CN-121984920-A - Service quality perception multi-target load balancing method and system for computing network

CN121984920ACN 121984920 ACN121984920 ACN 121984920ACN-121984920-A

Abstract

The invention provides a service quality perception multi-target load balancing method and a system for an algorithm network, comprising the steps of obtaining service flow data, and carrying out service quality prediction on the service flow data to obtain a service quality type; the method comprises the steps of obtaining a graph embedded representation of a network state from a pre-constructed graph structure, determining a plurality of forwarding sub-domains with consistent performance according to the graph embedded representation, constructing a plurality of forwarding paths with different performance characteristics by using a hierarchical path planning strategy according to the forwarding sub-domains, determining a scheduling state space according to the service quality type and the forwarding paths, and mapping the scheduling state space into a path selection strategy by using a pre-constructed multi-objective optimization strategy model. The invention supports the real-time QoS identification of flow, the construction of performance aware paths, the self-adaption of scheduling granularity and the multi-objective joint optimization capability, and improves the resource scheduling efficiency and the service guaranteeing capability of the computing network under the high concurrency of multiple services.

Inventors

ZHANG WEITING
ZHANG CHENGRUI
YANG DONG
GAO DEYUN
ZHANG HONGKE

Assignees

北京交通大学

Dates

Publication Date: 20260505
Application Date: 20260311

Claims (10)

1. The service quality perception multi-target load balancing method for the computing power network is characterized by comprising the following steps of: Acquiring service flow data, and carrying out service quality prediction on the service flow data to obtain a service quality type; Obtaining a graph embedded representation of a network state from a pre-constructed graph structure, wherein the graph structure is constructed based on network state information of each node; determining a plurality of forwarding sub-domains with consistent performance according to the graph embedded representation, and constructing a plurality of forwarding paths with different performance characteristics by using a hierarchical path planning strategy according to the forwarding sub-domains; And mapping the scheduling state space into a path selection strategy by utilizing a pre-constructed multi-objective optimization strategy model, wherein the multi-objective optimization strategy model is a model based on a deep reinforcement learning algorithm.
2. The method for computing-network-oriented quality of service aware multi-objective load balancing according to claim 1, wherein the traffic flow data comprises a first packet message of a traffic flow; the predicting the service quality of the service flow data to obtain the service quality type comprises the following steps: Extracting static characteristics from the first packet message, and determining time sequence characteristics based on the static characteristics of continuous T service flows; And inputting the time sequence characteristics into a transducer model, and predicting to obtain the service quality type.
3. The method for computing network oriented quality of service aware multi-objective load balancing of claim 2, wherein the transducer model is a self-attention mechanism based transducer model comprising a multi-headed self-attention mechanism and a feed forward network; the inputting the time sequence characteristic into a transducer model, predicting and obtaining the service quality type comprises the following steps: Performing leachable transformation projection on the time sequence characteristics to obtain projected characteristics; Capturing position information from the projected features, and taking the position information and the projected features as input features; and sequentially inputting the input characteristics into the multi-head self-attention mechanism and a feedforward network to obtain the quality of service type.
4. The method for power network oriented quality of service aware multi-objective load balancing of claim 1, wherein the graph structure comprises a set of nodes; the obtaining the graph embedded representation of the network state from the pre-constructed graph structure comprises the following steps: performing attention calculation on each node in the node set by using a graph attention mechanism to obtain an attention coefficient; obtaining graph attention embedding by utilizing the attention coefficient aggregation field information; Respectively generating a positive sample pair and a negative sample set according to the graph meaning force embedding; Utilizing a contrast learning mechanism to learn the positive sample pair and the negative sample set to obtain contrast learning embedding; and fusing the graph annotation force embedding with the contrast learning embedding to obtain the graph embedding representation.
5. The method of power network oriented quality of service aware multi-objective load balancing of claim 1, wherein the graph embedded representation comprises a graph attention embedding and a contrast learning embedding; said determining a plurality of forwarding sub-domains of performance consistency based on said graph embedded representation, comprising: Calculating to obtain node characteristic similarity factors between the graph embedded representation of the node i and the graph embedded representation of the node j by using the following formula : ; In the formula, The weight balance coefficients of the attention embedding and the contrast learning embedding of the graph; representing a normalized scale factor of the graph attention embedding; Representing the normalized scale factors embedded by contrast learning; is the graph-note force-embedded representation of node j , Is the graph meaning force embedding of the node j; is a contrast learning embedding of node i, Is the contrast learning embedding of the node j; based on the node characteristic similarity factor And calculating to obtain the enhancement factor by using the following enhancement factor function: ; According to the enhancement factor, the node Subdomain identification to which it belongs And Subdomain identification to which it belongs Correcting the modularity evaluation index to obtain a corrected modularity evaluation index : ; In the formula, Is the total number of edges in the graph structure; representing contiguous matrix elements, when a node And (3) with When an edge is present between the two, Take the value 1, otherwise Taking a value of 0; Representing a resolution parameter; And Representing nodes respectively And Degree of (3); As a Kronecker delta function when And (3) with When belonging to the same sub-field, Take the value 1, otherwise Taking a value of 0; according to the corrected module degree evaluation index And determining the forwarding sub-domain.
6. The method for power network oriented quality of service aware multi-objective load balancing according to claim 1, wherein said constructing a plurality of forwarding paths with different performance characteristics using hierarchical path planning policies according to the forwarding sub-domains comprises: Acquiring performance indexes of the forwarding subdomains, wherein the performance indexes comprise an aggregation bandwidth, an aggregation time delay index and an aggregation reliability index; According to the performance index, determining the bandwidth weight, the time delay weight and the reliability weight of the forwarding subdomain; And calculating each logic path by using a Dijkstra algorithm according to the bandwidth weight, the time delay weight and the reliability weight, and determining the forwarding path according to a calculation result.
7. The method of claim 1, wherein the multi-objective optimization policy model comprises a multi-objective rewards function, the multi-objective rewards function being: ; ; ; ; In the formula, Indicating time of day Is a comprehensive prize value for (1); 、、 weight coefficients respectively representing performance rewards, qoS matching rewards and load balancing rewards; representing a performance rewards component; Representing a QoS matching reward component; representing a load balancing rewards component; Representing a path Upper first A plurality of forwarding nodes; Indicating the type of performance index (performance index), Respectively representing bandwidth, time delay, queue length and packet loss rate; Representing performance index Importance weights of (2); Representing nodes At the index Current performance value on; Indicating index Is the minimum of (2); Indicating index Is the maximum value of (2); representing traffic flows At the node The flow rate of the flow rate; representing traffic flows Is a total flow of (1); Representing a set of paths A matching degree function with the QoS requirement of the service flow; representing candidate path sets, subscripts Representing QoS matching correlations; a minimum threshold representing a load level; A target threshold value representing a load level; Is the moment of time Is a full network load level indicator; Representing a set of paths Load balancing function of (a) subscript Indicating load balancing correlations.
8. The method of power network oriented quality of service aware multi-objective load balancing of claim 1, wherein the path selection policy comprises a scheduling granularity factor, the scheduling granularity factor being obtained by the following formula: ; ; In the formula, Representing a scheduling granularity factor at time t; , a learning rate or smoothing coefficient representing the granularity update; representing a scheduling granularity factor at time t-1; representing a baseline granularity factor; Represents the granularity adjustment amplitude; Indicating an out-of-order prize; Representing a set of paths Is a disorder control factor of (a); Representing an out-of-order penalty coefficient; Shannon entropy representing flow distribution ratio distribution; Representing allocation to paths The flow ratio of (2) satisfies ; Representing the maximum possible entropy value; A target threshold value representing a load level; Indicating time of day The whole network load level index of (2) is used for quantifying the whole utilization degree of the current network resource.
9. The method for power network oriented quality of service aware multi-objective load balancing of claim 1, wherein the path selection policy comprises an allocation proportion and a scheduling granularity factor; after mapping the scheduling state space to a path selection policy, the method further comprises: Classifying the service flow data into corresponding target flow categories according to the service quality types; And distributing the service flow data to each forwarding path according to the distribution proportion and by utilizing a target decision register corresponding to the target flow class, and controlling the dispatching according to the dispatching granularity factor.
10. A computing network oriented quality of service aware multi-objective load balancing system, the system comprising: The traffic prediction module is used for acquiring service flow data, and predicting the service quality of the service flow data to obtain a service quality type; the diagram embedding acquisition module is used for acquiring a diagram embedding representation of the network state from a pre-constructed diagram structure, wherein the diagram structure is constructed based on the network state information of each node; The path planning module is used for determining a plurality of forwarding sub-domains with consistent performance according to the graph embedded representation, and constructing a plurality of forwarding paths with different performance characteristics by using a hierarchical path planning strategy according to the forwarding sub-domains; and mapping the scheduling state space into a path selection strategy by utilizing a pre-constructed multi-objective optimization strategy model, wherein the multi-objective optimization strategy model is a model based on a deep reinforcement learning algorithm.

Description

Service quality perception multi-target load balancing method and system for computing network Technical Field The invention relates to the technical field of communication, in particular to a service quality perception multi-target load balancing method and system for a computing network. Background Along with the continuous improvement of the parameter scale of the large-scale artificial intelligent model, the demand of large-scale model training and reasoning on computational power resources is exponentially increased, and a single data center has difficulty in bearing ultra-large-scale calculation tasks. In conventional resource scheduling modes, computing tasks are typically focused on unified scheduling in a single or a few data centers, which face bottlenecks in three dimensions of computing power, storage capacity, and network bandwidth. Especially in the scene of computational power scheduling which relates to cross-domain and multi-land coordination, the traditional method can not carry out global optimization according to the multi-dimensional resource condition and service requirements, and the problems of unbalanced resource allocation, long task waiting time, low overall processing efficiency and the like are very easy to cause. In order to alleviate the problems of resource fragmentation, uneven computational load and the like, a computational power network is proposed, and aims to connect a geographically distributed data center through a network to form a resource pool, and realize dynamic coordination of computational power scheduling in a 'network as a service' mode. However, in practical applications, the traffic flow types carried in the power network are complex, and QoS requirements exhibit a high degree of differentiation. For example, gradient sync tasks require high bandwidth to speed training convergence, micro-service communication focuses on low latency to guarantee service interaction efficiency, while model backup tasks focus on high reliability to prevent data loss. In order to improve the utilization efficiency of link resources, a multi-path concurrent scheduling mode is commonly adopted in a computing network, and a large flow is split into a plurality of sub-flows and distributed on a plurality of paths to realize bandwidth aggregation. However, at present, rule-based flow classification mechanisms have difficulty in accurately identifying all critical traffic flows, and simple prioritization cannot meet multidimensional QoS requirements at the same time. In addition, although the bandwidth reservation mechanism can guarantee QoS, the resource utilization rate is reduced, and effective adaptation to dynamic network conditions is lacking. In addition, qoS aware multipath routing mechanisms are prone to severe packet out-of-order problems in heterogeneous networks. In summary, a multi-path scheduling method supporting the capabilities of real-time QoS identification of traffic, performance aware path construction, scheduling granularity self-adaptation and multi-objective joint optimization is needed. Disclosure of Invention The invention aims to overcome the defects of the prior art and provide a service quality perception multi-objective load balancing method and a service quality perception multi-objective load balancing system for a computing network, so as to support the real-time QoS identification of flow, the construction of performance perception paths, the self-adaption of dispatching granularity and multi-objective joint optimization capability, and improve the resource dispatching efficiency and service guarantee capability of the computing network under the high concurrency of multiple services. In order to achieve the above purpose, the present invention adopts the following technical scheme. In a first aspect, the present invention provides a method for balancing service quality perception multi-objective load for a computing network, including: Acquiring service flow data, and carrying out service quality prediction on the service flow data to obtain a service quality type; Obtaining a graph embedded representation of a network state from a pre-constructed graph structure, wherein the graph structure is constructed based on network state information of each node; determining a plurality of forwarding sub-domains with consistent performance according to the graph embedded representation, and constructing a plurality of forwarding paths with different performance characteristics by using a hierarchical path planning strategy according to the forwarding sub-domains; And mapping the scheduling state space into a path selection strategy by utilizing a pre-constructed multi-objective optimization strategy model, wherein the multi-objective optimization strategy model is a model based on a deep reinforcement learning algorithm. In some embodiments of the present invention, the service flow data includes a first packet of a service flow; the predicting the service q