, but peak and off-peak traffic times vary significantly across different regions. This difference stems primarily from factors such as different time zones, varying user habits, and differing local economic activity patterns. Because traditional scheduling systems lack targeted analysis and management mechanisms for regional traffic characteristics, they suffer from "selection difficulties" when faced with these regional traffic peak and off-peak variations. They cannot accurately schedule loads based on the actual traffic conditions in each region, making it difficult to allocate appropriate computing resources to the corresponding regions at the right time. This results in some regions experiencing insufficient computing resources during peak traffic periods, leading to a decline in service quality; while other regions experience significant idle computing resources during off-peak periods, wasting resources and further exacerbating the overall irrationality of resource allocation. Summary of the Invention
[0003] This invention mainly solves the above-mentioned problems and provides a method for scheduling computing power in a network cluster.
[0004] The technical solution adopted by this invention to solve its technical problem is a network cluster computing power scheduling method, comprising: obtaining the network access address of each network cluster management terminal; establishing a time series prediction model and inputting a traffic dataset to train the time series prediction model; obtaining and storing the total computing power value of each network cluster; obtaining and storing the latency of each network cluster; continuously obtaining and statistically analyzing the traffic records requested by the user terminal from the load balancing system, and inputting them into the trained time series prediction model to predict future traffic values; statistically analyzing the current total throughput of the network cluster, and when the current total throughput of the network cluster is less than the predicted traffic value, assigning weights to each network cluster, and having the load balancing system transfer traffic to each network cluster according to the weights of each network cluster.
[0005] As a preferred embodiment of the above scheme, the time series prediction model is built based on Temporal FusionTransformer.
[0006] As a preferred option of the above scheme, offline computing nodes are filtered out when obtaining the total computing power value of each network cluster.
[0007] As a preferred embodiment of the above scheme, the latency of each network cluster includes ping latency and TCP handshake latency.
[0008] As a preferred embodiment of the above scheme, the weight of the network cluster is the cluster's computing power divided by the cluster's latency and then divided by the sum of the reciprocals of the latency of all clusters.
[0009] As a preferred embodiment of the above scheme, when the total throughput of the current network cluster is less than the predicted traffic value, the load balancing system forwards traffic to the network clusters that have participated in the calculation based on the weight of each network cluster, and selects idle network clusters to join the calculation based on the weight of the difference between the total throughput of the current network cluster and the predicted traffic value.
[0010] As a preferred embodiment of the above scheme, when the total throughput of the current network cluster is greater than the predicted traffic value, the load balancing system forwards traffic to the network clusters that have participated in the calculation based on the weight of each network cluster, and removes the network cluster based on the weight according to the difference between the total throughput of the current network cluster and the predicted traffic value.
[0011] The present invention also provides a network cluster computing power scheduling system, including a load balancing system, an intelligent scheduling system, and a time-series prediction module; the time-series prediction module is used to predict the traffic value in a future time period; the intelligent scheduling system obtains the computing power value and latency of each network cluster and assigns weights to each network cluster; the load balancing system allocates computing power to each network cluster according to the predicted traffic value in the future time period and the network cluster weights.
[0012] As a preferred embodiment of the above scheme, when the predicted traffic value for a future time period is greater than the total computing power of the current network cluster set participating in the operation, the load balancing system selects an idle network cluster to add to the network cluster set according to its weight in descending order.
[0013] As a preferred embodiment of the above scheme, when the predicted traffic value for a future time period is less than the total computing power of the current network cluster set participating in the operation, the load balancing system selects network clusters to leave the network cluster set according to their weights in ascending order.
[0014] The advantages of this invention are: it effectively solves the problem of unbalanced resource allocation. Addressing the shortcomings of traditional solutions that cannot identify peak and off-peak traffic periods and cannot dynamically adjust, this invention introduces a time-series prediction model. By analyzing traffic variation patterns and predicting traffic data for the next time period, the intelligent scheduling system can plan computing resources in advance. Before peak traffic periods, the system can pre-allocate sufficient computing power to avoid single-point overload and response lag; during off-peak traffic periods, it can promptly reduce the number of running computing nodes to prevent multiple points of idleness and redundant operation. This mechanism significantly improves resource utilization, reduces hardware resource waste and energy costs, while ensuring system stability under different load conditions and improving user experience. Attached Figure Description
[0015] Figure 1 This is a diagram illustrating the collaboration process of each module.
[0016] Figure 2 Flowchart for obtaining cluster computing power values.
[0017] Figure 3 This is a flowchart for delayed data retrieval.
[0018] Figure 4 This is a flowchart of the scheduling algorithm. Detailed Implementation
[0019] The technical solution of the present invention will be further described below through embodiments and in conjunction with the accompanying drawings.
[0020] Example: This embodiment describes a network cluster computing power scheduling method, which requires the coordinated operation of various modules, such as... Figure 1 As shown, the intelligent scheduling system continuously plans the resource allocation weights for each network cluster based on the network cluster's floating-point computing power, cluster network latency, and predicted traffic volume, and synchronizes this information to the load balancing system. The load balancing system forwards traffic destined for itself to each network cluster according to the resource allocation weights continuously calculated by the intelligent scheduling system. During the calculation of resource allocation weights, the intelligent scheduling system also uses a time-series prediction model to predict traffic volume in future time periods, using this as a cluster computing power indicator to accurately control the computing power and resource utilization of each network cluster.
[0021] The workflow of the network cluster computing power scheduling method in this embodiment is as follows: Before launching the intelligent scheduling system, two configurations are required: First, configure the network access addresses for each computing cluster management terminal. Next, write the traffic dataset used to train the time-series prediction model. The dataset needs to contain key information such as time periods and the corresponding traffic volume for each time period. Furthermore, the time intervals for each record in the dataset must be equal, and the dataset must contain hundreds of records. To adapt to new application scenarios for the network cluster, simply write a new traffic dataset. This significantly improves the flexibility of this method of scheduling network cluster computing power using time-series prediction across multiple scenarios.
[0022] After the intelligent scheduling system starts, it will automatically train and use a time-series prediction model based on the TemporalFusion Transformer architecture based on the provided dataset. This architecture can capture various dependencies, thereby achieving more accurate predictions in complex scenarios. Once the dataset is started, the intelligent scheduling system will launch four background tasks: The first background task is as follows Figure 2 As shown, the intelligent scheduling system first establishes a connection with the configured management terminals of each computing power cluster. After the connection is established, the intelligent scheduling system will periodically send requests to the computing power cluster management terminals. Upon receiving the requests, the management terminals will collect the computing power values of all the computing nodes they manage and return them to the scheduling system. Since computing nodes can be dynamically added and deleted, the scheduling system also needs to filter out offline nodes. Finally, the computing power values of all computing nodes are summed to obtain the total computing power value of all online network clusters and stored.
[0023] The second background task is as follows Figure 3 As shown, the process begins by repeatedly sending ping packets to obtain the ping latency, representing the basic network connectivity, in each network cluster. Next, t1 is set to the current timestamp, a socket connection is created, and after creation, t2 is set to the current timestamp. Then, tcp_handshake_delay is set to t2-t1, representing the TCP handshake latency of the server's responsiveness. This operation is repeated multiple times. Subsequently, the impact of network fluctuations is offset by subtracting the maximum and minimum latency values and taking the average. Finally, the two latency values are weighted and summed to obtain the final result, which is then stored.
[0024] The third background task is responsible for establishing a connection with the load balancer server. If the load balancer server receives a request from a user, it reports the traffic records to the intelligent scheduling system, which will automatically compile statistics.
[0025] The fourth background task is as follows: Figure 4As shown, it realizes computing power scheduling and is executed regularly when the system time is close to the end of the dataset time period. At the start of the task, the system first uses the continuously statistical traffic history to predict the future traffic size and stores it as the forecasted variable. Subsequently, total_max_flow = 0 is set, and by traversing all the currently participating network clusters, total_max_flow is accumulated to obtain the total throughput of the currently participating network clusters.
[0026] By comparing total_max_flow and forecasted, it is determined whether computing power scheduling is required. When performing scheduling, it is also necessary to calculate the weights assigned to each network cluster. The weight of a network cluster is the computing power value of the cluster divided by the cluster latency and then divided by the sum of the reciprocals of the latencies of all clusters. This formula can achieve a higher score when the computing power value is higher and the latency is lower, thus accelerating the overall response speed.
[0027] When total_max_flow < forecasted, the load balancing system forwards traffic to the already participating network clusters based on the weights of each network cluster, and selects idle network clusters to join the operation based on the weight according to the difference between the total throughput of the current network cluster and the predicted traffic value. Ensure that the computing power of the network cluster meets the user's exact needs and ensure the user experience and system stability. When total_max_flow > forecasted, the load balancing system forwards traffic to the already participating network clusters based on the weights of each network cluster, and removes network clusters based on the weight according to the difference between the total throughput of the current network cluster and the predicted traffic value. Avoid wasting resources caused by excessive computing power.
[0028] Correspondingly, this embodiment also provides a network cluster computing power scheduling system, including a load balancing system, an intelligent scheduling system, and a timing prediction module; the timing prediction module is used to predict the traffic value in the future time period; the intelligent scheduling system is used to obtain the computing power value and latency of each network cluster and assign weights to each network cluster; the load balancing system is used to allocate computing power to each network cluster according to the predicted traffic value in the future time period according to the network cluster weights.
[0029] When the predicted traffic value in the future time period is greater than the total computing power of the current set of participating network clusters, the load balancing system selects idle network clusters to join the network cluster set in descending order of weight; when the predicted traffic value in the future time period is less than the total computing power of the current set of participating network clusters, the load balancing system selects network clusters to leave the network cluster set in ascending order of weight.
[0030] The network cluster computing power scheduling system and method in this embodiment effectively solves the problem of resource allocation imbalance. It introduces a time-series prediction model, analyzing traffic change patterns and predicting traffic data for the next time period, enabling the intelligent scheduling system to plan computing resources in advance. Before peak traffic periods, the system can pre-allocate sufficient computing power to avoid single-point overload and response lag; during off-peak traffic periods, it can promptly reduce the number of running computing nodes to prevent multiple points of idleness and redundant operation. This mechanism significantly improves resource utilization, reduces hardware resource waste and energy costs, while ensuring system stability under different load conditions and improving user experience.
[0031] The specific embodiments described herein are merely illustrative of the spirit of the invention. Those skilled in the art to which this invention pertains may make various modifications or additions to the described specific embodiments or use similar methods to substitute them, without departing from the spirit of the invention or exceeding the scope defined by the appended claims.
Claims
1. A network cluster computing power scheduling method, characterized by: include: Obtain the network access addresses of each network cluster management terminal; Establish a time series prediction model and train it using a traffic dataset; obtain and store the total computing power of each network cluster; obtain and store the latency of each network cluster; continuously obtain and count the traffic records of user requests to the load balancing system, and input them into the trained time series prediction model to predict future traffic values. The system calculates the total throughput of the current network cluster. When the total throughput of the current network cluster is less than the predicted traffic value, it assigns weights to each network cluster and the load balancing system forwards traffic to each network cluster according to their weights.
2. The network cluster computing power scheduling method according to claim 1, characterized in that: The time series prediction model is built on Temporal Fusion Transformer.
3. The network cluster computing power scheduling method according to claim 1, characterized in that: When obtaining the total computing power of each network cluster, offline computing nodes are filtered out.
4. The network cluster computing power scheduling method according to claim 1, characterized in that: The latency of each network cluster includes ping latency and TCP handshake latency.
5. The network cluster computing power scheduling method according to claim 1, characterized in that: The weight of the network cluster is the cluster's computing power divided by the cluster's latency, and then divided by the sum of the reciprocals of the latency of all clusters.
6. The network cluster computing power scheduling method according to claim 1, characterized in that: in When the current total throughput of the network cluster is less than the predicted traffic value, the load balancing system forwards traffic to the network clusters that have participated in the computation based on the weight of each network cluster, and selects idle network clusters to join the computation based on the weight of the difference between the current total throughput of the network cluster and the predicted traffic value.
7. The network cluster computing power scheduling method according to claim 1 or 6, characterized in that: in When the total throughput of the current network cluster is greater than the predicted traffic value, the load balancing system forwards traffic to the network clusters that have participated in the calculation based on the weight of each network cluster, and removes the network cluster based on the weight of the difference between the total throughput of the current network cluster and the predicted traffic value.
8. A network cluster computing power scheduling system, characterized in that: Includes a load balancing system, an intelligent scheduling system, and a time-series prediction module; The time-series prediction module is used to predict the flow rate for future time periods; The intelligent scheduling system obtains the computing power and latency of each network cluster and assigns weights to each network cluster. The load balancing system allocates computing power to each network cluster according to the network cluster weight based on the predicted traffic value for the future time period.
9. The network cluster computing power scheduling system according to claim 8, characterized in that: When the predicted traffic value for a future time period exceeds the total computing power of the current network cluster set participating in the operation, the load balancing system selects an idle network cluster to add to the network cluster set according to its weight in descending order.
10. The network cluster computing power scheduling system according to claim 8, characterized in that: When the predicted traffic value for a future time period is less than the total computing power of the current network cluster set participating in the operation, the load balancing system selects a network cluster to leave the network cluster set according to its weight in ascending order.