Non-homogeneous information joint cognition and scheduling transmission method

Abstract

The invention provides a non-homogeneous information joint cognition and scheduling transmission method, and the method comprises the steps: enabling an optical converter to be integrated into a DCN topological structure, and endowing a network with the intrinsic capability of carrying out the self-adjustment according to a traffic mode; compared with an existing layering or expansion graph scheme, the method has the capability of switching the network topology more flexibly, and switching can be completed even in the process that the data center service is in progress.

Description

Technical Field

[0001] The present invention belongs to the technical field of network communication and data processing, and in particular relates to a method for joint recognition and scheduling transmission of non-homogeneous information. Background Art

[0002] Data centers widely adopt topologies such as CLOS, BCube, and Xpander to manage their large device networks. Data center networks (DCNs) rely on traditional topologies, which are static in nature and lack flexibility to changes in traffic patterns. With the development of technology, researchers have recently designed dynamic DCN topologies based on multimodal converters, which provide an efficient solution without expanding existing communication resources. However, making full use of the heterogeneous information of DCN and realizing the full-network joint optimization of converters remains a major challenge in algorithm design. This patent uses graph representation learning algorithms for DCN topology optimization for the first time and proposes an innovative solution, MorpheusClos.

[0003] In recent years, in order to meet the growing needs of emerging applications, data center networks (DCNs) have developed rapidly. The new applications they carry, such as cloud computing and large-scale data learning, require more powerful network communication technology support. These emerging applications usually require more intensive data exchange and parallel processing services. It is worth noting that these services usually change over time. On the other hand, traditional DCNs rely on electrical connections to coordinate switch devices to establish a fixed topology. However, the fixed topology cannot meet the needs of emerging applications for topology awareness and cannot respond to changing traffic patterns in the network. As the latest research results, optical switch converters with adjustable connections provide a pioneering solution for design, which can build a network topology that truly adapts to needs.

[0004] Unlike traditional electrical connections, these converters provide a more flexible solution for switch connections. Figure 1The top half of the figure shows three modes of a 6-port switch that implements port switching through its physical optical splitter. Before the advent of these switches, the development of DCN topology design faced two conflicting goals: ease of deployment and good performance. CLOS, with its tree-like structure based on device clusters (Pods), has long been considered the gold standard for DCN topology design. Its advantages are that the system can be deployed using simple rules and it has inherent scalability when additional servers are introduced. However, the multi-layer structure of CLOS increases the number of routing hops between any two servers, thereby reducing network efficiency in terms of throughput and latency. In contrast, academia has introduced flatter extended graph architectures such as Jellyfish and Xpander, which reduce the number of routing hops between servers by randomly interconnecting them. A recent study showed that extended architectures can achieve similar throughput performance to CLOS while using 25% fewer switches. However, these flat structures are also more complex. Worse, the extended graph framework introduces the possibility of forming routing loops in its topology, which can cause priority-based flow control (PFC) deadlocks. Therefore, when considering the performance benefits of extended graphs, the risk of PFC deadlocks must also be considered.

[0005] Considering these conflicting aspects, many researchers have strived to find a balance between the rigid hierarchical structure of CLOS-like architectures and the randomly hierarchical flat configuration of extended graph-type designs. Important works in this area include Diamond, SlimFly, and FatClique, but a widely accepted general solution has not yet been found. This is because time-varying traffic patterns have their preferred network topology, and it is difficult to achieve a dynamic balance between hierarchical and flat solutions. As mentioned earlier, the use of adjustable switches presents a potential solution. Figure 1 The lower half of the figure shows the CLOS architecture before and after the integration of the converter. The dotted lines represent possible connections, but due to the inherent limitations of physical optical splitters, the optical converter cannot be connected to all ports at the same time. These converters replace electrical connection lines in the network, and the goal of this patent is to enable the DCN topology to dynamically adapt to network traffic patterns, allowing the network's topology to be reshaped in real time based on traffic demand. However, a key question remains: How do these converters adapt to changing network traffic? As data volumes surge and networks expand, the ability of these converters to accurately read and respond to various types of traffic becomes critical. Summary of the invention

[0006] In order to solve the above problems, the present invention provides a method for joint recognition and scheduling transmission of non-homogeneous information, which can switch network topology more flexibly.

[0007] A method for joint recognition and scheduling transmission of non-homogeneous information, wherein traffic is transmitted from a sending server to a receiving server via a meta-path composed of switches located at different layers of a heterogeneous data center network, wherein the switches on the heterogeneous data center network are divided into core layer switches, aggregation layer switches, and edge layer switches, and adjacent switches on the meta-path are connected through converters, and the number of converters is the smaller value of the number of aggregation layer switches and edge layer switches. At the same time, different working modes of the converters correspond to different switch connection modes, and the working mode of the converter changes with the traffic transmission demand of the heterogeneous data center network. When it is necessary to change the traffic transmission meta-path of the heterogeneous data center network in the next time slot, the method for obtaining the working mode of each converter in the next time slot is as follows:

[0008] S1: Obtaining status information of a heterogeneous data center network in a current time slot, wherein the heterogeneous data center network is composed of multiple converters, multiple switches, and multiple servers, and the status information includes service control information of each converter and topological location information in a current time slot, service control information of each server and topological location information in a current time slot, and topological location information and dynamic operation information of each switch in a current time slot;

[0009] S2: Input the state information into the trained decision system to obtain the optimal working mode of each converter in the next time slot.

[0010] Furthermore, in step S2, different working modes of switches, servers and converters in the heterogeneous data center network are used as nodes of the heterogeneous data center network topology; meanwhile, the decision system includes a plurality of cascaded Dyta modules, multi-layer perceptrons MLP, and PUCB models, wherein each Dyta module is composed of a Dyta layer and a causal convolutional layer;

[0011] Except for the Dyta layer of the first-level Dyta module, the Dyta layers of the remaining Dyta modules are used to process the first node feature information stored in the form of a heterogeneous data center network topology structure obtained by the previous-level Dyta module based on the HDCN mutual attention mechanism to obtain the second node feature information; the causal convolution layer is used to fuse the second node feature information obtained by the Dyta layer of the same Dyta module with the second node characteristic information of different time slots in history to obtain the first node feature information used to describe the causal relationship between topology scheduling and network performance; wherein the Dyta layer of the first-level Dyta module is used to convert the state information into the second node feature information based on the HDCN mutual attention mechanism, and the first node feature information obtained by the causal convolution layer of the last-level Dyta module is output to the multi-layer perceptron MLP;

[0012] The multi-layer perceptron MLP is used to obtain the prior probability that various working modes of each converter in the next time slot are the optimal working mode according to the first node feature information output by the previous Dyta module;

[0013] The PUCB model is used to calculate the corresponding confidence upper bounds of each converter when it is in various working modes in the next time slot based on the corresponding prior probabilities of each converter when it is in various working modes in the next time slot and the remaining link capacity percentage of each converter in the previous time slot, and use the working mode with the maximum confidence upper bound value among multiple modes of each converter as the working mode of each converter in the next time slot.

[0014] Furthermore, the second node feature information obtained by the Dyta layer of any level Dyta module is composed of the target node output features corresponding to each node in the network, and the method for obtaining the target node output features corresponding to any network node is:

[0015] Take each network node as the target node and perform the following steps to obtain the updated target node output features corresponding to each network node:

[0016] S21: Get the target node v t The adjacent nodes and the network nodes with similar traffic patterns in the previous time slot are taken as source nodes v s ;

[0017] S22: Each source node v s The service control information and topological location information of the previous time slot are used as the original features of the source node, and the original features of the source node are reduced to obtain the source node input features H l-1 [v s ], where l represents the source node v s The hierarchical number of the Dyta module where it is located; the source node input feature of the first-level Dyta module is defined as H 0 [v s ];

[0018] S23: Set each target node v t The service control information and topological location information of the previous time slot are used as the original features of the target node, and the original features of each target node are reduced to obtain the input features of the target node H l-1 [v t ]; the target node input feature of the first-level Dyta module is defined as H 0 [v t ];

[0019] S24: Input each target node into the feature H through the query vector linear projection function Q-linear l-1 [v t] are mapped to multiple vector spaces respectively, and the query vector mapping Q of each vector space is obtained h (v t ):

[0020] Q h (v t )=Q-linear h (H (l-1) [v t ])

[0021] Wherein, the vector space serial number h=1,2,...,H, H represents the number of vector spaces;

[0022] S25: Input the source node into the feature H through the key vector linear projection function K-linear l-1 [v s ] is mapped to multiple vector spaces, and the key vector mapping K of each vector space is obtained h (v s ):

[0023] K h (v s )=K-linear h (H (l-1) [v s ])

[0024] S26: Input the source node into the feature H through the value vector linear projection function V-linear l [v s ] is mapped to multiple vector spaces, and the value vector mapping V of each vector space is obtained h (v s ):

[0025] V h (v s )=V-linear h (H (l-1) [v s ])

[0026] S26: Map K according to the key vector h (v s ) and query vector mapping Q h (v t ) Get the source node v s For each target node v t The importance of Attention(v s ,v t ):

[0027]

[0028] Among them, ATT_headh (v s ,v t ) represents any target node v in the hth vector space t and source node v s The attention weight between is calculated as:

[0029]

[0030] Where d represents the target node input feature H l-1 [v t ] dimension, Indicates the capture source node v s and the target node v t The transformation matrix of the characteristic relationship between Represents the source node v s and the target node v t The attention adaptation scaling factor between , T represents transposition;

[0031] S27: Using linear projection Map the value vector V h (v s ) is converted to a value message ValueMessage(v s ,v t ):

[0032]

[0033] Among them, MSG-head h (v s ,v t ) indicates that the transformation matrix is ​​set Mapping V with value vector h (v s ) after multiplying, Indicates that all vector spaces correspond to the MSG-head h (v s ,v t ) added;

[0034] S28: The importance of Attention (v s ,v t ) and value message ValueMessage(v s ,v t ) to obtain each target node v t Message vector

[0035]

[0036] Among them, SN(t) represents the set of source nodes selected in the current time slot t, represents the product Attention(v) corresponding to the node pairs consisting of all source nodes and their paired target nodes in the set SN(t) s ,v t )·ValueMessage(v s ,v t ) added;

[0037] S29: Get the target node v according to the nonlinear activation ReLu function t Output feature H at the target node output by the current level Dyta module (l) [t]:

[0038]

[0039] Among them, σ is the weight of the nonlinear activation ReLu function, H (l-1) [t] is the target node v t Output features at the target node output by the previous level Dyta module.

[0040] Furthermore, the method for the causal convolution layer of any level Dyta module to obtain the feature information of the first node is as follows:

[0041]

[0042] in, is the output feature of the dth causal convolution of the current causal convolution layer at the current time slot t, is the first node feature information finally obtained by the current causal convolution layer at the current time slot t, D is the expansion factor used to characterize the number of layers in the causal convolution layer, K is the number of convolutions performed in each layer of causal convolution, is the weight of the causal convolution layer of the causal convolution layer in the current level l Dyta module when performing the i-th convolution, G t-i×d represents the graph data within the expansion time step, wherein the method for obtaining the graph data is: taking the different working modes of each switch, each server and converter as a node respectively, thereby forming a topological graph; taking the state information of the K×D topological graphs of the historical K×D time slots including the current time slot t in the current l-th level Dyta module as the second node feature information obtained as the graph data; σ represents the activation function, and ReLU represents the linear rectification function.

[0043] Furthermore, the method for determining the optimal working mode of each converter in the next time slot is as follows:

[0044] A multi-level search tree of the PUCB model is constructed, wherein the last level leaf node of the search tree corresponds to the next time slot, and each level node of the remaining nodes corresponds to a historical time slot, and each node represents a converter working mode and its related statistical information, wherein the statistical information includes the number of times selected as the optimal working mode, the percentage of the remaining link capacity in the previous time slot, and the prior probability; at the same time, except for the root node and the leaf node, the remaining nodes are both parent nodes and child nodes;

[0045] The search tree is traversed according to the set search rule, and the confidence upper bounds of the converter working modes corresponding to the nodes at each level are calculated level by level, wherein the set search rule is that the nodes with fewer selection times are selected first, and the nodes with larger confidence upper bounds are selected first; at the same time, the confidence upper bound of the converter working mode corresponding to any node is obtained as follows:

[0046]

[0047] Among them, U(v,t) represents the confidence upper bound of the converter working mode corresponding to the current node v in the time slot t corresponding to its own level, N(v,t) is the number of times the converter working mode corresponding to the current node v is selected as the optimal working mode before the end of the time slot t corresponding to its own level, P(v,t) is the highest priority probability of the converter working mode corresponding to the current node v in the time slot t corresponding to its own level, c puct is a setting constant that balances exploration and exploitation;

[0048] For each parent node, select the child node with the largest Q(v,t)+U(v,t) from each child node as the optimal child node:

[0049] Best child node ← argmax v (Q(v,t)+U(v,t))

[0050] Where Q(v,t) is the percentage of the remaining link capacity of the converter working mode corresponding to the current node v in the previous time slot of the time slot t corresponding to its own level;

[0051] The newly selected optimal child node is used as the new parent node, and the number of times the converter working mode corresponding to all nodes of the search tree is selected as the optimal working mode is updated. Then the confidence upper bounds of each child node of the selected new parent node are calculated, and the updated parent node is selected according to the confidence upper bound; and so on, the confidence upper bound of the last-level leaf node is obtained, and the converter working mode corresponding to the leaf node with the largest confidence upper bound is used as the optimal working mode of the current converter in the next time slot.

[0052] Beneficial effects:

[0053] 1. The present invention provides a method for joint recognition and scheduling transmission of non-homogeneous information. By integrating optical converters into the DCN topology structure, the network is given the inherent ability to self-adjust according to the traffic pattern. Compared with the existing layered or extended graph solutions, the present invention has the ability to switch network topology more flexibly, even when the data center service is in progress.

[0054] 2. The present invention provides a method for joint recognition and scheduling transmission of non-homogeneous information, which cleverly integrates graph-based learning methods into DCN topology design and explores new possibilities for adapting to the dynamic characteristics of data center networks. The learning method can recognize non-homogeneous network information and make real-time scheduling decisions. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 A comparison diagram of the traditional CLOS topology and the converter enhanced integration;

[0056] Figure 2 An example of an extended graph architecture and a PFC deadlock caused by CBD;

[0057] Figure 3 for network node information and its cognitive representation;

[0058] Figure 4 Schematic diagram of the overall framework of network cognition and switch scheduling decision process in MorpheusClos. DETAILED DESCRIPTION

[0059] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application.

[0060] To address the above challenges, we propose MorpheusClos, a name inspired by the change and adaptability of Morpheus in ancient Greek mythology. This innovative approach integrates converters into the existing CLOS architecture and uses fine-grained configuration strategies to improve DCN performance without the need for additional bandwidth resources. The core of MorpheusClos lies in its traffic pattern analysis method, which takes into account the topology of the network. Initially, we considered the configuration problem as a Steiner tree problem. However, given the complexity of directly solving this NP-complete problem, we developed a learnable framework that brings the set of converters as close to the optimal solution as possible. The core of the framework is a temporal self-supervised method designed to handle the dynamic characteristics of DCN. To the best of our knowledge, MorpheusClos is the first solution to explore the application of converters based on graph representation learning in DCN topology optimization.

[0061] The first contribution of this invention is the introduction of an innovative DCN topology that combines the converter with the existing CLOS structure. The design leverages the dynamic connection mode of the converter to achieve diverse cross-Pod connections and facilitate interconnection between different switch layers. By enhancing the adaptability of the CLOS network, it significantly improves the system's resilience to sudden traffic surges and enables the physical topology to intuitively adapt to network traffic patterns.

[0062] The second contribution of this paper is to introduce a graph representation learning framework designed specifically for DCN. Compared with traditional graph representation learning methods, this framework shows higher ability in handling dynamic heterogeneous graph structures. By combining the temporal self-supervised Transformer with the hierarchical aggregation of heterogeneous information, our model can effectively interpret and manage heterogeneous dynamic traffic information among multiple nodes.

[0063] Specifically, by integrating optical converters into the DCN topology, the present invention endows the network with the inherent ability to self-adjust according to traffic patterns. This section first explores the root causes of the performance differences between the traditional CLOS architecture and the extended graph when handling different traffic patterns, and emphasizes the key role of the adaptive topology facilitated by optical converters in resolving these differences.

[0064] The hierarchical tree structure of the CLOS architecture provides redundancy and fault tolerance through multiple paths, providing strong fault tolerance. However, its hierarchical routing design brings more routing hops and higher transmission delay.

[0065] Extended graph architectures, such as Jellyfish, Xpander, and SlimFly, are being explored as alternatives to CLOS. With their flat topology, they offer advantages in bandwidth and latency. Figure 2 As shown, using a four-port switch, the extended graph reduces the average number of hops for communication between servers through flat random interconnection instead of a hierarchical structure, thereby improving the speed of data transmission between servers compared to CLOS.

[0066] To compare the performance of CLOS and Extended Graph DCN architectures under different traffic scenarios, we created different traffic models using scripts of the iperf tool on the server. Our tests were conducted on a 40GbE data center with 100 switches and 1400 servers, focusing on two main traffic patterns:

[0067] 1. All to All: Each server communicates with all other servers, simulating high load and intensive data exchange to test network capacity and fault tolerance.

[0068] 2. Uniform Random: The server randomly selects the communication target, similar to typical DCN traffic, allowing realistic performance evaluation.

[0069] We summarize the performance of CLOS and Extended Graph topologies under different traffic and load conditions, highlighting their respective strengths. CLOS excels in “all-to-all” scenarios due to its redundant paths, making it suitable for high-load situations. On the other hand, the Extended Graph topology shows lower latency in simple traffic scenarios, but this advantage diminishes under higher loads, where CLOS performs better in terms of throughput. This shows that while Extended Graph is better suited for typical traffic, CLOS excels under high traffic demands. These results lead to a key conclusion: no single DCN topology is universally superior. The choice of topology should be closely related to the specific traffic patterns and load conditions expected in the network environment.

[0070] Case analysis of deadlock caused by existing PFC technology

[0071] Given the performance differences between different topologies, especially the degradation of scaling graph performance under high load, it is crucial to examine the impact of the upper layer protocols of DCN (the flow control algorithm at the transport layer and the routing algorithm at the network layer).

[0072] Data center operators such as Microsoft and Alibaba are increasingly adopting Remote Direct Memory Access (RDMA) to improve network performance. A key component of RDMA is the Priority Flow Control (PFC) algorithm, a flow control method designed to prevent packet loss in high-speed data networks. PFC relieves congestion by pausing specific data flows, ensuring that critical packets are not dropped.

[0073] Figure 2 An example of PFC deadlock introduced in the extended graph architecture is shown. The ring buffer dependency phenomenon (CBD) is a loop formed by the connections between value devices, which is avoided in the tree topology but widely exists in the flat extended graph network framework. Figure 2 The following figure shows a CBD example in a network framework consisting of switches A, B, and C (connected by yellow lines). Although routing algorithms such as ECMP and KSP can prevent a single flow from forming a loop at the routing level, loop deadlocks may still occur in some special cases. For example, in Figure 2In the figure, data flow 1 is transmitted through C→B→A (red dotted line), and the other flow is transmitted through A→C (blue dotted line), forming a CBD loop of three switches. When the real-time data inventory in the buffer reaches the PFC threshold, the CBD loop problem will become serious and may cause deadlock. Deadlock refers to the situation where all node devices in the CBD loop trigger the PFC threshold at the same time and suspend forwarding data. The deadlocked switch cannot recover itself and becomes an invalid node, thereby reducing the overall transmission efficiency of the DCN. In contrast, the CLOS architecture avoids the generation of CBD loops due to its tree structure and hierarchical "up and down" routing architecture, fundamentally preventing PFC deadlock and making it better at handling high-load traffic.

[0074] The above analysis shows that when running common upper layer protocols, the extended graph architecture may experience deadlock in its CBD loop structure when facing high-load traffic patterns. However, the occurrence of CBD in the extended graph is not inevitable, which is related to the traffic pattern at the time. And the existence of CBD does not necessarily mean that PFC-induced deadlock will be triggered. In non-extreme traffic conditions, such as uniform random traffic patterns, CBD is not common. Even if CBD exists, deadlock will not occur if all device caches in it do not reach the PFC threshold. On the contrary, in the absence of deadlock, the extended graph has a performance advantage over the CLOS architecture. It is worth noting that the verification experiment we conducted adopted the commonly used ECMP protocol, and similar trends were observed when testing with other routing algorithms (including KSP, VLB, etc.). From these observations, we conclude that when using switches to change the network topology to connect similarly to the extended graph, the traffic pattern and the number of switches in the flat connection must be considered to minimize the risk of deadlock caused by PFC.

[0075] Further, the present invention will first briefly review the previous works on heterogeneous DCNs with switches.

[0076] Two paradigms of DCN topology design. In the DCN topology landscape, two different directions have emerged, each addressing specific network needs and challenges. The first direction is the traditional static network topology, represented by structures such as Fat-Tree, VL2, Diamond, and BCube. These architectures focus mainly on scalability. For example, Fat-Tree is a tree-like topology that optimizes uniform traffic distribution and high bidirectional bandwidth. VL2 creates a logically complete bipartite graph through the interconnection of intermediate switches and aggregation switches, which simplifies wiring compared to Fat-Tree. Diamond enhances Fat-Tree by reducing the average path length, while BCube provides a low-diameter, full-bandwidth architecture with multiple alternative paths.

[0077] Another direction is toward flat network architectures, such as Jellyfish, Xpander, and SlimFly, which represent a design that deviates from the traditional hierarchical structure. These topologies are based on extended graph theory and aim to create more flexible and cost-effective network designs. For example, Jellyfish uses a random graph topology to achieve a high level of connectivity and flexibility. Xpander provides a scalable network design based on this. SlimFly proposes a low-diameter, high-bandwidth network. This DCN topology design direction is characterized by the pursuit of enhanced network efficiency.

[0078] Dynamic network topologies. Modern DCNs increasingly add flexibility by introducing randomness into their static topologies. Key research, such as Jellyfish, addresses the challenge of incremental scaling by injecting randomness into network design. Switches, an emerging class of devices that emerged with the technological development of optical communications, bring dynamicity to network architectures through various configurations. Notable research, such as Flattree, demonstrates how switchable switches can transform traditional topologies into more random forms, thereby improving network performance. Dynamic architectures such as RotorNet and Opera embody this progress, with leaf-spine topologies with uplinks that are constantly reconfigured at the top-of-rack switches. However, the limitation of these “dynamic” approaches is that they rely solely on randomness and lack the ability to generate insights, analyze, and adapt to changing traffic patterns.

[0079] Graph-based learning methods. The development of graph embedding methods has opened up important opportunities for graph analysis. These methods have facilitated the understanding of complex graph structures by simplifying their representation. However, real-world networks are often composed of multimodal and multi-type nodes and relationships, which poses challenges in capturing their diverse characteristics. Recently, the focus in the field of graph representation learning has shifted to the design of graph transformers. These models have shown remarkable results in various graph representation learning tasks by leveraging fully connected global attention mechanisms. However, most existing graph transformers are mainly targeted at clustering and prediction based on existing data, making them less adept at handling dynamic real-world networks that are constantly evolving, such as DCNs. This poses a challenge to adapt these powerful models to networks that are not only complex but also constantly changing. In this evolving field, MorpheusClos attempts to cleverly integrate graph-based learning methods into DCN topology design and explore new possibilities for adapting to the dynamic characteristics of data center networks.

[0080] Specifically, a method for joint recognition and scheduling transmission of non-homogeneous information is provided, wherein traffic is transmitted from a sending server to a receiving server via a meta-path composed of switches located at different layers of a heterogeneous data center network, wherein the switches on the heterogeneous data center network are divided into core layer switches, aggregation layer switches, and edge layer switches, and adjacent switches on the meta-path are connected through converters, and the number of converters is the smaller value of the number of aggregation layer switches and edge layer switches. At the same time, different working modes of the converters correspond to different switch connection methods, and the working mode of the converter changes with the traffic transmission demand of the heterogeneous data center network. When it is necessary to change the traffic transmission meta-path of the heterogeneous data center network in the next time slot, the method for obtaining the working mode of each converter in the next time slot is as follows:

[0081] S1: Obtaining status information of a heterogeneous data center network in a current time slot, wherein the heterogeneous data center network is composed of multiple converters, multiple switches, and multiple servers, and the status information includes service control information of each converter and topological location information in a current time slot, service control information of each server and topological location information in a current time slot, and topological location information and dynamic operation information of each switch in a current time slot;

[0082] It should be noted that if Figure 3 As shown in the figure, the category of topological location information includes the connection and location information of DCN nodes, such as the core layer (C), aggregation layer (A), edge layer (E), and server (S). This information is represented in the form of adjacency matrix and node attributes, capturing the complex relationships between nodes and dynamically reflecting the topological updates within the MorpheusClos network. Dynamic operational information: defined as time-sensitive information of switching nodes in the DCN that cannot be determined in advance. This information is critical for real-time operational decisions and adaptive responses to network conditions. 3. Service and control information: includes pre-obtainable information related to servers, such as protocol type, service type, and function type, as well as control information of converters. These attributes play a guiding role in the intelligence of the network.

[0083] The strategy of exploiting classified information is crucial in addressing the heterogeneity of network information. It is worth noting that traditional graph-based learning methods excel in deriving embeddings from static node features, but generally have difficulty extracting insights from dynamic information. Therefore, our approach is specifically designed to manage and exploit the dynamic behavior of DCNs in the MorpheusClos framework, which is important in terms of adaptability and real-time responsiveness.

[0084] A metapath represents a series of nodes that describes the behavior of data transmission. It includes a sending server, a receiving server, and a series of intermediate nodes determined by the routing algorithm. In traditional graph learning, information propagation is usually performed in units of a single hop. The concept of a metapath provides a higher-level perspective that covers multiple hops in a single unified path. This higher-level abstraction helps to more easily understand traffic events in the network because it encapsulates the entire data transmission process from the sender to the receiver.

[0085] S2: Input the state information into the trained decision system to obtain the optimal working mode of each converter in the next time slot.

[0086] It should be noted that the heterogeneous data center network (HDCN) is modeled as a dynamic undirected graph Where V represents the switches, converters (that is, Figure 1 b) and servers. The set of connected edges between these elements is denoted as E. The adjacency matrix It can be dynamically allocated according to the real-time network conditions. At any time step t, the state of the network is determined by the graph signal , where f represents the number of node-specific features.

[0087] In MorpheusClos, the converter dynamically adjusts the network topology without requiring additional resources in the traditional CLOS architecture. A Aggregation layer switches and Num E edge layer switches, we define Num p =min(Num A ,Num E ). It represents the aggregation by placing a switch between p pairs of switches. Figure 4 As shown in the upper left corner, the default mode of a single switch in an AE switch pair is to replace the connection to the core layer switches and servers with the switch. The remaining two ports of each switch are randomly connected to other switches in the network. The internal connections of the switch are reconfigurable, enabling cross-layer connections, such as between edge and core switches or between switches in different Pods. For a MorpheusClos network with p switches, the state of each switch can be represented as a two-dimensional matrix M:

[0088]

[0089] Each entry M i,j represents the state of the i-th converter in its j-th mode. The default mode of each converter is denoted as M i,1,When all converters are in default mode, the DCN topology is equivalent to a hierarchical CLOS architecture.

[0090] Furthermore, in step S2, different working modes of switches, servers and converters in the heterogeneous data center network are used as nodes of the heterogeneous data center network topology; meanwhile, the decision system includes a plurality of cascaded Dyta modules, multi-layer perceptrons MLP, and PUCB models, wherein each Dyta module is composed of a Dyta layer and a causal convolutional layer;

[0091] Except for the Dyta layer of the first-level Dyta module, the Dyta layers of the remaining Dyta modules are used to process the first node feature information stored in the form of a heterogeneous data center network topology structure obtained by the previous-level Dyta module based on the HDCN mutual attention mechanism to obtain the second node feature information; the causal convolution layer is used to fuse the second node feature information obtained by the Dyta layer of the same Dyta module with the second node characteristic information of different time slots in history to obtain the first node feature information used to describe the causal relationship between topology scheduling and network performance; wherein the Dyta layer of the first-level Dyta module is used to convert the state information into the second node feature information based on the HDCN mutual attention mechanism, and the first node feature information obtained by the causal convolution layer of the last-level Dyta module is output to the multi-layer perceptron MLP;

[0092] The multilayer perceptron MLP is used to obtain the prior probability that various working modes of each converter in the next time slot are the optimal working modes according to the received first node feature information;

[0093] The PUCB model is used to calculate the corresponding upper confidence bounds of each converter when it is in various working modes in the next time slot based on the corresponding prior probabilities of each converter when it is in various working modes in the next time slot and the remaining link capacity percentage of each converter in the previous time slot, and the working mode corresponding to the maximum value of the lower confidence upper bound of each converter is used as the working mode of each converter in the next time slot.

[0094] Furthermore, the second node feature information obtained by the Dyta layer of any level Dyta module is composed of the target node output features corresponding to each converter, and the method for obtaining the target node output features corresponding to any network node is:

[0095] Take each network node as the target node and perform the following steps to obtain the updated target node output features corresponding to each network node:

[0096] S21: Get the target node v t The adjacent nodes and the network nodes with similar traffic patterns in the previous time slot are taken as source nodes vs ;

[0097] S22: Each source node v s The service control information and topological location information of the previous time slot are used as the original features of the source node, and the original features of the source node are reduced to obtain the source node input features H l-1 [v s ], where l represents the source node v s The hierarchical number of the Dyta module where it is located; the source node input feature of the first-level Dyta module is defined as H 0 [v s ];

[0098] S23: Set the target node v t The service control information and topological location information of the previous time slot are used as the original features of the target node, and the original features of each target node are reduced to obtain the input features of the target node H l-1 [v t ]; the target node input feature of the first-level Dyta module is defined as H 0 [v t ];

[0099] S24: Input each target node into the feature H through the query vector linear projection function Q-linear l-1 [v t ] are mapped to multiple vector spaces respectively, and the query vector mapping Q of each vector space is obtained h (v t ):

[0100] Q h (v t )=Q-linear h (H (l-1) [v t ])

[0101] Wherein, the vector space serial number h=1,2,...,H, H represents the number of vector spaces;

[0102] S25: Input the source node into the feature H through the key vector linear projection function K-linear l-1 [v s ] is mapped to multiple vector spaces, and the key vector mapping K of each vector space is obtained h (v s ):

[0103] K h (v s )=K-linear h (H (l-1) [v s ])

[0104] S26: Input the source node into the feature H through the value vector linear projection function V-linear l [v s ] is mapped to multiple vector spaces, and the value vector mapping V of each vector space is obtained h (v s ):

[0105] V h (v s )=V-linear h (H (l-1) [v s ])

[0106] S26: Map K according to the key vector h (v s ) and query vector mapping Q h (v t ) Get the source node v s For each target node v t The importance of Attention(v s ,v t ):

[0107]

[0108] Among them, ATT_head h (v s ,v t ) represents any target node v in the hth vector space t and source node v s The attention weight between is calculated as:

[0109]

[0110] Where d represents the target node input feature H l-1 [v t ] dimension, Indicates the capture source node v s and the target node v t The transformation matrix of the characteristic relationship between Represents the source node v s and the target node v t The attention adaptation scaling factor between , T represents transposition;

[0111] S27: Using linear projection Map the value vector V h (v s ) is converted to a value message ValueMessage(v s ,v t ):

[0112]

[0113] Among them, MSG-head h (v s ,v t ) indicates that the transformation matrix is ​​set Mapping V with value vector h (v s ) after multiplying, Indicates that all vector spaces correspond to the MSG-head h (v s ,v t ) added;

[0114] S28: The importance of Attention (v s ,v t ) and value message ValueMessage(v s ,v t ) to obtain each target node v t Message vector

[0115]

[0116] Among them, SN(t) represents the set of source nodes selected in the current time slot t, represents the product Attention(v) corresponding to the node pairs consisting of all source nodes and their paired target nodes in the set SN(t) s ,v t )·ValueMessage(v s ,v t ) added;

[0117] S29: Get the target node v according to the nonlinear activation ReLu function t Output feature H at the target node output by the current level Dyta module (l) [t]:

[0118]

[0119] Among them, σ is the weight of the nonlinear activation ReLu function, H (l-1) [t] is the target node v t Output features at the target node output by the previous level Dyta module.

[0120] Furthermore, the method for the causal convolution layer of any level Dyta module to obtain the feature information of the first node is as follows:

[0121]

[0122] in, is the output feature of the dth causal convolution of the current causal convolution layer at the current time slot t, is the first node feature information finally obtained by the current causal convolution layer at the current time slot t, D is the expansion factor used to characterize the number of layers in the causal convolution layer, K is the number of convolutions performed in each layer of causal convolution, is the weight of the causal convolution layer of the causal convolution layer in the current level l Dyta module when performing the i-th convolution, G t-i×d represents the graph data within the expansion time step, wherein the method for obtaining the graph data is: taking the different working modes of each switch, each server and converter as a node respectively, thereby forming a topological graph; taking the state information of the K×D topological graphs of the historical K×D time slots including the current time slot t in the current l-th level Dyta module as the second node feature information obtained as the graph data; σ represents the activation function, and ReLU represents the linear rectification function.

[0123] Furthermore, the method for determining the optimal working mode of each converter in the next time slot is as follows:

[0124] A multi-level search tree of the PUCB model is constructed, wherein the last level leaf node of the search tree corresponds to the next time slot, and each level node of the remaining nodes corresponds to a historical time slot, and each node represents a converter working mode and its related statistical information, wherein the statistical information includes the number of times selected as the optimal working mode, the percentage of the remaining link capacity in the previous time slot, and the prior probability; at the same time, except for the root node and the leaf node, the remaining nodes are both parent nodes and child nodes;

[0125] The search tree is traversed according to the set search rule, and the confidence upper bounds of the converter working modes corresponding to the nodes at each level are calculated level by level, wherein the set search rule is that the nodes with fewer selection times are selected first, and the nodes with larger confidence upper bounds are selected first; at the same time, the confidence upper bound of the converter working mode corresponding to any node is obtained as follows:

[0126]

[0127] Among them, U(v,t) represents the confidence upper bound of the converter working mode corresponding to the current node v in the time slot t corresponding to its own level, N(v,t) is the number of times the converter working mode corresponding to the current node v is selected as the optimal working mode before the end of the time slot t corresponding to its own level, P(v,t) is the highest priority probability of the converter working mode corresponding to the current node v in the time slot t corresponding to its own level, c puct is a setting constant that balances exploration and exploitation;

[0128] For each parent node, select the child node with the largest Q(v,t)+U(v,t) from each child node as the optimal child node:

[0129] Best child node ← argmax v (Q(v,t)+U(v,t))

[0130] Where Q(v,t) is the percentage of the remaining link capacity of the converter working mode corresponding to the current node v in the previous time slot of the time slot t corresponding to its own level;

[0131] The newly selected optimal child node is used as the new parent node, and the number of times the converter working mode corresponding to all nodes of the search tree is selected as the optimal working mode is updated. Then the confidence upper bounds of each child node of the selected new parent node are calculated, and the updated parent node is selected according to the confidence upper bound; and so on, the confidence upper bound of the last-level leaf node is obtained, and the converter working mode corresponding to the leaf node with the largest confidence upper bound is used as the optimal working mode of the current converter in the next time slot.

[0132] The basic principle of the present invention is described in detail below.

[0133] 1. Problem Definition

[0134] The classic Steiner tree problem, which is considered NP-hard, is a basic combinatorial optimization problem. It is described as follows: Given a weighted graph G = (V, E) and a set of terminal nodes The goal is to find a minimum cost tree that contains all terminal nodes Te. Non-terminal nodes in the graph that do not belong to Te are called Steiner nodes.

[0135] In a specific scenario in the MorpheusClos framework, the switches in the network, denoted as V Te , is considered as the terminal node, and the various modes of the converter, denoted as V C =VV Te , are considered as potential Steiner nodes. The goal is to optimize the network configuration to minimize the maximum link utilization, which is a key indicator for load balancing in a dynamic and heterogeneous DCN environment.

[0136] 2. MorpheusClos framework definition

[0137] In this section, we detail our network optimization strategy within the MorpheusClos framework, focusing on the role of transformers. Our approach is based on a comprehensive analysis of DCN topologies and exploits the rich graph structure of these networks. The key benefits of this exploration include: (1) Graph-driven insights: We study the ability of graph learning methods to solve Steiner tree-like problems, leveraging the structural complexity of DCNs. (2) Improved accuracy: By leveraging the inherent heterogeneous information in DCN nodes and their connections, our approach goes beyond traditional decision-making processes based on raw data. (3) Adaptive online decision-making: Our strategy uses a weighted adjacency matrix to dynamically represent node relationships, enabling responsive and evolving decisions under real-time network conditions.

[0138] Figure 4 (a) shows the overall process of converter mode decision in the MorpheusClos framework.

[0139] First, we generate a preliminary adjacency matrix to represent the topological localization information of the graph. Node attributes and most of the service and control information are one-hot encoded to form the initial attributes. Since dynamic operation information is time-sensitive, it is not included in the node attributes. Instead, it is used to calculate node similarity scores and integrated with the adjacency matrix as weights. Formally, we represent the HDCN data sequence for time period tT to t as follows:

[0140]

[0141] The adjacency matrix and graph information of these $T$ time periods are input into a network consisting of DyTa (Dynamic Traffic Attention) blocks. Subsequently, the embeddings processed by the DyTa layer and the causal convolutional layer are concatenated and aggregated along the time dimension to form a comprehensive representation of the DCN node information and traffic patterns at time $t$:

[0142]

[0143] The next step is to Through a multi-layer perceptron (MLP) layer, the effective prior probability of each converter mode node is obtained:

[0144]

[0145] These probabilities are transformed into the converter configuration for the next time period t+1 through the Probabilistic Update Confidence Bounds (PUCB) model. We adopt an iterative update scheme for adaptive decision making, treating the HDCN atlas as a FIFO stack with depth T, which is updated after the topology is adjusted. This allows the converter decision to be executed cyclically during the framework operation.

[0146] 2.1 Enhanced Graph Analysis Using Real-time Network Data

[0147] Dynamic weighting of the adjacency matrix: In a dynamic network environment, nodes continuously generate operational information with significant time sensitivity. Traditional graph-based learning models have difficulty in effectively processing this information. To ensure that traffic patterns and other operational data effectively guide decision-making, we incorporate these elements into the adjacency matrix of HDCN. For each terminal node, we consider its operational information attribute sequence (o t-T,n ,…,o t,n ), from (O t-T ,…,O t ) creates an overall representation of the node's dynamic operation information attributes:

[0148]

[0149] u n Represents the aggregate operation information of the nth terminal node (the switch is the terminal), based on the derived aggregate weight p τ,n Here, τ indexes the time period (tT,t). The aggregation weight p τ,n Reflects the traffic pattern of a specific node (o τ,n ) and the overall traffic pattern (u n ). The correlation between is a learnable parameter vector for the transformation, where D represents the dimension of the terminal node attributes.

[0150] We propose to estimate the degree of heterogeneity between nodes, which reflects the variability of their traffic distribution over time. This can be mathematically expressed as follows:

[0151]

[0152] Here, the higher q m,n indicates that the traffic patterns between nodes m and n are more similar. This metric is integrated as a weight in the adjacency matrix of graph G to form a weighted adjacency matrix

[0153] 3 Joint Cognition via Heterogeneous Graph Transformers

[0154] We exploit the structural properties of the graph to compute mutual attention between nodes and facilitate information transfer. This approach makes it possible to learn embeddings of transformer nodes in the context of the network.

[0155] 3.1DyTa layer

[0156] like Figure 4 As shown in (b), the DyTa layer is a key component of the MorpheusClos framework, and its unique ability lies in generating attention masks for heterogeneous information during message passing.

[0157] Adaptive Node Selection: In DCN data, traffic often travels through multiple nodes, resulting in long meta-paths. Effectively capturing these extended relationships is a challenge for graph-based learning tools. To address this issue, we introduce a node selection module before the message passing process. This module selects the source node for message passing based on two criteria: physical connectivity with the current node and presence of a node in the weighted adjacency matrix. The mathematical formula of the node selection process is as follows:

[0158]

[0159] The selection criteria for message passing nodes are based on two key aspects. First, N(i) represents the set of physical neighbor nodes of node i, ensuring that local structural connections play a role in the message propagation process. Second, the inclusion of node k is determined by its weighted adjacency matrix The high operational similarity with node i in is determined. In this matrix, and denote the mean and standard deviation of the similarity scores of the i-th row, respectively. Nodes with similarity scores exceeding two standard deviations from the mean are considered to be significantly similar to node i. This threshold is chosen to capture nodes related to operations, ensuring that only nodes with significant similarity in traffic patterns to node i are included.

[0160] HDCN mutual attention mechanism: The DyTa layer aims to update the embedding of the target node. Figure 4 As shown in (b), the embedding of the target node is converted into a query vector (Q) through a linear layer, while the embedding of the selected node as the source node is converted into a key vector (K) and a value vector (V). This mechanism is different from the traditional transformer because it takes into account the heterogeneity of node information. Each pair of node relationships has its own unique projection weight, which is shared between pairs of nodes of the same type.

[0161] The embedding of network nodes in different DyTa layers is represented as H l , where l represents the lth DyTa layer. In the DyTa layer, the embedding of a node is affected by its selected node set SN. To describe this relationship, we refer to the nodes in the selected node set as source nodes v s , the affected node is called the target node v t . Target node v t and the embedding of the source node at layer l are denoted as H l [v t ] and H l [v s ]. In order to efficiently manage the information flow within the network, the DyTa layer adopts a multi-head attention mechanism. This mechanism t and vs The embedding of is divided into h heads, allowing the model to focus on information in different embedding subspaces in parallel. This parallel processing not only reduces computation time, but also allows a more detailed understanding of the dynamics of the network.

[0162] These node embeddings are mapped into the h-th vector space using three linear projection functions:

[0163] Q h (v t )=Q-linear h (H (l-1) [v t ])

[0164] K h (v s )=K-linear h (H (l-1) [v s ])

[0165] V h (v s )=V-linear h (H (l-1) [v s ])

[0166] To calculate the node v t and v s mutual attention between them, we introduce an attention operator to estimate each v s v t Importance:

[0167]

[0168] The attention function can be described as mapping a query vector and a set of key-value pairs to each pair of nodes (v s ,v t ) output. t and v s The overall attention of ATT_head is the concatenation of the attention weights of all heads, processed by a softmax function. h (v s ,v t ) represents the hth head v t and v s The attention weight between is calculated as follows:

[0169]

[0170] here, is a transformation matrix used to capture feature relationships, and d is the output H of the previous DyTa layer l-1 [vt ] dimension, as an adaptive scaling factor for attention. The last step involves concatenating the h attention heads to form an attention vector for each pair of nodes. Subsequently, for the target node t, the attention vectors of all selected nodes SN(t) are assembled and normalized by a softmax operation. This ensures that the sum of the attention weights of all nodes in SN(t) is equal to one, following the principle of attention distribution.

[0171] Message passing in the DyTa layer: Figure 4 As shown in (b), when calculating all node pairs (v s ,v t ), these weights are combined with the value vector (V) to update the target node v t The value message for each source-target node pair is defined as follows:

[0172]

[0173] To construct the i-th message header MSG-headi(s,t), the features of the source node s are projected to the i-th message vector using linear projection Then through the matrix Adjust the original V h (v s ) feature vector. The last step is to aggregate all h message headers to obtain the value vector for each pair of nodes (s, t). This method uses a method similar to that of generating attention weights, introducing the transformation matrix To balance the differences in feature distribution of heterogeneous nodes.

[0174] After computing the heterogeneous multi-head attention and value messages, we need to aggregate them from the source nodes to the target nodes (see Figure 4 (b)). Since the softmax process in the formula normalizes the sum of the attention vectors to one, we can simply use the attention vectors as weights to average the corresponding messages from the source nodes to update the vector as follows:

[0175]

[0176] This aggregates information from the selected set of nodes SN(t) to the target node v t We update the vector Apply the nonlinear activation ReLu function as follows:

[0177]

[0178] In this way, we get the output H of the target node t in the DyTa layer (l)[t]. By stacking modules containing DyTa layers, the information of each node is allowed to reach most nodes in the HDCN graph. Therefore, the DyTa layer generates a representation of each node that contains global traffic pattern features for downstream tasks that predict converter priors.

[0179] Throughout the DyTa layer architecture, we rely heavily on the weighted adjacency matrix Generate attention weights. This can be interpreted as leveraging the traffic patterns to jointly update the embeddings of multiple nodes to the target node. Compared with the standard graph transformer, our DyTa layer distinguishes the relationship weights between heterogeneous nodes and thus better handles the distribution differences in HDCN.

[0180] 3.2 Temporal Aggregation via Causal Convolution

[0181] As part of the DyTa block, the task of the causal convolution layer is to process the temporal HDCN dataset as shown in formula (1), ensuring the preservation of the causal order of the provided data. For one-dimensional temporal data, this operation can be achieved by ordinary convolution, but to achieve a sufficiently large receptive field, multiple layers of convolution need to be stacked. In order to reduce the computational cost, dilated convolution is used. Using a certain dilation step size d, the advantage of dilated convolution is that it can exponentially expand the receptive field while maintaining high computational efficiency. The process of dilated causal convolution can be defined as:

[0182]

[0183] Represents the output features of time t and layer d. The dilated causal convolution operation is generalized, where K is the convolution kernel size and D is the dilation factor. is the convolution weight of layer l, G t-i×d represents the graph data at the dilated time step. ReLU introduces nonlinearity in the convolution process. Finally, σ(·) is the nonlinear activation function applied to the convolution output, ensuring that the model captures the temporal dynamics of the graph data while preserving causal relationships.

[0184] Compared to recurrent neural networks (RNNs), dilated causal convolution allows parallel computation and alleviates the problem of gradient vanishing or exploding due to its non-recursive nature. Considering its superiority in handling time series problems, we use this module to process the time evolution information of the graph. Figure 4 As shown in (c), past graph snapshots are aggregated into a single heterogeneous data center information topology set, which encapsulates the information of the past T time periods, so that the MorpheusClos framework can effectively process the information of multiple past time periods and dynamically understand the traffic trends in the DCN.

[0185] 4 Graph-enhanced PUCB for converter mode decision

[0186] In the previous section, we used the graph transformer technique to obtain prior probabilities for identifying the best transformer patterns (Steiner nodes). A naive approach is to rank the patterns of each transformer according to these prior probabilities. However, this approach may be unreliable because the learning-based model has only one chance to calculate the best solution per time period, and it does not backtrack or reverse decisions. In addition, as the number of transformers increases, the decision space grows exponentially, and decisions cannot be made in parallel because the decisions between different transformers affect each other.

[0187] This section outlines the decision process of the prior probability output that guides the mode decision of the converter in the MorpheusClos framework. We introduce the exploration and exploitation balance principle of proportional confidence upper bound (PUCB) in this step. We construct a search tree containing the mode nodes of the converter embedded with the prior probabilities, ensuring that different mode nodes of the same converter cannot be visited simultaneously. Each node of the search tree in time period $t$ is represented as a triple (N(v,t),Q(v,t),P(v,t)), where N(v,t) is the number of visits, Q(v,t) is the remaining link capacity percentage of the current node, and P(v,t) is the highest prior probability of the node.

[0188] PUCB decision process: In the context of the Steiner tree problem, our goal is to explore configurations that minimize link utilization. The decision process consists of several stages: Selecting a strategy: The process starts from the root node, and each step is based on a strategy that balances exploration (visiting the converter pattern suggested by the prior probability) and utilization (visiting the converter pattern of the previous time period). We define the overall reward of a node as a combination of its prior probability and the percentage of remaining link capacity. The method for determining the child node (actually refers to the converter pattern, not the server, switch, converter above) is as follows:

[0189] childrennode←argmax v (Q(v,t)+U(v,t))

[0190]

[0191] Among them, c puctis a constant that balances exploration and exploitation. Here, U(v,t) represents the confidence upper bound of node v. Recursive decision: After selecting a child node, the decision is translated into a change in the DCN topology. The remaining nodes in the search tree are then re-evaluated. The selected child node becomes the root node for the next stage of the search process, which involves recursively calculating the confidence upper bounds of its child nodes and selecting the next node based on these calculations. State update: Once all switches make a decision, the statistics of each node in the search tree are updated, including the number of visits N(v), the remaining link capacity percentage Q(v), and the prior probability P(v).

[0192] Scheduling and Adaptation: Our framework optimizes the switch configuration to evenly distribute network traffic. Figure 4 (d) shows how changes in the confidence upper bounds of the switch nodes can lead to topological adjustments, such as the red switch on the right changing its connection pattern in response to a decrease in the confidence upper bound of its current mode node. This approach is particularly effective in dynamic and heterogeneous distributed communication networks.

[0193] 4 Practical considerations

[0194] The workflow includes the following stages. First, the server actively sends transmission requests and service information to the switch and controller. Subsequently, the switch and the converter collaborate on routing, and the switch sends dynamic operation information to the converter. Third, the controller collects information from the converter to update the heterogeneous graph information set. Finally, the controller propagates the converter mode decision information, and the routing in the switch is reconfigured in response to the topology change.

[0195] Decision cycle: First, considering the certain stability of data center workloads, we choose to sense and make decisions periodically. Based on experience, we set each cycle to 90 seconds, including a control phase and a longer uptime, about 50 times that of the control phase, to achieve a balance between overhead and adaptability. In the control phase, the optical path link of the converter is temporarily disconnected, and the traffic needs to be buffered and wait for the next allocation.

[0196] Time cost of topology change: Time cost is another key factor affecting the overall performance of DCN. This includes the time cost of adjusting OpenFlow rules and the physical reconfiguration time cost of switches. We observed through experiments that the core bandwidth of the network drops after the topology change and quickly returns to a stable state within 1.5 to 2.5 seconds. Considering the benefits of network bandwidth brought by topology changes, the impact of time cost on bandwidth is about 2%, which is acceptable.

[0197] In summary, the first innovation of this invention is the introduction of an innovative DCN topology that combines the converter with the existing CLOS structure. This design uses the dynamic connection mode of the converter to achieve diversified cross-Pod connections and facilitate interconnection between different switch layers. By enhancing the adaptability of the CLOS network, it significantly improves the system's resilience to sudden traffic surges and enables the physical topology to intuitively adapt to network traffic patterns.

[0198] The second innovation of this paper is the introduction of a graph representation learning framework designed specifically for non-homogeneous networks. Compared with traditional graph representation learning methods, this framework shows higher ability in handling dynamic heterogeneous graph structures. By combining the temporal self-supervised Transformer with the hierarchical aggregation of heterogeneous information, our model can effectively interpret and manage heterogeneous dynamic traffic information among multiple nodes.

[0199] Of course, the present invention may have many other embodiments. Without departing from the spirit and essence of the present invention, those skilled in the art may certainly make various corresponding changes and modifications based on the present invention, but these corresponding changes and modifications should all fall within the scope of protection of the claims attached to the present invention.

Claims

1. A method for joint recognition and scheduling transmission of non-homogeneous information, characterized in that: Traffic is transmitted from the sending server to the receiving server via a meta-path composed of switches located at different layers of the heterogeneous data center network, wherein the switches on the heterogeneous data center network are divided into core layer switches, aggregation layer switches, and edge layer switches. Adjacent switches on the meta-path are connected through converters, and the number of converters is the smaller value of the number of aggregation layer switches and edge layer switches. At the same time, different working modes of the converters correspond to different switch connection methods, and the working mode of the converter changes with the traffic transmission demand of the heterogeneous data center network. When it is necessary to change the traffic transmission meta-path of the heterogeneous data center network in the next time slot, the method for obtaining the working mode of each converter in the next time slot is as follows: S1: Obtaining status information of a heterogeneous data center network in a current time slot, wherein the heterogeneous data center network is composed of multiple converters, multiple switches, and multiple servers, and the status information includes service control information of each converter and topological location information in a current time slot, service control information of each server and topological location information in a current time slot, and topological location information and dynamic operation information of each switch in a current time slot; S2: Input the state information into the trained decision system to obtain the optimal working mode of each converter in the next time slot.

2. A method for joint recognition and scheduling transmission of non-homogeneous information as claimed in claim 1, characterized in that: In step S2, different working modes of switches, servers and converters in the heterogeneous data center network are used as nodes of the heterogeneous data center network topology; at the same time, the decision system includes multiple cascaded Dyta modules, multi-layer perceptrons MLP, and PUCB models, wherein each Dyta module consists of a Dyta layer and a causal convolutional layer; Except for the Dyta layer of the first-level Dyta module, the Dyta layers of the remaining Dyta modules are used to process the first node feature information stored in the form of a heterogeneous data center network topology structure obtained by the previous-level Dyta module based on the HDCN mutual attention mechanism to obtain the second node feature information; the causal convolution layer is used to fuse the second node feature information obtained by the Dyta layer of the same Dyta module with the second node characteristic information of different time slots in history to obtain the first node feature information used to describe the causal relationship between topology scheduling and network performance; wherein the Dyta layer of the first-level Dyta module is used to convert the state information into the second node feature information based on the HDCN mutual attention mechanism, and the first node feature information obtained by the causal convolution layer of the last-level Dyta module is output to the multi-layer perceptron MLP; The multi-layer perceptron MLP is used to obtain the prior probability that various working modes of each converter in the next time slot are the optimal working mode according to the first node feature information output by the previous Dyta module; The PUCB model is used to calculate the corresponding confidence upper bounds of each converter when it is in various working modes in the next time slot based on the corresponding prior probabilities of each converter when it is in various working modes in the next time slot and the remaining link capacity percentage of each converter in the previous time slot, and use the working mode with the maximum confidence upper bound value among multiple modes of each converter as the working mode of each converter in the next time slot.

3. A method for joint recognition and scheduling transmission of non-homogeneous information as claimed in claim 2, characterized in that: The second node feature information obtained by the Dyta layer of any level Dyta module is composed of the target node output features corresponding to each node in the network, and the method for obtaining the target node output features corresponding to any network node is: Take each network node as the target node and perform the following steps to obtain the updated target node output features corresponding to each network node: S21: Get the target node v t The adjacent nodes and the network nodes with similar traffic patterns in the previous time slot are taken as source nodes v s ; S22: Each source node v s The service control information and topological location information of the previous time slot are used as the original features of the source node, and the original features of the source node are reduced to obtain the source node input features H l-1 [v s ], where l represents the source node v s The hierarchical number of the Dyta module where it is located; the source node input feature of the first-level Dyta module is defined as H 0 [v s ]; S23: Set each target node v t The service control information and topological location information of the previous time slot are used as the original features of the target node, and the original features of each target node are reduced to obtain the input features of the target node H l-1 [v t ]; the target node input feature of the first-level Dyta module is defined as H 0 [v t ]; S24: Input each target node into the feature H through the query vector linear projection function Q-linear l-1 [v t ] are mapped to multiple vector spaces respectively, and the query vector mapping Q of each vector space is obtained h (v t ): Q h (v t )=Q-linear h (H (l-1) [v t ]) Wherein, the vector space serial number h=1,2,...,H, H represents the number of vector spaces; S25: Input the source node into the feature H through the key vector linear projection function K-linear l-1 [v s ] is mapped to multiple vector spaces, and the key vector mapping K of each vector space is obtained h (v s ): K h (v s )=K-linear h (H (l-1) [v s ]) S26: Input the source node into the feature H through the value vector linear projection function V-linear l [v s ] is mapped to multiple vector spaces, and the value vector mapping V of each vector space is obtained h (v s ): V h (v s )=V-linear h (H (l-1) [v s ]) S26: Map K according to the key vector h (v s ) and query vector mapping Q h (v t ) Get the source node v s For each target node v t The importance of Attention(v s ,v t ): Among them, ATT_head h (v s ,v t ) represents any target node v in the hth vector space t and source node v s The attention weight between is calculated as: Where d represents the target node input feature H l-1 [v t ] dimension, Indicates the capture source node v s and the target node v t The transformation matrix of the characteristic relationship between Represents the source node v s and the target node v t The attention adaptation scaling factor between , T represents transposition; S27: Using linear projection Map the value vector V h (v s ) is converted to a value message ValueMessage(v s ,v t ): Among them, MSG-head h (v s ,v t ) indicates that the transformation matrix is ​​set Mapping V with value vector h (v s ) after multiplying, Indicates that all vector spaces correspond to the MSG-head h (v s ,v t ) added; S28: The importance of Attention (v s ,v t ) and value message ValueMessage(v s ,v t ) to obtain each target node v t Message vector Among them, SN(t) represents the set of source nodes selected in the current time slot t, represents the product Attention(v) corresponding to the node pairs consisting of all source nodes and their paired target nodes in the set SN(t) s ,v t )·ValueMessage(v s ,v t ) added; S29: Get the target node v according to the nonlinear activation ReLu function t Output feature H at the target node output by the current level Dyta module (l) [t]: Among them, σ is the weight of the nonlinear activation ReLu function, H (l-1) [t] is the target node v t Output features at the target node output by the previous level Dyta module.

4. A method for joint recognition and scheduling transmission of non-homogeneous information as claimed in claim 2, characterized in that: The method for obtaining the first node feature information of the causal convolution layer of any level Dyta module is as follows: in, is the output feature of the dth causal convolution of the current causal convolution layer at the current time slot t, is the first node feature information finally obtained by the current causal convolution layer at the current time slot t, D is the expansion factor used to characterize the number of layers in the causal convolution layer, K is the number of convolutions performed in each layer of causal convolution, is the weight of the causal convolution layer of the causal convolution layer in the current level l Dyta module when performing the i-th convolution, G t-i×d represents the graph data within the expansion time step, wherein the method for obtaining the graph data is: taking the different working modes of each switch, each server and converter as a node respectively, thereby forming a topological graph; taking the state information of the K×D topological graphs of the historical K×D time slots including the current time slot t in the current l-th level Dyta module as the second node feature information obtained as the graph data; σ represents the activation function, and ReLU represents the linear rectification function.

5. A method for joint recognition and scheduling transmission of non-homogeneous information as claimed in claim 2, characterized in that: The optimal operating mode of each converter in the next time slot is determined as follows: A multi-level search tree of the PUCB model is constructed, wherein the last level leaf node of the search tree corresponds to the next time slot, and each level node of the remaining nodes corresponds to a historical time slot, and each node represents a converter working mode and its related statistical information, wherein the statistical information includes the number of times selected as the optimal working mode, the percentage of the remaining link capacity in the previous time slot, and the prior probability; at the same time, except for the root node and the leaf node, the remaining nodes are both parent nodes and child nodes; The search tree is traversed according to the set search rule, and the confidence upper bounds of the converter working modes corresponding to the nodes at each level are calculated level by level, wherein the set search rule is that the nodes with fewer selection times are selected first, and the nodes with larger confidence upper bounds are selected first; at the same time, the confidence upper bound of the converter working mode corresponding to any node is obtained as follows: Among them, U(v,t) represents the confidence upper bound of the converter working mode corresponding to the current node v in the time slot t corresponding to its own level, N(v,t) is the number of times the converter working mode corresponding to the current node v is selected as the optimal working mode before the end of the time slot t corresponding to its own level, P(v,t) is the highest priority probability of the converter working mode corresponding to the current node v in the time slot t corresponding to its own level, c puct is a setting constant that balances exploration and exploitation; For each parent node, select the child node with the largest Q(v,t)+U(v,t) from each child node as the optimal child node: Best child node ← argmax v (Q(v,t)+U(v,t)) Where Q(v,t) is the percentage of the remaining link capacity of the converter working mode corresponding to the current node v in the previous time slot of the time slot t corresponding to its own level; The newly selected optimal child node is used as the new parent node, and the number of times the converter working mode corresponding to all nodes of the search tree is selected as the optimal working mode is updated. Then the confidence upper bounds of each child node of the selected new parent node are calculated, and the updated parent node is selected according to the confidence upper bound; and so on, the confidence upper bound of the last-level leaf node is obtained, and the converter working mode corresponding to the leaf node with the largest confidence upper bound is used as the optimal working mode of the current converter in the next time slot.

Images