A wired-wireless hybrid interconnected star-mesh matrix data center network architecture and scheduling method

Abstract

The application relates to the field of data center networks, and discloses a wired-wireless hybrid interconnected star-moon matrix data center network architecture and a scheduling method. The architecture comprises a plurality of modularized star-moon units, all the star-moon units are connected through horizontal matrix expansion and vertical hierarchical expansion to form a three-dimensional star-moon matrix network. The architecture has strong compatibility and scene adaptation capability, can be used for full deployment of a newly-built super-large-scale computing center, can be used for local modification and upgrading of an existing data center, can be widely applied to various application scenes such as cloud computing, artificial intelligence training, high-performance computing and edge data center, and has high popularization value and industrial application prospect.

Description

Technical Field

[0001] This invention relates to the field of data center networks, and in particular to a wired and wireless hybrid interconnected star matrix data center network architecture and scheduling method. Background Technology

[0002] With the large-scale deployment of technologies such as cloud computing, big data, generative artificial intelligence, and high-performance computing, data centers, as the core carriers of the digital economy and computing infrastructure, are experiencing explosive growth in east-west traffic between their internal servers. This places unprecedentedly stringent demands on the bandwidth capacity, transmission latency, dynamic scalability, operational reliability, and maintenance flexibility of data center networks (DCNs). Traditional data center network architectures are no longer adequate for the communication needs of ultra-large-scale computing clusters, becoming a core bottleneck restricting the release of computing power.

[0003] Currently, the mainstream network architecture of commercial data centers is built on wired fiber optic communication (FOC) technology, with the Spine-Leaf two-level architecture being the most widely used. This architecture, through full connectivity of spine and leaf switches, improves the forwarding efficiency of east-west traffic to some extent. However, it still has inherent flaws that cannot be overcome: First, the architecture's performance is strictly limited by the number of ports on the physical switches. As the scale of server nodes increases, the core spine switch is prone to becoming a bandwidth bottleneck, making it impossible to achieve linear growth in network bandwidth with the number of nodes. Second, the topology relies entirely on fixed fiber optic cabling, which is a static physical topology. When the data center needs to expand server capacity, adjust rack layout, or dynamically schedule computing resources, the construction cost of re-laying and splicing fiber optic cables is high and the deployment cycle is long, making it completely unsuitable for the dynamic resource allocation and plug-and-play requirements in computing pooling scenarios. Third, traffic forwarding is highly dependent on dedicated switching equipment, posing a single point of failure risk. Once the core spine switch malfunctions, it will lead to a large-scale network outage. At the same time, slow routing convergence and high forwarding hop counts in large-scale networking further exacerbate transmission latency and congestion risks.

[0004] To overcome the static limitations of wired architectures, Free-Space Optical Communication (FSO) technology has begun to be introduced into data center networking scenarios. FSO technology uses lasers as the information carrier, achieving high-speed data transmission without a physical transmission medium. It also boasts advantages such as flexible deployment, resistance to electromagnetic interference, no cabling required, and low latency, providing a technological possibility for the dynamic reconfiguration of data center networks. Existing technologies have begun to explore hybrid interconnected data center architectures combining FOC and FSO. Mainstream solutions often adopt a hierarchical model of "wired backbone + wireless supplement" or a primary / backup model of "wired primary + wireless backup," attempting to combine the high bandwidth stability of wired links with the deployment flexibility of wireless links.

[0005] However, in practical applications and technical research, existing wired and wireless hybrid interconnection data center network technologies still suffer from the following core technical shortcomings that are difficult to resolve: First, the topology design does not achieve deep integration of wired and wireless technologies, resulting in insufficient dynamic reconfiguration capabilities. Existing hybrid architectures mostly involve simply overlaying wireless links onto traditional wired topologies, failing to achieve collaborative design of the two communication technologies at the underlying topology structure. Wireless links serve only as supplementary channels, unable to leverage their core advantage of flexible networking. The network topology remains constrained by the static limitations of wired cabling, offering limited flexibility in the face of node expansion and layout adjustments, and failing to achieve dynamic topology reconfiguration.

[0006] Second, the lack of a heterogeneous link coordination mechanism results in low traffic scheduling efficiency and insufficient reliability assurance. Existing solutions do not design differentiated scheduling mechanisms for the performance characteristics of FOC and FSO links. Most only treat wireless links as emergency backups for wired links, failing to achieve resource complementarity and load balancing between the two types of links. At the same time, facing the link quality degradation caused by vibration and obstruction of FSO links, as well as the congestion and interruption problems of wired links, existing technologies lack efficient seamless switching and dynamic compensation mechanisms. When links fail, problems such as service interruption, latency jitter, and increased packet loss rate are likely to occur, making it difficult to guarantee the transmission continuity of high-priority services.

[0007] Third, the architecture lacks standardized and modular design, resulting in severely insufficient scalability in large-scale scenarios. Existing hybrid interconnection topologies are mostly designed for single racks and small-scale clusters, failing to form replicable and scalable standardized basic units. As the number of nodes increases, the complexity of network connections increases exponentially, and the overhead of routing calculations and network management increases sharply. At the same time, the existing architecture has not yet overcome the performance bottleneck of core switching nodes. As the cluster size increases, the traffic pressure on core nodes increases dramatically, which can easily lead to network-wide congestion. It cannot achieve linear scaling of network performance with the number of nodes and is difficult to adapt to the deployment needs of ultra-large-scale data centers with tens of thousands of nodes.

[0008] Fourth, the routing mechanism is inefficient and the control plane overhead is too high. The existing hybrid architecture topology lacks regularity and symmetry, making it unsuitable for efficient static route planning. It can only rely on dynamic routing protocols such as OSPF for addressing. In large-scale networking scenarios, the route learning and convergence process consumes a lot of computing and bandwidth resources, which not only increases the end-to-end latency of data forwarding but also reduces the self-healing efficiency after network failures, making it difficult to meet the transmission requirements of high real-time services.

[0009] In addition, existing patented technologies for wired and wireless hybrid data center networks, such as Chinese patent CN115801792A, propose a structure and method for interconnecting wireless and wired converged data centers. However, it still adopts a hierarchical architecture, fails to overcome the performance bottleneck of the core switching node, and does not design a modular topology that can be scaled up. Chinese patent CN113726879A proposes a hybrid data center network structure based on VLC links, but it is only for short-range communication scenarios within racks and cannot achieve large-scale network expansion at the system level. It also fails to solve the problems of heterogeneous link collaborative scheduling and routing optimization.

[0010] In summary, existing technologies still cannot simultaneously meet the comprehensive requirements of hyperscale data centers for low latency, high bandwidth, easy scalability, high reliability, and easy operation and maintenance. There is an urgent need for a new data center network architecture that can achieve deep integration of wired and wireless networks, has linear scalability, and efficient intelligent scheduling. Summary of the Invention

[0011] In order to overcome the shortcomings of the prior art, one of the objectives of this invention is to provide a wired and wireless hybrid interconnected star matrix data center network architecture and scheduling method.

[0012] One of the objectives of this invention is achieved through the following technical solution: A wired and wireless hybrid interconnected star matrix data center network architecture includes several modular star units, all of which form a three-dimensional star matrix network through horizontal matrix expansion and vertical hierarchical expansion. Each of the starburst units is a topology structure composed of three interwoven nodes, including one central wired server node in the innermost layer, six wireless server nodes in the middle layer, and six peripheral wired server nodes in the outermost layer; wherein, the six wireless server nodes are arranged in a regular hexagon, the central wired server node is deployed at the center of the regular hexagon, and the six peripheral wired server nodes are deployed on the periphery of the regular hexagon. The central wired server node, wireless server node, and peripheral wired server nodes are connected to form a highly redundant quasi-fully connected local communication network through wired fiber optic communication FOC links and wireless free space optical communication FSO links; wherein, the wired server nodes are equipped with FOC communication modules and the wireless server nodes are equipped with FSO communication modules. In the horizontal direction, multiple star-shaped units are arranged in an N×M array. Adjacent star-shaped units are interconnected through the FSO link of the wireless server node to establish a torus topology interconnection, realizing point-to-point data penetration between adjacent units. In the vertical direction, the star-shaped matrix network supports N-layer deep expansion. Wireless server nodes at corresponding positions in different layers establish cross-layer communication channels through the FSO link. The star matrix network is a decentralized architecture, with each server node integrating routing and forwarding functions. The communication bandwidth increases linearly with the increase in the number of nodes. Each server node maintains the same connectivity with its neighboring nodes within its own star unit, forming a symmetrical bidirectional ring-shaped self-healing ring structure, which is used for path redirection and local fault self-healing in the event of link failure.

[0013] As a further improvement to the above technical solution: Within the starburst unit, the central wired server node communicates with six wireless server nodes via FOC links; the six wireless server nodes communicate with each other via FSO links in a ring topology; and each wireless server node communicates with its corresponding peripheral wired server node via an FOC link.

[0014] The FOC link is used to carry the first type of service flow that exceeds the preset data packet size threshold, and the FSO link is used to carry the second type of service flow that is smaller than the preset data packet size threshold. At the same time, the FSO link serves as a redundant backup link for the FOC link.

[0015] The Star Matrix Network pre-calculates the shortest path across the entire network based on the Breadth-First Search (BFS) algorithm, generating a static routing map table for the entire network. This static routing map table assigns a unique point-to-point subnet address to each node, thereby reducing the computational overhead of network initialization and topology reconstruction.

[0016] When any single link within the Starburst Unit suffers physical damage, or when the backbone FOC link experiences congestion / physical damage, the system quickly redirects to a backup loop path or a backup path based on the FSO link using the BFS algorithm, achieving seamless data switching and network self-healing.

[0017] A scheduling method for a wired-wireless hybrid interconnected star matrix data center network, implemented based on the star matrix data center network architecture described in any one of claims 1-5, the method comprising the following steps: S1. Topology construction steps: Build a star matrix network composed of several star units, complete the deployment of heterogeneous links within the star units, and the horizontal and vertical interconnection between star units. S2. Traffic classification steps: Identify the data packet size of the business flow in real time, and classify the business flow into the first type of business flow and the second type of business flow based on the preset data packet size threshold; S3. Link matching and scheduling steps: Assign the first type of service flow to the FOC link for transmission, and assign the second type of service flow to the FSO link for transmission; S4. Routing and Self-Healing Guarantee Steps: Based on the topological characteristics of the star matrix network, path planning and redundancy backup are completed through the BFS algorithm. When the main link fails or becomes congested, the system quickly switches to the backup path to complete data transmission.

[0018] As a further improvement to the above technical solution: The S1 topology construction steps specifically include: S101. Constructing the basic starburst unit: Deploy 6 wireless server nodes arranged in a regular hexagon, deploy a central wired server node at the center of the regular hexagon, and deploy 6 peripheral wired server nodes around the regular hexagon to establish a quasi-full connection link between FOC and FSO within the unit. S102, System-level expansion deployment: Arrange multiple starburst units horizontally in an N×M array, and construct a horizontal toroidal topology through the FSO links of the wireless server nodes; replicate the single-layer starburst matrix network, and connect the wireless server nodes at corresponding positions of different levels through FSO links to complete the vertical hierarchical expansion and construct a three-dimensional matrix communication architecture.

[0019] In the S2 traffic classification step, the preset data packet size threshold is 1MB. Service flows with data packet size ≥ 1MB are defined as the first type of service flow, namely the elephant flow, and service flows with data packet size < 1MB are defined as the second type of service flow, namely the ant flow.

[0020] In the S4 routing and self-healing guarantee step, the shortest path of the entire network is pre-calculated using the BFS algorithm to generate a static route mapping table and an XML-formatted static route table, completing the path planning in the network initialization phase, and ensuring that data packets are transmitted along the shortest path with the fewest hops.

[0021] In the S4 routing and self-healing assurance steps, when the primary data link detects congestion or physical failure, it quickly redirects to the backup redundant path based on the BFS algorithm to achieve load balancing and network self-healing, ensuring the continuity of service transmission.

[0022] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention significantly reduces the average number of hops for data forwarding through a compact topology design of modular starburst units. Within each unit, a quasi-fully connected structure enables high-speed local data exchange, while cross-unit communication utilizes wireless FSO links for point-to-point penetration, avoiding the multi-hop forwarding losses inherent in traditional architectures where traffic must pass through the core switch. Furthermore, based on the topology's regular symmetry, the BFS algorithm is used to pre-compute a static routing mapping table, completely avoiding the control plane overhead and forwarding latency of dynamic routing protocols. Experimental data shows that the single-layer starburst matrix of this invention reduces end-to-end latency by 29.8% under full load conditions compared to a traditional Spine-Leaf architecture of the same scale, and the three-layer large-scale starburst matrix reduces latency by 47.4% compared to a Spine-Leaf architecture of the same scale. Even when scaling from a single layer to a three-layer architecture, doubling the node size, the average latency increase under full service intensity is only 10.5%, maintaining extremely low latency levels even in ultra-large-scale networking scenarios. This greatly improves the real-time response capability of data center networks, perfectly adapting to the low-latency-sensitive service requirements of AI large-scale model training and high-performance computing.

[0023] 2. This invention uses standardized starburst units as the smallest modular networking unit and designs a three-dimensional networking mechanism that combines horizontal matrix expansion and vertical hierarchical expansion, completely eliminating the dependence of traditional architectures on dedicated core switches. This architecture adopts a decentralized design, with all server nodes integrating routing and forwarding functions. The total forwarding capacity and communication bandwidth of the network increase linearly with the increase in node scale. Experimental data verifies that the full-load throughput of the three-layer starburst matrix architecture of this invention reaches 6.13Gbps, approximately three times that of a single-layer architecture, achieving a near-ideal bandwidth multiplication effect; compared to the traditional Spine-Leaf architecture of the same scale, the throughput is increased by 250.3%, completely solving the core pain point that the performance of traditional architectures cannot grow synchronously with scale expansion. Simultaneously, based on the networking characteristics of wireless FSO links, no new fiber optic cables need to be laid during data center expansion and layout adjustments, significantly reducing the construction cost and cycle of expansion. It enables plug-and-play and dynamic deployment of computing nodes, perfectly adapting to the resource scheduling needs of computing power pooling.

[0024] 3. This invention achieves deep integration of FOC wired links and FSO wireless links at the top-level, rather than simply superimposing and supplementing them. It designs a differentiated traffic scheduling mechanism based on the performance characteristics of the two types of links: the high-bandwidth, low-latency FOC links are used to carry large-capacity "elephant streams," ensuring stable transmission of core services; the highly flexible and easily deployed FSO links are used to carry small data packet "ant streams," enabling flexible forwarding across units and layers, while also alleviating the traffic pressure on wired links. This mechanism achieves complementary advantages and optimal resource allocation for heterogeneous links, avoiding congestion bottlenecks in wired links while fully leveraging the flexibility of wireless links. It achieves stable linear throughput growth under all service load gradients, avoiding the link saturation and performance drop problems common in traditional architectures, significantly improving the overall network resource utilization efficiency and load balancing capabilities.

[0025] 4. This invention constructs a multi-dimensional backup path system for the network through a symmetrical bidirectional ring-shaped self-healing ring structure within the starburst unit, combined with a fully redundant interconnection topology between units and layers. Simultaneously, based on the BFS algorithm, a fast rerouting and seamless switching mechanism is designed. When a link experiences congestion, physical damage, or interruption, the system can complete the redirection of the backup path within milliseconds, achieving seamless service switching and seamless recovery. Specialized experiments verify that the average path switching latency for a single link failure is only 1.28ms, and the service interruption duration is less than 5ms. Even with 10 concurrent link failures, the self-healing success rate remains 100%, and the impact of the failure is limited to the affected starburst unit, without causing network-wide performance fluctuations. This completely solves the problems of poor reliability and slow fault recovery in existing hybrid architecture heterogeneous link collaboration, providing a solid guarantee for the continuous and stable operation of data center services.

[0026] 5. This invention employs a standardized and modular starburst unit design, resulting in a network topology with high regularity and symmetry. Routing planning can be achieved through pre-computed static routing tables, eliminating the need for complex dynamic routing protocols. This significantly reduces routing computation overhead and control plane bandwidth consumption during large-scale networking, and substantially improves the efficiency of network initialization and topology reconfiguration. Simultaneously, the modular design greatly simplifies network deployment, operation, and expansion processes. It eliminates the need to redesign topology and routing rules for clusters of different sizes, and management complexity does not increase exponentially with node size, effectively reducing the operational difficulty and costs of ultra-large-scale data centers.

[0027] Furthermore, the architecture of this invention has strong compatibility and scenario adaptability. It is suitable for full deployment in newly built ultra-large-scale computing centers, as well as for partial transformation and upgrading of existing data centers. It can be widely adapted to various application scenarios such as cloud computing, artificial intelligence training, high-performance computing, and edge data centers, and has extremely high promotional value and industrial application prospects.

[0028] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described in detail below with reference to the accompanying drawings. Attached Figure Description

[0029] Figure 1 This is a schematic diagram of the starburst unit in this embodiment; Figure 2 This is a schematic diagram of a 3x3 star matrix in this embodiment; Figure 3 Comparison of average end-to-end latency of star matrix networks at different service intensities; Figure 4 Comparison of total throughput of star matrix networks at different levels under different service intensities; Figure 5 Comparison of average end-to-end latency between the starburst matrix and the traditional Spine-Leaf architecture; Figure 6 Comparison of total network throughput between the star matrix and the traditional Spine-Leaf architecture. Detailed Implementation

[0030] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments. It should be noted that, without conflict, the various embodiments or technical features described below can be arbitrarily combined to form new embodiments.

[0031] It should be noted that when a component is described as "fixed to" another component, it can be directly on the other component or may have a component in between. When a component is considered "connected to" another component, it can be directly connected to the other component or may have a component in between. When a component is considered "set on" another component, it can be directly set on the other component or may have a component in between. The terms "vertical," "horizontal," "left," "right," and similar expressions used in this document are for illustrative purposes only.

[0032] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0033] Example 1: Specific Implementation of the Modular Starburst Unit This embodiment is the smallest basic modular unit of the entire star matrix network, and it is the core carrier for realizing heterogeneous link fusion and local fault self-healing. The specific implementation details of this embodiment are as follows: 1. Node composition and spatial layout A single Starlight Unit (SU) consists of 13 server nodes interwoven across three layers, specifically including: Innermost layer: 1 central wired server node, hereinafter referred to as C node, deployed at the very center of the unit; Middle layer: 6 wireless server nodes, hereinafter referred to as W nodes, numbered W1-W6 in sequence, are evenly arranged in a standard regular hexagon. The geometric center of the regular hexagon coincides with the deployment position of the C node, and the spacing between adjacent W nodes is equal. Outermost layer: 6 peripheral wired server nodes, hereinafter referred to as P nodes, numbered P1-P6 in sequence, deployed on the outside of the 6 sides of the regular hexagon, with each P node corresponding to 1 W node.

[0034] 2. Configuration of heterogeneous communication links All wired server nodes (C nodes and P nodes) are equipped with wired fiber optic communication FOC modules, with a basic bandwidth configuration of 10Gbps per module, a one-way propagation delay of 100ns, and a link error rate of 1e-12. They use single-mode fiber for physical connection. All wireless server nodes (W nodes) are equipped with free-space optical communication FSO modules, with a basic bandwidth configuration of 1Gbps per module, a one-way propagation delay of 50ns, and a link error rate of 1e-9. They use indoor unobstructed laser transmission for wireless interconnection.

[0035] 3. Intra-unit topological connections The starburst unit internally constructs a highly redundant quasi-fully connected local communication network through FOC and FSO links, with the specific connection rules as follows: Node C establishes full-connection communication with six nodes W1-W6 via the FOC link, enabling high-speed data interaction between the central node and the intermediate layer nodes; The six W nodes, W1-W6, are connected bidirectionally in a regular hexagonal path via FSO links, namely W1 to W2, W2 to W3, W3 to W4, W4 to W5, W5 to W6, and W6 to W1, forming a closed bidirectional ring topology. Each W node establishes a dedicated communication connection with its corresponding peripheral P node through the FOC link, i.e., W1 connects to P1, W2 connects to P2, and so on, to achieve high-speed data interaction between peripheral wired nodes and intermediate wireless nodes.

[0036] 4. Local self-healing ring function realized. Based on the aforementioned symmetrical ring topology, each server node maintains the same connectivity with its neighboring nodes within the unit, forming a bidirectional self-healing ring structure. When any single-point link within the unit suffers physical damage or interruption, the system can quickly redirect the backup path within the unit based on the breadth-first search (BFS) algorithm.

[0037] For example, when the FSO main link between W1 and W2 is interrupted, the system can quickly match multiple backup paths through the BFS algorithm, including the primary backup path W1→C→W2 and the secondary backup path W1→W6→W5→W4→W3→W2. The system prioritizes the shortest path with the fewest hops to complete data forwarding, achieving millisecond-level self-healing of local faults without the need for cross-unit scheduling, thus ensuring the continuity of service transmission.

[0038] Example 2: Specific Implementation of a Single-Layer 3×3 Starburst Matrix Network This embodiment, based on the starburst unit described in Embodiment 1, achieves horizontal matrix expansion to construct a single-layer planar starburst matrix network. Specific implementation details are as follows: 1. Cell array arrangement rules Using the starburst unit described in Example 1 as the smallest expandable module, the nine starburst units are arranged in a 3x3 array in a plane. Each starburst unit is denoted as SU(i,j), where i is the row number (values ​​1, 2, 3) and j is the column number (values ​​1, 2, 3). That is, the array contains a total of nine units: SU(1,1), SU(1,2), SU(1,3), SU(2,1), SU(2,2), SU(2,3), SU(3,1), SU(3,2), and SU(3,3).

[0039] 2. Interconnection rules between adjacent units To achieve point-to-point data penetration between adjacent star-shaped units, a horizontal torus topology interconnection is constructed based on the FSO link of the W node. The specific interconnection rules are as follows: The orientation of nodes W is predefined: W1 is the node directly above the regular hexagon, W2 is the node at the top right, W3 is the node at the bottom right, W4 is the node directly below, W5 is the node at the bottom left, and W6 is the node at the top left. Interconnection between adjacent units in the same row: Within the same row, the W2 and W3 nodes of the left star-shaped unit SU(i,j) establish point-to-point bidirectional communication connections with the W5 and W6 nodes of the adjacent star-shaped unit SU(i,j+1) on the right through the FSO link. Interconnection between adjacent units in the same column: Within the same column, the W3 and W4 nodes of the upper starburst unit SU(i,j) establish point-to-point bidirectional communication connections with the W1 and W6 nodes of the lower adjacent starburst unit SU(i+1,j) respectively through the FSO link; Toroidal Closed Interconnection: The star-shaped cells at the edge of the array are interconnected with the star-shaped cells on the opposite edge through FSO links, forming a closed toroidal topology. For example, the W1 node of the first row SU(1,j) is interconnected with the W4 node of the third row SU(3,j), and the W6 node of the first column SU(i,1) is interconnected with the W2 node of the third column SU(i,3), further improving the link redundancy of the entire network.

[0040] 3. Decentralized architecture implementation The single-layer 3×3 star matrix network constructed in this embodiment lacks the core switching equipment found in traditional architectures. Each server node (C node, W node, P node) integrates routing and forwarding functions, allowing nodes to directly make local data forwarding decisions without needing to aggregate to the core switching equipment. This embodiment has a total of 9×13=117 server nodes. The forwarding capabilities of all nodes are cumulative, and the overall network communication bandwidth increases linearly with the increase in node scale, fundamentally solving the port bottleneck problem of the core switch in traditional architectures.

[0041] Example 3: Specific Implementation of a Multi-Layer Three-Dimensional Starburst Matrix Network This embodiment, based on the single-layer star matrix described in Embodiment 2, achieves vertical hierarchical expansion to construct a three-dimensional star matrix network, adapting to the deployment requirements of large-scale and ultra-large-scale data centers. Specific implementation details are as follows: 1. Vertical hierarchical expansion rules Taking the single-layer 3×3 star matrix described in Example 2 as a single-level unit, it supports N-layer deep expansion. This example uses the commonly used two-layer and three-layer architectures in the industry as examples for detailed explanation: Layer definition: The vertical layers are denoted as L1, L2, and L3 respectively. Each layer is an independent 3×3 star matrix network. The physical spaces between the layers are arranged in parallel, and the spatial coordinates of the star units and server nodes at corresponding positions correspond one-to-one. Cross-layer interconnection rules: In different layers, W nodes within corresponding star-shaped units establish vertical cross-layer communication channels through FSO links. For example, W1 node in SU(1,1) of layer L1 establishes bidirectional communication connections with W1 nodes in SU(1,1) of layer L2 and SU(1,1) of layer L3 through FSO links in sequence; and so on, all W nodes establish corresponding vertical cross-layer links to achieve direct data penetration between layers without having to go through wired links within the layer.

[0042] 2. Node configuration for different architecture scales Dual-layer star matrix network: It contains two 3×3 star matrix levels, with a total of 18 star units and a total of 18×13=234 server nodes, which is suitable for medium-sized data center scenarios. Three-layer star matrix network: It contains three 3×3 star matrix layers, with a total of 27 star units and a total of 27×13=351 server nodes, which is suitable for large-scale data center scenarios.

[0043] 3. Advantages of cross-layer routing and expansion The three-dimensional architecture constructed in this embodiment allows for intra-layer communication via the horizontal toroidal topology described in Embodiment 2, while cross-layer communication can be directly accomplished via vertical FSO links, without altering the original intra-layer topology and routing configuration. When the data center needs expansion, only a complete hierarchical matrix needs to be added, and vertical interconnection can be achieved through the FSO links of the corresponding W nodes, eliminating the need for re-laying fiber optic cabling, significantly reducing the cost and time of expansion, and enabling plug-and-play dynamic expansion.

[0044] Example 4: Specific Implementation of Traffic Differentiation Scheduling and Intelligent Routing Self-Healing Method This embodiment, based on the starburst matrix network architecture described in the above embodiments, implements the corresponding network scheduling method, fully realizing the entire process of traffic classification, link matching, intelligent routing, and fault self-healing. The specific implementation steps are as follows: 1. Implementation of Traffic Classification and Differentiated Scheduling Step 1: Preset traffic classification threshold. Configure the data packet size threshold of 1MB in the application layer and forwarding layer of the node as the classification standard for business flows; Step 2: Real-time business flow identification. The node parses the business data packets generated by the sender in real time, extracts the data packet size parameter, and completes the business flow classification: business flows with a data packet size ≥ 1MB are defined as the first type of business flow, i.e., elephant flow; business flows with a data packet size < 1MB are defined as the second type of business flow, i.e. ant flow. Step 3: Heterogeneous link collaborative matching. For elephant streams, the system automatically allocates them to high-throughput FOC wired links for transmission, ensuring the stability and low latency of large-capacity data transmission. For ant streams, the system automatically allocates them to highly elastic FSO wireless links for transmission. By leveraging the flexible forwarding capabilities of wireless links, the system alleviates the traffic pressure on core wired links and enables efficient data interaction across units and levels.

[0045] 2. Implementation of Static Route Pre-computation Based on BFS Algorithm Step 1: Topology information acquisition. Obtain complete topology information of the star matrix network, including the location of all nodes, the link connection relationship between nodes, link type and bandwidth parameters; Step 2: Shortest path pre-calculation. Considering the regular and symmetrical topological characteristics of the star matrix, a breadth-first search (BFS) algorithm is used to traverse all nodes in the network and pre-calculate the shortest path from any source node to all destination nodes in the network, with the core optimization objective being to minimize the number of forwarding hops. Step 3: Static routing table generation. Using a Python script, based on the calculation results of the BFS algorithm, a unique point-to-point subnet IP address is assigned to each node in the entire network, generating a corresponding XML format static routing table file. The routing table contains core information such as the outgoing port, next-hop node, and link type for each destination address. Step 4: Loading and Activating the Routing Table. During the network initialization phase, the generated static routing table is preloaded into the routing module of the corresponding node. When the node starts up, it directly reads the routing table to complete the forwarding rule configuration, without having to run dynamic routing protocols such as OSPF, which greatly reduces the routing calculation overhead and control plane bandwidth consumption during large-scale network initialization.

[0046] 3. Seamless switching and self-healing implementation of heterogeneous links Step 1: Real-time monitoring of link status. Each node periodically checks the operating status of all its links. Monitoring indicators include link connectivity, end-to-end latency, packet loss rate, and bandwidth utilization. Preset fault trigger conditions: physical link interruption, packet loss rate > 1%, one-way latency > 10ms, and link bandwidth utilization > 95% for more than 100ms. Step 2: Backup path redirection. When the primary link meets the failure triggering conditions, the node immediately triggers the BFS algorithm. Based on the currently available links in the entire network, it quickly calculates the backup redundant path from the source node to the destination node and prioritizes the path with the fewest hops and the highest available bandwidth. Step 3: Seamless switching and seamless recovery. The system seamlessly switches the business flow to be transmitted to the backup path for transmission. For business sessions that are already in transmission, the session status is kept uninterrupted. The business interruption time during the switching process is controlled within 5ms, so as to achieve seamless data recovery. Step 4: Load balancing optimization. When high load congestion is detected on the backbone FOC link, the system can dynamically divert some non-core elephant and ant flows to idle FSO links to achieve traffic load balancing across the entire network and avoid overall performance degradation caused by local link congestion.

[0047] Experimental verification process I. Experimental Objective The performance of the star matrix network architecture of this invention in terms of latency and throughput under different service intensities was verified, and the performance stability after the layer expansion was verified. By comparing the performance difference between the present invention and the mainstream traditional Spine-Leaf architecture in data centers at the same node scale, the technical advantages of the present invention are quantified. To verify the fault self-healing capability of the present invention, and to test the switching latency and service continuity assurance effect under link failure scenarios; To verify the effectiveness of the traffic differentiation scheduling mechanism of the present invention, the performance difference between heterogeneous link collaborative scheduling and traditional scheduling methods is compared.

[0048] II. Experimental Environment and Parameter Configuration (I) Simulation Platform and Hardware Environment Simulation software: OMNeT++ version 5.7, paired with the INET 4.4 network simulation framework, which is the industry's common standard platform for data center network simulation; Operating environment: Linux Ubuntu 22.04LTS operating system, hardware platform is Intel Xeon 8375C CPU, 128GB DDR4 memory, 2TB SSD storage, to ensure the computational stability and data read and write efficiency of the simulation process.

[0049] (II) Configuration of core simulation parameters All scenarios in this experiment use a unified set of basic parameters to eliminate the interference of non-core variables on the experimental results. The specific parameters are as follows: ; III. Experimental Design and Control Group Setup This experiment uses the controlled variable method. All experimental groups and control groups have the same total number of server nodes, business traffic model, and basic node capabilities. The only difference is the network architecture and scheduling method, which ensures the comparability of the experimental results.

[0050] (I) Experimental Group (Technical Solution of the Invention) ; (II) Control Group (Traditional Mainstream Technical Solution) Using the Spine-Leaf architecture, the most widely used architecture in the data center industry, as a control, and configured with a node scale that perfectly matches the experimental group, as follows: ; (III) Setting Business Intensity Gradients To verify performance under different load scenarios, 10 gradients of normalized business intensity were set from 0.1 to 1.0. The parameter configurations for each gradient are shown in the table below: Table 1 Simulation parameter configurations for different business intensities ; IV. Complete Experimental Procedure Step 1: Network Topology Model Construction 1. Experimental Group Topology Construction: Using the NED language of OMNeT++, a modular model of the starburst unit was first defined, encapsulating node types and internal link connections. Based on the starburst unit module, single-layer, double-layer, and triple-layer 3×3 starburst matrix topologies were built respectively, defining horizontal FSO interconnection links between adjacent units and vertical FSO interconnection links across layers, thus completing the model construction of the three experimental groups.

[0051] 2. Control group topology construction: Spine switch and Leaf switch modules were defined using NED language, following the full connectivity rules of the traditional Spine-Leaf architecture. Spine switches were fully interconnected with all Leaf switches, and Leaf switches were fully interconnected with their downstream servers. Three network models for the control group were built.

[0052] Step 2: Implementation of Routing and Scheduling Mechanisms 1. Experimental group function implementation: Implement the BFS shortest path pre-calculation algorithm based on Python scripts, generate a static routing table in XML format for each experimental group, and load it into the corresponding node; implement traffic classification function in the node, classify and allocate links for elephant flow and ant flow according to the 1MB threshold; implement link status monitoring and fault self-healing switching function, and complete the deployment of the whole process scheduling mechanism.

[0053] 2. Control group functionality: The OSPF dynamic routing protocol, commonly used in traditional data centers, is adopted to enable route learning and updates between switches; the standard Equal Cost Multipath (ECMP) algorithm is used to achieve traffic load balancing and a non-differentiated traffic scheduling mechanism, consistent with actual commercial deployment solutions.

[0054] Step 3: Simulation Scenario Configuration and Execution The simulation parameters for all scenarios are uniformly configured using the OMNeT++ INI configuration file, including application layer modules, business intensity gradients, simulation duration, and statistical indicators. Simulation experiments are executed sequentially according to three main categories: performance comparison within the experimental group, cross-architecture performance comparison between the experimental group and the control group, and fault self-healing performance verification, ensuring that each scenario runs independently and avoiding data interference.

[0055] Step 4: Data Acquisition and Processing Using the built-in analysis tool .anf file in OMNeT++, the core statistical data of all simulation scenarios are exported, including indicators such as end-to-end latency, total network throughput, and packet loss rate. The arithmetic mean of the results of three repeated experiments is taken, and standardized comparison tables and trend analysis data are generated to complete the final result verification.

[0056] V. Detailed Experimental Data and Results Analysis (I) Latency performance experiments of different layer architectures of the present invention This experiment verifies the latency stability of the starburst matrix architecture of this invention after hierarchical expansion. Detailed data on the average end-to-end latency under different service intensities are shown in the table below: Table 2. Average end-to-end latency of star matrix networks at different levels under different service intensities (unit: ms) ; Results analysis: 1. In low-load scenarios, the architecture of this invention has excellent basic forwarding efficiency. When the service intensity is 0.1, the latency of a single-layer star matrix is ​​only 4.84ms. Even when expanded to a three-layer large-scale architecture, the initial latency is only 5.91ms, with no significant increase in initial overhead. 2. Stable performance under high load scenarios. When the business intensity reaches full load of 1.0, the latency of the single-layer star matrix is ​​only 16.93ms and the latency of the three-layer architecture is only 18.53ms. It increases linearly and steadily with the increase of business intensity, without exponential degradation. 3. The latency increase during hierarchical expansion is controllable. When expanding from a single layer to a three-layer architecture, the node scale doubles, and the latency only increases by 1.6ms under full load. The average latency increase under the full service intensity gradient is only about 10.5%, which has extremely strong stability for scaling up and solves the pain point of the dramatic increase in latency after the scale of traditional architectures expands.

[0057] (II) Throughput performance experiments of different layer architectures of the present invention This experiment verifies the linear scalability of the architecture of this invention. Detailed data on the total network throughput under different service intensities are shown in the table below: Table 3 Total throughput of each level of the star matrix network under different service intensities (unit: Gbps) ; Results analysis: 1. It has near-ideal linear bandwidth growth characteristics. Under any service intensity, the throughput of the three-layer architecture is about 3 times that of the single-layer architecture, and the throughput of the two-layer architecture is about 2 times that of the single-layer architecture. It realizes the linear accumulation of communication bandwidth with the increase of node scale, and completely breaks through the port number limitation of the traditional architecture. 2. No performance bottleneck under high load. Under full load, the throughput of the three-layer star matrix reaches 6.13Gbps. As the business intensity increases from 0.1 to 1.0, the throughput always maintains a steady linear growth. There is no link saturation or sudden drop in throughput that is common in traditional architectures. The system resources are used in a balanced and efficient manner.

[0058] (III) Latency Comparison Experiment between the Invention and the Traditional Spine-Leaf Architecture This experiment verifies the latency performance advantage of this invention compared to traditional mainstream architectures. Detailed latency comparison data for the same node scale is shown in the table below: Table 4. Average end-to-end latency (in ms) of the star matrix and Spine-Leaf network under different service intensities. ; Results analysis: 1. Basic forwarding capabilities have a generational advantage. In low-load (service intensity 0.1) scenarios, the latency of a single-layer star matrix is ​​only 68.5% of that of a Spine-Leaf architecture of the same scale, a latency reduction of 31.5%; the initial latency of a three-layer large-scale architecture is only 66.3% of that of a Spine-Leaf architecture of the same scale, with significant advantages. 2. It has extremely strong resistance to degradation under high load. Under full load (service intensity 1.0) scenarios, the latency of a single-layer star matrix is ​​reduced by 29.8% compared to the Spine-Leaf architecture of the same scale; the latency of a three-layer star matrix is ​​only 18.53ms, compared to 35.20ms of the Spine-Leaf architecture of the same scale, a latency reduction of 47.4%, which solves the problem of increased latency caused by core switch congestion under high load in traditional architecture; 3. The larger the scale, the more obvious the advantages. When the traditional Spine-Leaf architecture is scaled from small to large scale, the full-load latency increases from 24.12ms to 35.20ms, an increase of 45.9%. However, when the present invention is scaled from a single layer to a three-layer architecture, the full-load latency only increases from 16.93ms to 18.53ms, an increase of 9.4%. It has an overwhelming performance advantage in ultra-large-scale data center scenarios.

[0059] (iv) Throughput comparison experiment between the present invention and the traditional Spine-Leaf architecture This experiment verifies the throughput and scalability advantages of the present invention compared to traditional mainstream architectures. Detailed throughput comparison data under the same node scale is shown in the table below: Table 5. Total throughput of the Star Matrix and Spine-Leaf network under different service intensities (unit: Gbps) ; Results analysis: 1. Throughput performance has achieved a leapfrog improvement. Under full load, the throughput of a single-layer star matrix is ​​43.7% higher than that of a Spine-Leaf architecture of the same scale; the throughput of a three-layer star matrix reaches 6.13Gbps, which is 250.3% higher than that of a Spine-Leaf architecture of the same scale, achieving a leap in performance by orders of magnitude. 2. Completely solves the port bottleneck of traditional architecture. The throughput of traditional Spine-Leaf architecture is limited by the number of downlink ports of Spine switches. Even if the node scale is increased by 3 times, the full-load throughput only increases from 1.42Gbps to 1.75Gbps, an increase of only 23.2%. In contrast, the throughput of the architecture of this invention increases linearly with the node scale. From a single layer to a three-layer architecture, the throughput increases from 2.04Gbps to 6.13Gbps, an increase of 200.5%, perfectly adapting to the bandwidth growth requirements of large-scale data centers. 3. Resource scheduling efficiency is significantly improved. This invention achieves full utilization of resources in FOC and FSO links through differentiated scheduling of elephant flow and ant flow. Under the full service intensity gradient, the throughput maintains linear growth without performance saturation trend. In contrast, the traditional Spine-Leaf architecture shows a significant slowdown in throughput growth after the service intensity exceeds 0.7, resulting in a bottleneck of link saturation.

[0060] (v) Specialized experiment on fault self-healing performance This experiment verifies the link fault self-healing capability of the present invention. The test environment is a three-layer 3×3 star matrix network. Under full-load service scenarios, 10 FOC main links are randomly interrupted. The core self-healing indicators are tested, and detailed data are shown in the table below: Table 6. Test data on the fault self-healing performance of the star matrix network. ; Results Analysis: This invention achieves millisecond-level seamless switching of link failures through the local self-healing ring structure of the starburst unit, heterogeneous link redundancy backup, and BFS fast rerouting algorithm. Even if multiple links fail concurrently, it can complete 100% self-healing, and the continuity and stability of service transmission are not significantly affected. This solves the problems of low collaborative efficiency and insufficient reliability guarantee of heterogeneous links in the existing hybrid architecture.

[0061] The above embodiments are merely preferred embodiments of the present invention and should not be construed as limiting the scope of protection of the present invention. Any non-substantial changes and substitutions made by those skilled in the art based on the present invention shall fall within the scope of protection claimed by the present invention.

Claims

1. A wired-wireless hybrid interconnected star-matrix data center network architecture, characterized in that, It includes several modular starburst units, and all starburst units form a three-dimensional starburst matrix network through horizontal matrix expansion and vertical hierarchical expansion. Each of the starburst units is a topology structure composed of three interwoven nodes, including one central wired server node in the innermost layer, six wireless server nodes in the middle layer, and six peripheral wired server nodes in the outermost layer; wherein, the six wireless server nodes are arranged in a regular hexagon, the central wired server node is deployed at the center of the regular hexagon, and the six peripheral wired server nodes are deployed on the periphery of the regular hexagon. The central wired server node, wireless server node, and peripheral wired server nodes are connected to form a highly redundant quasi-fully connected local communication network through wired fiber optic communication FOC links and wireless free space optical communication FSO links; wherein, the wired server nodes are equipped with FOC communication modules and the wireless server nodes are equipped with FSO communication modules. In the horizontal direction, multiple star-shaped units are arranged in an N×M array. Adjacent star-shaped units are interconnected through the FSO link of the wireless server node to establish a torus topology interconnection, realizing point-to-point data penetration between adjacent units. In the vertical direction, the star-shaped matrix network supports N-layer deep expansion. Wireless server nodes at corresponding positions in different layers establish cross-layer communication channels through the FSO link. The star matrix network is a decentralized architecture, with each server node integrating routing and forwarding functions. The communication bandwidth increases linearly with the increase in the number of nodes. Each server node maintains the same connectivity with its neighboring nodes within its own star unit, forming a symmetrical bidirectional ring-shaped self-healing ring structure, which is used for path redirection and local fault self-healing in the event of link failure.

2. The wired-wireless hybrid interconnected star matrix data center network architecture according to claim 1, characterized in that, Within the starburst unit, the central wired server node communicates with six wireless server nodes via FOC links; the six wireless server nodes communicate with each other via FSO links in a ring topology; and each wireless server node communicates with its corresponding peripheral wired server node via an FOC link.

3. The wired-wireless hybrid interconnected star matrix data center network architecture according to claim 1, characterized in that, The FOC link is used to carry the first type of service flow that exceeds the preset data packet size threshold, and the FSO link is used to carry the second type of service flow that is smaller than the preset data packet size threshold. At the same time, the FSO link serves as a redundant backup link for the FOC link.

4. The wired-wireless hybrid interconnected star matrix data center network architecture according to claim 1, characterized in that, The Star Matrix Network pre-calculates the shortest path across the entire network based on the Breadth-First Search (BFS) algorithm, generating a static routing map table for the entire network. This static routing map table assigns a unique point-to-point subnet address to each node, thereby reducing the computational overhead of network initialization and topology reconstruction.

5. The wired-wireless hybrid interconnected star matrix data center network architecture according to claim 1, characterized in that, When any single link within the Starburst Unit suffers physical damage, or when the backbone FOC link experiences congestion / physical damage, the system quickly redirects to a backup loop path or a backup path based on the FSO link using the BFS algorithm, achieving seamless data switching and network self-healing.

6. A network scheduling method for a starburst matrix data center with wired and wireless hybrid interconnection, characterized in that, Based on the starburst matrix data center network architecture described in any one of claims 1-5, the method includes the following steps: S1. Topology construction steps: Build a star matrix network composed of several star units, complete the deployment of heterogeneous links within the star units, and the horizontal and vertical interconnection between star units. S2. Traffic classification steps: Identify the data packet size of the business flow in real time, and classify the business flow into the first type of business flow and the second type of business flow based on the preset data packet size threshold; S3. Link matching and scheduling steps: Assign the first type of service flow to the FOC link for transmission, and assign the second type of service flow to the FSO link for transmission; S4. Routing and Self-Healing Guarantee Steps: Based on the topological characteristics of the star matrix network, path planning and redundancy backup are completed through the BFS algorithm. When the main link fails or becomes congested, the system quickly switches to the backup path to complete data transmission.

7. The network scheduling method for a star-matrix data center with wired and wireless hybrid interconnection according to claim 6, characterized in that, The S1 topology construction steps specifically include: S101. Constructing the basic starburst unit: Deploy 6 wireless server nodes arranged in a regular hexagon, deploy a central wired server node at the center of the regular hexagon, and deploy 6 peripheral wired server nodes around the regular hexagon to establish a quasi-full connection link between FOC and FSO within the unit. S102, System-level expansion deployment: Arrange multiple starburst units horizontally in an N×M array, and construct a horizontal toroidal topology through the FSO links of the wireless server nodes; replicate the single-layer starburst matrix network, and connect the wireless server nodes at corresponding positions of different levels through FSO links to complete the vertical hierarchical expansion and construct a three-dimensional matrix communication architecture.

8. The network scheduling method for a star-matrix data center with wired and wireless hybrid interconnection according to claim 6, characterized in that, In the S2 traffic classification step, the preset data packet size threshold is 1MB. Service flows with data packet size ≥ 1MB are defined as the first type of service flow, namely the elephant flow, and service flows with data packet size < 1MB are defined as the second type of service flow, namely the ant flow.

9. The network scheduling method for a star-matrix data center with wired and wireless hybrid interconnection according to claim 6, characterized in that, In the S4 routing and self-healing guarantee step, the shortest path of the entire network is pre-calculated using the BFS algorithm to generate a static route mapping table and an XML-formatted static route table, completing the path planning in the network initialization phase, and ensuring that data packets are transmitted along the shortest path with the fewest hops.

10. The wired-wireless hybrid interconnected star matrix data center network scheduling method according to claim 6, characterized in that, In the S4 routing and self-healing assurance steps, when the primary data link detects congestion or physical failure, it quickly redirects to the backup redundant path based on the BFS algorithm to achieve load balancing and network self-healing, ensuring the continuity of service transmission.

Images