A method for constructing a hybrid optical interconnection data center network architecture based on a ternary recursive clustering mechanism

Abstract

The application provides a hybrid optical interconnection data center network architecture based on a ternary recursive clustering mechanism, comprising: constructing a bottom layer hexagonal wired optical network in a data center, and taking the bottom layer hexagonal wired optical network as a basic physical layer for data transmission; dividing the physical network of the basic physical layer into multiple sparse logical layers through a ternary recursive clustering algorithm; constructing an upper layer FSO interconnection network based on an adaptive distance threshold; and deploying a static routing strategy between the logical layers and the upper layer FSO interconnection network based on a shortest path. The architecture takes a logical hexagonal topology as a basis, introduces a ternary recursive clustering mechanism to hierarchically group the computing units, thereby guiding the construction and mapping of the FSO enhanced link, orderly replacing part of the wired links under the premise of ensuring the basic connectivity, reducing the wiring complexity and reducing the communication hop count and time delay.

Description

Technical Field

[0001] This invention relates to the field of data center network communication technology, and in particular to a method for constructing a hybrid optical interconnect data center network architecture based on a ternary recursive clustering mechanism. Background Technology

[0002] With the rapid development of technologies such as cloud computing, artificial intelligence, and big data analytics, traffic within data centers (especially east-west traffic) is growing exponentially. This demands that data center networks (DCNs) possess higher bandwidth, lower latency, and more flexible scalability. Currently, mainstream data center networks widely adopt the Fat-Tree architecture. This type of architecture achieves good split-pass bandwidth through a large number of switching devices and wired fiber optic interconnections. However, with the expansion of data center scale and the continuous growth of east-west traffic, its engineering deployment and operation and maintenance face prominent problems such as extremely high cabling complexity, limited scalability, and multi-hop latency bottlenecks. To ensure full connectivity and high bandwidth, a large number of wired links need to be configured, resulting in dense cabling within the server room, difficult cable management, and high deployment and maintenance costs. At the same time, the scale of this architecture is strictly limited by the number of switch ports, and incremental expansion or partial modification often requires a large range of link adjustments, resulting in insufficient flexibility. In addition, communication across racks or pods requires forwarding through multiple levels of switching devices, increasing the end-to-end path hop count, making it difficult to further reduce latency, and easily forming bottleneck links and hotspots.

[0003] In recent years, the introduction of free-space optical communication (FSO) or millimeter-wave technology to construct hybrid data center networks has been proposed. FSO technology uses laser beams to transmit data through the air, offering advantages such as no cabling required, high bandwidth, and flexible deployment, making it considered an ideal solution to the "flying wires" problem in data centers. However, existing hybrid data center networks typically only supplement existing wired networks with a small number of direct FSO links, lacking systematic topology planning and scalable packet / scheduling mechanisms to coordinate with the underlying wired network. This makes it difficult to systematically replace some wired links at the architectural level and control the number of links, port resource usage, and space deployment constraints. Simultaneously, wired links still dominate within the data center, resulting in high cable density, high cabling management, maintenance, and expansion costs. Cross-regional communication still requires multi-hop forwarding, limiting end-to-end latency and throughput performance. Furthermore, existing technologies also have the following drawbacks: 1. Extremely difficult cable management: To maintain high-bandwidth interconnects, traditional data center architectures require the laying of massive amounts of fiber optic cables. These cables are not only costly but also occupy a large amount of space in the server room, resulting in messy cabling. 2. Limited scalability: The scale of traditional data center architecture is strictly limited by the number of ports on the switch, making it difficult to support flexible modular and smooth expansion; 3. Insufficient routing optimization: Unable to provide differentiated low-latency "direct" paths based on the logical clustering relationship of network nodes, resulting in low efficiency in cross-level communication. Summary of the Invention

[0004] The purpose of this invention is to provide a method for constructing a hybrid optical interconnect data center network architecture based on a ternary recursive clustering mechanism. By constructing a logical hexagonal topology and combining it with the ternary recursive clustering mechanism, the complexity of cabling, the number of communication hops, and the latency can be reduced.

[0005] To achieve the above objectives, the present invention provides the following solution: A method for constructing a hybrid optical interconnect data center network architecture based on a ternary recursive clustering mechanism includes: A hexagonal wired optical network is built inside the data center, and this hexagonal wired optical network is used as the basic physical layer for data transmission. The physical network of the basic physical layer is divided into multiple sparse logical levels using a ternary recursive clustering algorithm; An upper-layer FSO interconnection network is constructed based on an adaptive distance threshold. Static routing strategies are deployed between the logical layer and the upper-layer FSO interconnection network based on the shortest path.

[0006] Optionally, a hexagonal wired optical network is constructed within the data center, and this hexagonal wired optical network serves as the basic physical layer for data transmission, including: The data center server racks are divided into several basic computing units; each basic computing unit consists of 7 standard racks with identical structures. The seven standard server racks are physically arranged in a hexagonal structure; the hexagonal structure has one standard server rack in the center and six standard server racks at the edges. The underlying hexagonal wired optical network is formed by connecting standard cabinets located at the edge of adjacent basic computing units with optical fibers; the underlying hexagonal wired optical network is honeycomb-shaped.

[0007] Optionally, the physical network of the basic physical layer is divided into multiple sparse logical levels using a ternary recursive clustering algorithm, including: The physical network of the basic physical layer is clustered and labeled, and the basic computing units are mapped to the state set to obtain state labels; Group basic computational units with the same status label into logical clusters; Repeat the clustering labeling and merging operations within the logical cluster until the number of nodes in the logical cluster meets the termination condition, thus obtaining the logical hierarchy.

[0008] Optionally, an upper-layer FSO interconnection network is constructed based on an adaptive distance threshold, including: Deploy free-space optical communication transceivers on top of a standard cabinet; Calculate the physical distance between all basic computing units within the logical cluster, and determine the minimum distance within the logical cluster; The minimum distance is used as the connection threshold for free-space optical communication transceivers, and free-space optical communication direct links are established only between basic computing units whose physical distance does not exceed the connection threshold within the logical cluster, resulting in the upper-layer FSO interconnection network.

[0009] Optionally, a static routing policy is deployed between the logical layer and the upper-layer FSO interconnection network based on the shortest path, including: Construct a network topology model based on the logical hierarchy and the upper-layer FSO interconnection network; Initialize the source nodes in the network topology model and determine the priority queue; In each iteration, the node with the smallest distance in the priority queue is taken out and determined as the determined set; The system performs backtracking and reverse ordering on the established set to obtain the shortest path, and then generates a static routing strategy based on the shortest path.

[0010] According to specific embodiments provided by the present invention, the following technical effects are disclosed: The method for constructing a hybrid optical interconnect data center network architecture based on a ternary recursive clustering mechanism provided by the present invention includes: constructing a bottom-level hexagonal wired optical network within the data center, and using the bottom-level hexagonal wired optical network as the basic physical layer for data transmission; dividing the physical network of the basic physical layer into multiple sparse logical levels using a ternary recursive clustering algorithm; constructing an upper-level FSO interconnection network based on an adaptive distance threshold; and deploying a static routing strategy between the logical levels and the upper-level FSO interconnection network based on the shortest path. This architecture, based on a logical hexagonal topology, introduces a ternary recursive clustering mechanism to hierarchically group computing units, thereby guiding the construction and mapping of FSO enhanced links. While ensuring basic connectivity, it orderly replaces some wired links, reducing cabling complexity and communication hop count and latency. Attached Figure Description

[0011] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 This is a flowchart of the method for constructing a hybrid optical interconnect data center network architecture based on a ternary recursive clustering mechanism according to the present invention. Figure 2This is a schematic diagram of the overall large-scale hybrid optical interconnect network of the present invention; Figure 3 This is a schematic diagram of Layer 0 of the clustering process of this invention; Figure 4 This is a schematic diagram of Layer 1 of the clustering process in this invention; Figure 5 This is a time delay comparison analysis diagram of an embodiment of the present invention under different load intensities; Figure 6 This is a comparative analysis chart of throughput under different load intensities in an embodiment of the present invention. Detailed Implementation

[0013] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0014] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0015] like Figure 1 As shown, this invention provides a method for constructing a hybrid optical interconnect data center network architecture based on a ternary recursive clustering mechanism, comprising: Step 100: Construct a bottom-layer hexagonal wired optical network inside the data center, and use the bottom-layer hexagonal wired optical network as the basic physical layer for data transmission; Step 200: Divide the physical network of the basic physical layer into multiple sparse logical levels using the ternary recursive clustering algorithm; Step 300: Construct the upper-layer FSO interconnection network based on the adaptive distance threshold; Step 400: Deploy static routing policies between the logical layer and the upper-layer FSO interconnection network based on the shortest path.

[0016] Specifically, such as Figure 2 As shown, the implementation process of step 100 includes: The data center server racks are divided into several basic computing units (Pods), and each Pod consists of 7 standard racks with the same structure. The seven standard server racks are physically arranged in a hexagonal structure; the hexagonal structure has one standard server rack in the center and the other six standard server racks at the edges. By connecting standard racks located at the edge of the Pod with fiber optic cables, a bottom-layer hexagonal wired optical network covering the entire data center is formed, namely a cellular basic wired network, which ensures basic full connectivity of all nodes, with the transmission rate of the wired fiber optic links configured at 10Gbps.

[0017] Preferably, the physical network of the basic physical layer is divided into multiple sparse logical levels using a ternary recursive clustering algorithm, including: The physical network of the basic physical layer is clustered and labeled, and the basic computing units are mapped to the state set to obtain state labels; Group basic computational units with the same status label into logical clusters; Repeat the clustering labeling and merging operations within the logical cluster until the number of nodes in the logical cluster meets the termination condition, thus obtaining the logical hierarchy.

[0018] In the specific implementation process, step 200 first uses a ternary labeling method (graph partitioning / graph coloring, 3-color labeling) to perform the first round of clustering labeling on the entire network graph of the cellular basic wired network, mapping all Pods to the state set [0,1,2]. The clustering constraint is that physically adjacent Pods must have different state labels. Then, Pods with the same state label are grouped into the same logical cluster (e.g., all Pods labeled 0 are grouped into cluster V0). Finally, the clustering labeling process is repeated within each logical cluster Vi until the number of nodes within the cluster meets a preset termination condition, generating a hierarchical logical view of the network, thus obtaining the logical hierarchy. Figure 3 As shown, Layer 0 in the hierarchical logical view is the original network graph G0, with nodes being Pods and edges representing physical adjacency relationships. Figure 4 As shown, the k-th layer (Layerk, k>=1) is generated based on the (k-1)-th layer. The generation process is as follows: for each logical cluster Vi in the (k-1)-th layer, take the Pods within the cluster and their adjacent relationships to form a subgraph Gi. Perform a clustering labeling and logical cluster merging operation on Gi to divide Vi into several sub-clusters. Then, summarize all the sub-clusters of Vi to obtain Layerk. Finally, repeat the process until the number of nodes within the cluster is less than or equal to a preset threshold, and generate Layer1, ..., Layerk. Figure 3 and Figure 4 Hexagonal structures represented by the same shape (such as circles, triangles, and squares) belong to the same logical cluster.

[0019] Preferably, constructing an upper-layer FSO interconnection network based on an adaptive distance threshold includes: Deploy free-space optical communication transceivers on top of a standard cabinet; Calculate the physical distances between all basic computing units in the computing logic cluster, and determine the minimum distance within the logic cluster; Take the minimum distance as the connection threshold of the free-space optical communication transceiver device, and establish a free-space optical communication direct link only between the basic computing units whose physical distance does not exceed the connection threshold within the logic cluster, so as to obtain an upper-layer FSO interconnection network.

[0020] In the specific implementation process, in step 300, free-space optical communication (FSO) transceivers are first deployed on the top of the cabinet. Then, for each logic cluster at the k-th layer, the physical distance between all Pod pairs in the cluster is calculated, and the minimum distance δ in the cluster is found, which is specifically as follows: each Pod has its own physical coordinate (x,y), for any two Pods in the cluster i and Pod j , the physical distance d ij is calculated according to the Euclidean distance, and the calculation formula is: , after calculating for all Pod pairs (i<j) in this logic cluster , the minimum value is taken as δ. Finally, the FSO connection threshold Threshold=δ is determined. In the same cluster, FSO direct links are established only for Pod pairs whose physical distance ≤Threshold, forming an ordered spatial optical network, that is, the upper-layer FSO interconnection network, wherein the transmission rate of the free-space optical communication (FSO) link is configured as 10Gbps.

[0021] Preferably, the static routing strategy between the logical hierarchy and the upper-layer FSO interconnection network is deployed based on the shortest path, including: Construct a network topology model according to the logical hierarchy and the upper-layer FSO interconnection network; Initialize the source node in the network topology model and determine the priority queue; In each iteration process, the node with the smallest distance in the priority queue is taken out and determined as the determined set; Backtracking and reverse processing are performed on the determined set to obtain the shortest path, and a static routing strategy is generated according to the shortest path.

[0022] In the specific implementation process, step 400 uses Dijkstra's algorithm to determine the shortest path in a hybrid network topology containing underlying wired fiber and upper-layer FSO links, and generates a static routing table accordingly. The specific steps are as follows: the hybrid network topology is abstracted as a graph G=(V,E), where nodes are network devices or hosts, and edges are link connections. The minimum hop count strategy is used to set the weight of any link edge to 1. For the source node s, dist[s]=0 is set, and for the other nodes v, dist[v]=+∞. The predecessor array prev[] and the determined set visited[] are maintained, and a priority queue is extracted according to dist from smallest to largest, thereby completing the initialization of the source node. Here, dist[s]=0 means that when executing Dijkstra's algorithm for initialization, the initial minimum hop count (distance) from the source node s (starting point) to itself is set to 0, which serves as the reference starting point for subsequent iterations to calculate the shortest route path. Then, in each round, the node u with the smallest dist value is taken from the queue, and visited[u] is set to true. Since the edge weight is non-negative (1 here), when u is taken out and set to determined, dist[u] is the final minimum hop count from s to u. Then, the adjacent nodes v of u are traversed. If dist[u] + 1 is less than dist[v], dist[v] and prev[v] are updated to u, and v is added to the priority queue. Finally, for any destination node t, when t is taken out of the priority queue and visited[t] is set to true, dist[t] is finally determined. The shortest path is obtained by backtracking from t along prev[] to s in reverse order. If there are paths with the same number of hops, the shortest path is selected according to the deterministic rule (the next hop ID / interface number is the smallest), and a static routing table is directly generated based on the shortest path.

[0023] In one specific embodiment, hybrid optical interconnect networks of different scales (N=3, N=5, and N=7, with 259, 637, and 1183 server nodes respectively) were constructed based on the OMNeT++ simulation platform and the INET framework for comparative analysis and discussion of network scalability. The wired optical fiber connections used 10Gbps, and the FSO (Fiber Optic Sockets) operated at 60GHz with a transmission rate of 10Gbps. The simulation experiment in this embodiment was set to a total duration of 500 seconds, and all packet transmission intervals were determined by an exponentially distributed random function with a mean parameter of 480 milliseconds (exponential(480ms)).

[0024] Then, a network model is built using OMNeT++'s NED file. Modules are dragged and dropped graphically in the Design interface, or the network structure is defined using the NED language in the Source interface to build the topology. Next, parameters such as simulation time, packet size, and transmission interval are set using the INI file, configuring UDPBasicApp as the sender and UDPSink as the receiver to generate a static routing table. Key data such as end-to-end delay and throughput are recorded, and the simulation results are exported using an OMNeT++ analysis file (.anf).

[0025] To facilitate data statistics, normalized simulation data for service intensity was used. The relationship between service intensity and data packet size is shown in Table 1. Service intensity ranges from 0.1 to 1.0, with 10 gradients representing service levels; higher values ​​indicate busier service. All service intensities use an exponentially distributed random function with a mean parameter of 480 milliseconds. Data packet size varies with service intensity from 0.1 to 1.0, corresponding to 300k, 600k, 900k, 1200k, 1500k, 1800k, 2100k, 2400k, 2700k, and 3000k bytes respectively; higher service intensity results in larger data packets.

[0026] Table 1 Simulation data with normalized load intensity set

[0027] Latency refers to the time difference between the transmission of data from the sender to the destination, and it is one of the key indicators affecting network performance. For example... Figure 5 As shown, simulation data results are used to summarize the latency changes of three different networks under different network intensities, with the vertical axis representing latency and the horizontal axis representing the normalized load intensity increasing sequentially from left to right. The simulation verifies the latency changes of the three different networks under different load intensities. In the network architecture with N=3, the latency can reach approximately 10.11ms at the lowest load intensity of 0.1. As the load level gradually increases, the average end-to-end latency also increases accordingly. When the load level is 1.0, the average end-to-end latency reaches its highest value of 93.31ms. In the large-scale network architecture with N=7, the highest load intensity is 104.41ms. Although the average end-to-end latency increases with the increase of the load level, it shows a stable linear growth trend overall, without an explosive increase due to the increase in load level. It is particularly noteworthy that when expanding from N=3 to N=7, the network size increases significantly, but the latency only increases slightly (e.g., only about 11% at the highest load). This demonstrates that the network architecture has good stability and scalability.

[0028] Throughput is a core metric for measuring network performance and transmission efficiency, reflecting the network's ability to transmit data in practical use. For example... Figure 6 As shown, simulation data results summarize the throughput changes of three different networks under different network intensities, presented as a line graph. The vertical axis represents throughput, and the horizontal axis represents the normalized load intensity, increasing sequentially from left to right. When the load intensity decreases, the network load lightens, contention decreases, congestion lessens, and fewer data packets are transmitted, resulting in a corresponding decrease in actual throughput because the total amount of data transmitted in the network decreases. Under low load intensity, link performance may be relatively stable, with less interference and congestion, leading to a higher success rate for individual data packet transmission. As the load intensity increases, the amount of data transmitted in the network increases, and throughput increases accordingly. However, as the load intensity continues to increase, network resources may become congested, leading to problems such as data packet loss, thus limiting the growth of throughput. Figure 6 It can be seen that, under different network architectures (N=3, N=5, N=7), the throughput shows a near-linear growth trend as the load intensity gradually increases. The network with N=7 achieves a throughput of approximately 591.29 Gbps under the highest load, significantly higher than the 129.44 Gbps of N=3. This indicates that as the network scales up, the newly added nodes and links effectively contribute bandwidth, and the core layer does not become a bottleneck. Therefore, this network architecture demonstrates excellent data transmission capabilities.

[0029] The beneficial effects of this invention are as follows: 1) Through the cellular physical topology of hexagonal Pods and the logical layering of ternary recursive clustering, it supports "plug and play" smooth expansion. It can be integrated into the network simply by connecting adjacent edge cabinets. Unlike Fat-Tree, it does not require the reconstruction of the core switching layer, thus realizing the orderly construction and efficient interconnection of large-scale networks. Combined with the dual protection of the FSO layer, it significantly improves the fault tolerance of the network. 2) By introducing adaptive FSO links as a supplement to the wired network, some core layer optical fibers were replaced, reducing a large number of long-distance cables, alleviating cable congestion in the data center, effectively reducing cabling complexity and cost, and significantly improving the network's flexibility, scalability and communication efficiency. While ensuring high throughput and low latency performance, it also supports flexible inter-node communication, providing strong technical support for the development of large-scale data center networks. 3) The unique recursive clustering and adaptive threshold mechanism avoids spatial intersection and occlusion of FSO beams from the topology design level, realizing the feasibility of large-scale FSO networking.

[0030] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0031] Specific examples have been used to illustrate the principles and implementation methods of this invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of this invention. Furthermore, those skilled in the art will recognize that, based on the ideas of this invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this invention.

Claims

1. A method for constructing a hybrid optical interconnect data center network architecture based on a ternary recursive clustering mechanism, characterized in that, include: A bottom-layer hexagonal wired optical network is constructed inside the data center, and the bottom-layer hexagonal wired optical network is used as the basic physical layer for data transmission; The underlying hexagonal wired optical network is composed of multiple basic computing units; The physical network of the basic physical layer is divided into multiple sparse logical levels using a ternary recursive clustering algorithm; The specific steps include: clustering and labeling the physical network of the basic physical layer and assigning three types of state labels, with adjacent basic computing units having different state labels; The basic computing units with the same state label are grouped into logical clusters; Repeat the clustering labeling and merging operations within the logical cluster until the number of nodes in the logical cluster meets the termination condition, thus obtaining the logical hierarchy; The upper-layer FSO interconnection network is constructed based on an adaptive distance threshold. Specific steps include: deploying free-space optical communication transceivers on the top of a standard rack; calculating the physical distance between all basic computing units within the logical cluster and determining the minimum distance within the logical cluster; using the minimum distance as the connection threshold for the free-space optical communication transceivers; and establishing direct free-space optical communication links only between the basic computing units within the logical cluster whose physical distance does not exceed the connection threshold, thereby obtaining the upper-layer FSO interconnection network. Static routing strategies are deployed between the logical layer and the upper-layer FSO interconnection network based on the shortest path.

2. The method for constructing a hybrid optical interconnect data center network architecture based on a ternary recursive clustering mechanism according to claim 1, characterized in that, A bottom-layer hexagonal wired optical network is constructed within the data center, and this bottom-layer hexagonal wired optical network serves as the basic physical layer for data transmission, including: The data center server rack is divided into several basic computing units; each basic computing unit consists of seven identical standard racks. The seven standard server racks are physically arranged in a hexagonal structure; the hexagonal structure consists of one standard server rack at the center and six standard server racks at the edges. The underlying hexagonal wired optical network is formed by connecting the standard racks located at the edge of the adjacent basic computing units via optical fibers; the underlying hexagonal wired optical network is honeycomb-shaped.

3. The method for constructing a hybrid optical interconnect data center network architecture based on a ternary recursive clustering mechanism according to claim 2, characterized in that, Deploying static routing policies between the logical layer and the upper-layer FSO interconnection network based on the shortest path includes: Construct a network topology model based on the logical hierarchy and the upper-layer FSO interconnection network; Initialize the source nodes in the network topology model and determine the priority queue; In each iteration, the node with the smallest distance in the priority queue is extracted and determined as the determined set; The determined set is backtracked and reversed to obtain the shortest path, and the static routing strategy is generated based on the shortest path.

Images