Photoelectric hybrid interconnection system for large-scale intelligent computing center

By introducing optoelectronic hybrid interconnection systems into the data center interconnection architecture, combining the advantages of electrical switches and optical switches, the bandwidth limitation and delay bottleneck problems during large model training are solved, and efficient and flexible data exchange and optical path connection are achieved, meeting the high scalability needs of intelligent computing centers.

CN120200990AActive Publication Date: 2025-06-24BEIJING UNIV OF POSTS & TELECOMM
View PDF 4 Cites 0 Cited by

Patent Information

Application Number
CN202510346818.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2025-03-21
Publication Date
2025-06-24
Estimated Expiration
2045-03-21

AI Technical Summary

Technical Problem

The existing data center interconnection architecture has shortcomings in the high bandwidth, low latency and high scalability requirements when training large models, including bandwidth limitations, delay bottlenecks, large energy consumption and switching capacity limitations.

Method used

A hybrid optoelectronic interconnection system for large-scale intelligent computing centers was designed. By combining in-group switching systems and inter-group switching systems, the advantages of electrical switches and optical switches are utilized to achieve flexible data exchange and efficient optical path connections.

Benefits of technology

It realizes a topological structure that supports large-scale connections under low-cost conditions, provides flexible exchange methods and large-capacity and low-latency data transmission channels, meets the efficient data processing needs of smart computing centers, and has good scalability and smooth upgrade capabilities.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN120200990A_ABST
    Figure CN120200990A_ABST
Patent Text Reader

Abstract

The invention provides a photoelectric hybrid interconnection system for a large-scale intelligent computing center. The photoelectric hybrid interconnection system comprises an intra-group switching system with a plurality of intra-group switching units and an inter-group switching system comprising an optical switch, each intra-group switching unit comprises a first number of electric switches and at least one server, each server is at least provided with a first number of computing nodes corresponding to the first number of electric switches, the computing nodes are mutually connected, and each electric switch is connected with the corresponding computing node through a downlink port; each optical switch in the inter-group switching system comprises a plurality of pairs of double-side ports and is used for providing a plurality of optical path connections between different-side ports of the double-side ports through mapping configuration of the different-side ports. Two different-side ports connected with each optical path of the same switch can be respectively connected with uplink ports of the electric switches of different intra-group switching units based on the connection relationship between the optical switch ports and the electric switch ports, and the mapping configuration enables inter-group optical path switching channels to be formed between the different intra-group switching units.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of communication technologies, and particularly to an optical-electric hybrid interconnection system for a large-scale intelligent computing center. Background Art

[0002] With the continuous development of large model technologies, accelerating the development of the artificial intelligence industry and building a strong computing power data center (or called intelligent computing center) that supports large models have become important research and development directions in the future. The key characteristics of large model training tasks are ultra-large amounts of data and distributed computing methods, which pose new requirements for the networking of intelligent computing centers carrying such new services: (1) Facing large-scale connections: Large model training services require the participation of a large number of computing nodes, and the networking of intelligent computing centers urgently needs to support the efficient communication and collaboration of tens of thousands of computing nodes; (2) Supporting large-capacity transmission: Large model training tasks require training a large number of data sets, and at the same time, the parameters of the model are numerous. The network of the intelligent computing center needs to improve the capacity and rate of channel transmission; (3) Ensuring low-latency synchronization: Frequent parameter synchronization and data interaction between computing nodes make communication latency a bottleneck for the service. The networking of intelligent computing centers needs to ensure that the latency of data interaction is low, so as to ensure the high efficiency of large-scale cluster training tasks.

[0003] Traditional data center architectures based on electrical switching have problems such as bandwidth limitations and latency bottlenecks, and it is difficult to meet the explosive growth of traffic and high synchronization requirements during large model training. Data center architectures based on electrical packet switching include Spine-Leaf networks, DragonFly networks, and Tours networks. In a Spine-Leaf network, all Leaf layer switches are equally connected to all Spine layer switches, forming a fully interconnected network. Although the Spine-Leaf network has excellent bisection bandwidth performance, its network diameter is large and the communication hops between intra-cluster and inter-cluster nodes are inconsistent. During the AI training process, parameter synchronization needs to wait for the last node to complete before the next iteration can be carried out, so it may lead to waste of computing resources. The DragonFly network completes the full connection between groups through direct connection. Although it can effectively reduce latency, it is difficult to meet scalability and ensure the communication performance of various synchronization algorithms under the service requirements of a new type of intelligent computing center. The Torus architecture is a multi-dimensional topological structure with advantages such as simple structure and multiple paths. However, as the number of network dimensions increases, the number of internal links and delays in the network gradually increase; in addition, the ring structure of the Torus itself will bring complex deadlock problems. These interconnection architectures are difficult to support the explosive growth of internal traffic in the data center during large model training in terms of performance, and there are problems including bandwidth limitations, high energy consumption, latency bottlenecks, and switching capacity limitations. It is necessary to design a more efficient data center interconnection solution to solve the bottlenecks existing in the electrical switching network.

[0004] Compared with electrical packet switching, optical switching does not parse data packets during the process of forwarding data packets and does not require the installation of transceiver modules for the conversion between optical and electrical signals. In large-scale networks, this feature can effectively reduce the frequent optical-electrical conversion overhead in electrical switching networks and reduce communication latency. However, the prices of optical switching devices and components are relatively expensive, with a large investment and high operation and maintenance difficulty. Therefore, currently, the optical-electrical hybrid switching architecture that replaces some electrical packet switches with optical switches is considered a more suitable architecture for large-scale high-computing-power data centers and has been actually applied. It combines the advantages of electrical packet switching and optical switches and balances performance and cost. For example, both the Helios architecture and the ReSaw architecture adopt a two-layer tree topology. Their access switches (electrical switches) are directly connected to servers and are responsible for local communication. The core switches are composed of two parts: an electrical packet switch and an optical switch. Some uplink ports of the access switches are connected to the electrical packet switch, and some uplink ports are connected to the optical switch. For short-term bursty traffic, the electrical packet switching mode is adopted; for traffic with a long duration and high bandwidth requirements, the optical switching mode is adopted. Such architectures have more switching levels and larger switching latency, but can achieve data aggregation at a lower cost and are very suitable for the traditional data center services with "coexistence of large and small flows, more north-south traffic, and less east-west traffic". However, in the intelligent computing center, it is necessary to complete the arbitrary interconnection of large-scale, same-level nodes, with high requirements for east-west communication, showing the new characteristics of less north-south traffic and more east-west traffic. The traditional optical-electrical hybrid architecture is not applicable. In addition, the traditional optical-electrical hybrid architecture is difficult to achieve flexible reconstruction, with weak scalability and smooth upgrade capabilities of the architecture, and it is difficult to support large-scale cluster interconnection training services, thus making it difficult to efficiently process complex data processing requirements. Therefore, how to meet the requirements of large model training, achieve the interconnection of a large number of GPU cards, build a flat and non-blocking intelligent computing center network architecture, with both deployment cost advantages and the flexible adaptation ability of rigid pipelines has become an important challenge for current intelligent computing centers. Summary of the Invention

[0005] In view of this, embodiments of the present invention provide a novel optical-electrical hybrid interconnection system for large-scale intelligent computing centers to eliminate or improve one or more defects existing in the prior art.

[0006] One aspect of the present invention provides an optical-electrical hybrid interconnection system for large-scale intelligent computing centers. The system includes: an intra-group switching system and an inter-group switching system. The intra-group switching system includes a plurality of intra-group switching units, and the inter-group switching system includes at least one optical switch;

[0007] Each in-group switching unit includes a first number of electrical switches and at least one server. Each server has at least a first number of computing nodes corresponding to the first number of electrical switches. Among them, the first number of computing nodes are interconnected with each other, and each electrical switch is connected to the corresponding computing node among the first number of computing nodes through a downlink port;

[0008] Each optical switch in the inter-group switching system includes multiple pairs of bilateral ports. The multiple pairs of bilateral ports are used to provide multiple optical path connections between different pairs of opposite-side ports through mapping configurations between different pairs of opposite-side ports. For each optical path connection of the same switch, the two opposite-side ports can be respectively connected to the uplink ports of the electrical switches of different in-group switching units based on the connection relationship between the optical switch ports and the electrical switch ports. The mapping configuration enables an inter-group optical path switching channel to be formed between different in-group switching units.

[0009] In some embodiments of the present invention, the first number of computing nodes of the server in the in-group switching unit are interconnected through a high-speed interconnection technology.

[0010] In some embodiments of the present invention, a single-fiber bidirectional optical module is used for interconnection between the in-group switching system and the inter-group switching system for data transmission.

[0011] In some embodiments of the present invention, the number of downlink ports of each electrical switch in the in-group switching unit is greater than or equal to the number of servers in the corresponding in-group switching unit.

[0012] In some embodiments of the present invention, a correspondence is established between the electrical switches in the in-group switching unit and the computing nodes inside the server through the mapping relationship between the electrical switch numbers and the computing node numbers.

[0013] In some embodiments of the present invention, the connection relationship between the optical switch ports and the electrical switch ports includes: the same pair of bilateral ports in the optical switch are respectively connected to different uplink ports of the electrical switches in the same in-group switching unit; the mapping configuration includes establishing a one-to-one mapping relationship between different pairs of opposite-side ports of the bilateral ports.

[0014] In some embodiments of the present invention, the total number of pairs of bilateral ports of the optical switches in the inter-group switching system is greater than or equal to a second number, and at least one optical path connection is established between any two in-group switching units, where the second number is N*(N - 1) / 2; the number of optical switches in the inter-group switching system satisfies the following formula: Alternatively, the total number of pairs of bilateral ports of the optical switches in the inter-group switching system is less than the second number, and optical path connections are established between selected in-group switching units; the number of optical switches in the inter-group switching system satisfies the following formula:

[0015] Among them, N is the number of intra-group switching units in the intra-group switching system, O is the number of optical switches, and n o is the number of pairs of bilateral ports in the o-th optical switch.

[0016] In some embodiments of the present invention, the mapping configuration of the opposite-side ports of the optical switches in the inter-group switching system can be switched according to the service transmission requirements of different intra-group switching units, so as to realize the switching of the connection optical paths.

[0017] In the embodiments of the present invention, the computing node is used to send and receive data information that needs to be interacted; the electric switches of the same intra-group switching unit are used to receive the data information sent from the corresponding computing node and forward it to the connected optical switch of the inter-group switching system, or to receive the data information sent from the connected optical switch of the inter-group switching system and forward it to the corresponding computing node; the optical switch is used to receive the data information from the electric switch and forward the received data information between different intra-group switching units.

[0018] For the optoelectronic hybrid interconnection system for large-scale intelligent computing centers of the present invention, in terms of flexibility, the electric switches of the intra-group switching system provide a flexible switching method for the computing nodes in the same group. At the same time, the inter-group optical switches provide a large-capacity and low-latency data transmission channel for the computing nodes across groups, and can flexibly establish optical paths through a service-oriented dynamic resource scheduling method, providing an adaptable connection and switching ability for services. In terms of construction cost, the invested optical switches have the ability of smooth upgrade. When the number of optical path connections provided by the actually invested optical switches is greater than the basic optical path connection number requirement, the present invention can provide a larger-capacity transmission channel between intra-group switching units; when the number of optical path connections provided by the actually invested optical switches is less than the basic optical path connection number requirement, the flexible switchable ability of the optical switches can be used to construct an adaptable optical path for services. Compared with the network interconnection systems of other data centers, the system of the present invention not only has a low construction cost, but also is more flexible.

[0019] The additional advantages, objects, and features of the present invention will be partially described below, and will become partially apparent to those of ordinary skill in the art after studying the following text, or can be learned from the practice of the present invention. The objects and other advantages of the present invention can be realized and obtained through the structures specifically pointed out in the specification and the drawings.

[0020] Those skilled in the art will understand that the objects and advantages that can be achieved by the present invention are not limited to the above specifically described, and the above and other objects that the present invention can achieve will be more clearly understood according to the following detailed description. Description of the Drawings

[0021] The accompanying drawings described herein are used to provide a further understanding of the present invention, form a part of this application, and do not limit the present invention.

[0022] Figure 1 It is a schematic structural diagram of an intra-group switching unit in an optical and electrical hybrid interconnection system according to an embodiment of the present invention.

[0023] Figure 2 It is a schematic interconnection diagram between computing nodes of a server according to an embodiment of the present invention.

[0024] Figure 3 It is a schematic diagram of the internal optical path connection of an optical switch in an inter-group switching system according to an embodiment of the present invention.

[0025] Figure 4 It is a schematic diagram of the interconnection between an intra-group switching unit of an intra-group switching system and an optical switch of an inter-group switching system in an optical and electrical hybrid interconnection system according to an embodiment of the present invention.

[0026] Figure 5 It is a schematic diagram of the interconnection between an intra-group switching unit of an intra-group switching system and an optical switch of an inter-group switching system in an optical and electrical hybrid interconnection system according to another embodiment of the present invention.

[0027] Figure 6 It is a schematic diagram of the interconnection between an intra-group switching unit of an intra-group switching system and an optical switch of an inter-group switching system in an optical and electrical hybrid interconnection system according to another embodiment of the present invention.

[0028] Figure 7 It is a schematic diagram of the interconnection between an intra-group switching unit of an intra-group switching system and an optical switch of an inter-group switching system in an optical and electrical hybrid interconnection system according to another embodiment of the present invention. Detailed implementation manners

[0029] To make the objectives, technical solutions, and advantages of the present invention clearer and more understandable, the present invention will be further described in detail below in conjunction with the implementation manners and the accompanying drawings. Herein, the illustrative implementation manners of the present invention and their descriptions are used to explain the present invention, but do not limit the present invention.

[0030] Herein, it should also be noted that in order to avoid obscuring the present invention due to unnecessary details, only the structures and / or processing steps closely related to the solution according to the present invention are shown in the accompanying drawings, while other details less related to the present invention are omitted.

[0031] It should be emphasized that the term "including / containing" when used herein refers to the presence of features, elements, steps, or components, but does not exclude the presence or addition of one or more other features, elements, steps, or components.

[0032] Here, it should also be noted that, unless otherwise specified, the term "connection" in this text can refer not only to direct connection but also to indirect connection with intermediaries.

[0033] In the following, embodiments of the present invention will be described with reference to the accompanying drawings. In the drawings, the same reference numerals represent the same or similar components, or the same or similar steps.

[0034] Aiming at the difficulty of existing intelligent computing centers in meeting the requirements of large-scale connection, large bandwidth, and low latency in the training of large models under time-varying conditions, the present invention proposes a novel optical-electrical hybrid interconnection system for large-scale intelligent computing centers to achieve a low-cost topology structure that supports large-scale connections under the rigid pipeline characteristics of optical interconnection. The novel optical-electrical hybrid interconnection system proposed by the present invention consists of an intra-group switching system and an inter-group switching system. The intra-group switching system includes a plurality of intra-group switching units, and each intra-group switching unit includes a server and a plurality of electrical switches (also referred to as electrical packet switches). The server has a plurality of interconnected computing nodes. The inter-group switching system includes one or more optical switches. The electrical switches in the intra-group switching unit provide flexible data synchronization for the computing nodes. The optical switches between groups use the characteristics of large bandwidth and low latency of optical links to provide a rigid channel for the inter-group data exchange of the computing nodes, and at the same time can flexibly adapt the services by using the flexible reconfiguration ability of the optical switches. Preferably, the intra-group switching unit has a plurality of servers, and the plurality of servers are interconnected. The interconnection of the computing nodes within the server is used to improve the data processing and communication efficiency within a single server; the interconnection between the servers can ensure efficient data exchange between the servers within the same group.

[0035] The optical-electrical hybrid interconnection system for large-scale intelligent computing centers provided by the present invention includes an intra-group switching system and an inter-group switching system. The intra-group switching system includes a plurality of intra-group switching units (or referred to as a plurality of groups), and the inter-group switching system includes at least one optical switch.

[0036] Each intra-group switching unit includes a first number of electrical switches and at least one server. Each server has at least a first number of computing nodes corresponding to the first number of electrical switches, and among them, the first number of computing nodes are interconnected. Any electrical switch in each intra-group switching unit is connected to one computing node on all the intra-group servers. For example, a corresponding relationship can be preset so that each electrical switch in the intra-group switching unit is connected to the corresponding computing node among the first number of computing nodes in the intra-group through a downlink port.

[0037] The number of intra-group switching units can be represented as N, where N is a natural number. As an example, N = 5, that is, the intra-group switching system includes 5 intra-group switching units, or in other words, the intra-group switching system is divided into 5 groups, numbered G1, G2, …, G i ,…,G N , Gi denotes the i-th group. Each switching unit within a group includes M electric switches numbered E1, E2, …, E m , …, E M and L servers numbered S1, S2, …, S l , …, S L , where E i denotes the i-th electric switch and S i denotes the i-th server. Each electric switch within the group may include P switching ports, and each server includes M computing nodes numbered C1, C2, …, C m , …, C M , where C i denotes the i-th computing node. Each computing node numbered C i can be respectively connected to the downstream ports of the electric switches numbered E i within the group through network interfaces.

[0038] As Figure 1 shown, M = 8 and L = 32, that is, each switching unit within a group of the in-group switching system includes 8 electric switches (E1 to E8) and 32 servers (S1 to S 32 ), the number of ports P of the electric switch is, for example, 64, and each server includes 8 computing nodes (C1 to C8, Figure 1 in the example of, C1 = 1, …, C8 = 8), the computing nodes can be GPUs inside the server, and the computing nodes within the same service are interconnected. As Figure 2 shown, this connection can be a direct connection or an indirect connection. As an example, the computing nodes within the same service can achieve interconnection communication through high-speed interconnection technology (such as NVLink). In the embodiments of the present invention, any electric switch of 1 in-group switching unit is connected to 1 computing node on all in-group servers.

[0039] The inter-group switching system may include one or more optical switches. Each optical switch includes multiple pairs of bilateral ports (such as n pairs), and the multiple pairs of bilateral ports are used for pre-mapping configuration to provide multiple optical path connections between the opposite-side ports of the bilateral ports through the mapping configuration. The two opposite-side ports of each optical path connection of the same optical switch can be respectively connected to the upstream ports of the electric switches of different in-group switching units, and the mapping configuration enables an inter-group optical path switching channel to be formed between different in-group switching units. In the embodiments of the present invention, the in-group switching system and the inter-group switching system use a single-fiber bidirectional optical module for interconnection to transmit data. In the embodiments of the present invention, the mapping configuration includes establishing a one-to-one mapping relationship between the opposite-side ports of different pairs of the bilateral ports.

[0040] The in-group switching unit and the inter-group optical switch can be respectively used for:

[0041] The computing nodes are used to send and receive data information that needs to be interacted.

[0042] The electric switches of the switching units within the same group are used to receive the data information sent from the corresponding computing nodes and forward it to the connected optical switches of the inter-group switching system, or to receive the data information sent from the connected optical switches of the inter-group switching system and forward it to the corresponding computing nodes.

[0043] The optical switches are used to receive the data information from the electric switches and forward the received data information between different intra-group switching units.

[0044] In the above-mentioned novel optoelectronic hybrid interconnection system for large-scale intelligent computing centers, among the P ports of each electric switch in one intra-group switching unit, L downlink ports are used to connect to the computing nodes, and (P - L) uplink ports are used to interconnect with the inter-group optical switches. In some embodiments of the present invention, the number L of downlink ports of each electric switch in the intra-group switching unit is greater than or equal to the number of servers in the corresponding intra-group switching unit. In addition to the downlink ports connected to the computing nodes in the servers and the uplink ports connected to the optical switches, the electric switches can also be provided with redundant or idle uplink / downlink ports, and the specific number is not limited in the present invention as long as it can meet the connection requirements with the computing nodes and the optical switches.

[0045] In the embodiments of the present invention, the electric switches of the intra-group switching units and the computing nodes inside the servers are corresponding through the mapping relationship between the electric switch numbers and the computing node numbers. The following method can be used to connect any electric switch in one intra-group switching unit to one computing node on all intra-group servers: Map the computing nodes in the intra-group servers S1, S2, S3, …, S L to the intra-group electric switches, and connect the computing node C i to the corresponding electric switch E i through the network interfaces respectively. As an example, a feasible mapping method is to connect the computing nodes with the same number on different servers to the same electric switch through the network interfaces, that is, connect all the computing nodes numbered 1 in 32 intra-group servers to the electric packet switch numbered 1 through the first network interface of each server; all the computing nodes numbered C2 (such as C2 = 2) are respectively connected to the electric packet switch numbered E2 (such as E2 = 2) through the second network interface of each server; and so on until all the computing nodes numbered C i (such as C i = i) are respectively connected to the electric switch numbered E i (such as E i = i) through the i-th network interface of each server. This mapping method is only an example, and the present invention is not limited thereto.

[0046] In an embodiment of the present invention, the inter-group switching system may include O optical switches numbered A1, A2, A3, …, A o , …, A O . Here, A o represents the o-th optical switch. Each optical switch may include multiple pairs of bilateral switching ports. For example, it may include n o pairs of bilateral ports, that is, there are n o ports on each side, and a total of 2n o bilateral ports. n o represents the number of unilateral ports of the o-th optical switch. The number of bilateral ports of different optical switches may be the same or different. As an example, if an inter-group optical switch has n pairs of bilateral ports, the ports on the left side may be represented as P L1 , P L2 , …, P Li , …, P Ln , and the ports on the right side may be represented as P R1 , P R2 , …, P Ri , …, P Rn . Here, P Li represents the i-th port on the left side, and P Ri represents the i-th port on the right side. P Li -P Ri is a pair of bilateral ports.

[0047] In an embodiment of the present invention, the multiple pairs of bilateral ports are used to form multiple groups of bilateral ports through mapping configurations between different pairs of opposite-side ports. These multiple groups of bilateral ports are used to provide multiple optical path connections between different pairs of opposite-side ports. That is, each group of bilateral ports is formed by opposite-side ports from different pairs of ports, and an optical path connection is formed between these opposite-side ports. Each group of bilateral ports of the same optical switch (i.e., the two opposite-side ports of each optical path connection) can be respectively connected to the upstream ports of the electrical switches of different intra-group switching units based on the connection relationship between the optical switch ports and the electrical switch ports. This mapping configuration enables the formation of an inter-group optical path switching channel between different intra-group switching units. As an example, inside the optical switch, if P Li -P Ri represents a pair of bilateral ports, then P Li -P RjIt can be used to represent a group of bilateral ports composed of the i - port on the left and the j - port on the right (i≠j). An optical path connection can be formed between this group of configured bilateral ports, and ports on the same side cannot be connected; at any moment, an optical switch with n port groups can provide n optical path connections. For the convenience of flexible mapping configuration of multiple groups of bilateral ports, in the embodiments of the present invention, preferably, the connection relationship between the optical switch ports and the electrical switch ports can be predetermined or determined in real - time. This connection relationship may include: the same pair of bilateral ports of the same optical switch can be respectively connected to different upstream ports of the electrical switch in the same group - internal switching unit. More specifically, the same pair of bilateral ports of the same optical switch can be connected to different upstream ports of the same electrical switch in the same group - internal switching unit, or can be connected to the upstream ports of different electrical switches in the same group - internal switching unit. In some embodiments of the present invention, in addition to the switching ports connected to the electrical switch, the optical switch can also be provided with redundant or idle switching ports. The specific number is not limited in the present invention, as long as the connection requirements with the electrical switch can be met.

[0048] Figure 3 Shown is an example of the internal optical path connection of an optical switch in the embodiments of the present invention. Figure 3 An optical switch with n = 5 is shown, which includes 2n bilateral ports, 5 ports on the left and 5 ports on the right. The left - and right - hand ports are numbered to match into 5 pairs of bilateral ports, which are respectively: P L1 -P R1 ,P L2 -P R2 ,P L3 -P R3 ,P L4 -P R4 ,P L5 -P R5 。These 5 pairs of bilateral ports form 5 groups of bilateral ports for providing optical path connections through the mapping configuration of different pairs of ports on the opposite sides. More specifically, the port mapping configuration is: these 5 groups of bilateral ports are mapped and configured as P L1 -P R2 ,P L2 -P R3 ,P L3 -P R4 ,P L4 -P R5 ,P L5 -P R1, that is, the left 1 port and the right 2 port form a group, and there is an optical path link between the ports; the left 2 port and the right 3 port form a group, and there is an optical path connection between the ports; the left 3 port and the right 4 port form a group, and there is an optical path connection between the ports; the left 4 port and the right 5 port form a group, and there is an optical path connection between the ports; the left 5 port and the right 1 port form a group, and there is an optical path connection between the ports, providing a total of 5 optical path connections. The optical switch transmits data with the in-group electrical packet switch through a single-fiber bidirectional optical module with the same wavelength. The optical path of the optical switch has flexible reconfigurability. To facilitate flexible mapping configuration of multiple groups of bilateral ports, Figure 3 In L1 , for the same pair of bilateral ports of the optical switch, such as P R1 , they can be respectively connected to different upstream ports of the same electrical switch or the upstream ports of different electrical switches in the same switching unit within the group.

[0049] In some embodiments of the present invention, the total number of pairs of bilateral ports of the optical switch in the inter-group switching system can be configured to be equal to N*(N - 1) / 2, where N is the number of switching units within the group in the intra-group switching system. In this case, an optical path connection can be established between any two switching units within the group. At this time, the number of optical switches in the inter-group switching system satisfies the following formula:

[0050]

[0051] where O is the number of optical switches, and n o is the number of pairs of bilateral ports in the o-th optical switch.

[0052] Figure 4 This is an example of a novel optoelectronic hybrid interconnection system provided by the present invention. As Figure 4 shown, the number of switching units within the group N = 5, the number of single-sided ports of each optical switch is equal, n = n o = 5, o = 1, 2,..., O, and the number of optical switches satisfies:

[0053]

[0054] When the number of single-sided ports of each optical switch is equal, the number of optical switches can be calculated as:

[0055] In this case, an optical path connection is established between any two switching units within the group through an optical switch, and an inter-group optical path switching channel can be provided for any two switching units within the group. Specifically, Figure 4In two optical switches, the first pair of bilateral ports of the two optical switches are respectively connected to 4 different upstream ports of the electrical switch in the in-group switching unit G1. Similarly, the second, third, fourth, and fifth pairs of bilateral ports of the two optical switches are respectively connected to 4 different upstream ports of the electrical switches in the in-group switching units G2, G3, G4, and G5. Through the one-to-one mapping relationship (mapping configuration) established between the different-side ports of different pairs of bilateral ports of the two optical switches, for example, the 5 groups of bilateral ports formed by one optical switch are P L1 -P R2 ,P L2 -P R3 ,P L3 -P R4 ,P L4 -P R5 ,P L5 -P R1 ,and the 5 groups of bilateral ports formed by the other optical switch are P L1 -P R3 ,P L2 -P R4 ,P L3 -P R5 ,P L4 -P R1 ,P L5 -P R2 ,an optical path connection is established between any two in-group switching units through the optical switch. For example, an optical path connection is established between the in-group switching units G1 and G2 through a group of ports P Figure 4 of the left optical switch in L1 -P R2 ,and an optical path connection is established between the in-group switching units G1 and G3 through a group of ports P Figure 4 of the right optical switch in L1 -P R3 ,It can be clearly seen from the figure that optical path connections are also established between other pairs of in-group switching units through the optical switch, which will not be listed one by one here.

[0056] In some other embodiments of the present invention, the total number of pairs of bilateral ports of the optical switch in the inter-group switching system can be configured to be greater than N*(N - 1) / 2. In this case, one or more optical path connections can be established between any two in-group switching units, and a larger inter-group optical circuit switching bandwidth can be provided. At this time, the number of optical switches in the inter-group switching system satisfies the following formula:

[0057]

[0058] When the number of single-side ports of each optical switch is equal (n o = n), the number of optical switches can be calculated as:

[0059] Figure 5 An example of a novel optoelectronic hybrid interconnection system provided for this application. As Figure 5 shown, the number of intra-group switching units N = 3, the number of single-sided ports of each optical switch is equal, n = n o = 3, the number of optical switches O = 2, that is, it satisfies That is

[0060] Figure 5 Among the two optical switches of, the first pair of bilateral ports of the two optical switches are respectively connected to 4 different uplink ports of the electrical switch in the intra-group switching unit G1. Similarly, the second and third pairs of bilateral ports of the two optical switches are respectively connected to 4 different uplink ports of the electrical switches in the intra-group switching units G2 and G3. Through the one-to-one mapping relationship (mapping configuration) established between the different-side ports of different pairs of bilateral ports of the two optical switches, the three groups of bilateral ports formed by each optical switch are respectively P L1 -P R2 ,P L2 -P R3 ,P L3 -P R1 And P L1 -P R3 ,P L2 -P R1 ,P L3 -P R2 ,An optical path connection is established between any two intra-group switching units for each group of bilateral ports mapped by each optical switch. Then, two optical path connections are established between any two intra-group switching units for the groups of bilateral ports mapped by the two optical switches, thereby providing more data transmission channel options for the computing nodes across groups. These transmission channels can be used simultaneously or partially, and these transmission channels can be used to construct an adapted optical path for different services through optical path switching inside the switch. In an alternative embodiment of the present invention, different numbers of optical path connections can also be established between different intra-group switching units based on service requirements, so as to establish more or fewer optical path connections between specific intra-group switching units. For example, only one optical path connection can be established between the intra-group switching units G1 and G2, and three optical path connections can be established between the intra-group switching units G2 and G3. This is only an example here, and the present invention is not limited to this.

[0061] In some other embodiments of the present invention, the total number of pairs of bilateral ports of the optical switch in the inter-group switching system can be configured to be less than N*(N - 1) / 2. In this case, optical path connections cannot be established simultaneously between any two intra-group switching units, but optical path connections can be established between selected intra-group switching units, and the flexible and switchable ability of the optical switch can be utilized to construct an adapted optical path for services. At this time, the number of optical switches in the inter-group switching system satisfies the following formula: o = 1, 2, …, O.

[0062] Figure 6 This is an example of a novel optoelectronic hybrid interconnection system provided by this application. As Figure 6 shown, the number of intra-group switching units N = 5, the number of unilateral ports of each optical switch is equal, n = n o = 5, and the number of optical switches satisfies:

[0063]

[0064] At this time, the number of optical switches can be calculated as: That is, O = 1;

[0065] where, n o = n = 5, o = 1.

[0066] Figure 6 The 1st to 5th pairs of bilateral ports of the optical switch of are respectively connected to 2 different upstream ports of the electrical switches in the intra-group switching units G1 to G5. Through the one-to-one mapping relationship established between the different-side ports of different pairs of bilateral ports of the optical switch configured by the left mapping in Figure 6 , the 5 groups of bilateral ports P Figure 6 formed by this optical switch L1 -P R2 , P L2 -P R3 , P L3 -P R1 , P L4 -P R5 and P L5 -P R1 respectively establish optical path connections between two intra-group switching units G1 and G2, G2 and G3, G3 and G4, G4 and G5, and G5 and G1, that is, optical path connections are established between specific intra-group switching units, but optical path connections are not established between any two intra-group switching units. In this case, the flexible switching ability of optical switching technology (OCS) can be utilized to construct an adapted optical path for services between other intra-group switching units through internal optical path switching of the switch, so as to realize service transmission between other intra-group switching units. For example, by changing the left mapping configuration in Figure 6 to Figure 6 Figure 6 the left mapping configuration in Figure 6 is changed to Figure 6The mapping configuration shown on the right side of the figure is used to switch the optical path. The modified mapping configuration forms 5 groups of double-sided ports P L1 -P R3 , P L2 -P R4 , P L3 -P R1 , P L4 -P R5 and P L5 -P R2 Optical path connections can be established between two selected intra-group switching units G1 and G3, between G2 and G4, between G3 and G1, between G4 and G5, and between G5 and G2, respectively, to meet the service transmission requirements between these intra-group switching units.

[0067] When there are multiple optical switches in the inter-group switching system, the numbers of switching ports of different optical switches may be the same or different. Figures 4 - 6 This is an example where different optical switches in the system have the same number of switching ports. Figure 7 This is an example of different optical switches in the system having different numbers of switching ports.

[0068] like Figure 7 As shown, the number of switching units in the group is N = 5, the system contains 3 optical switches (O = 3), the number of ports on one side of the optical switch is n1 = n2 = 3, n3 = 4, and the number of optical switches satisfies Attached Figure 7 The case where the number of optical switch ports is different is given. Figure 7 It can be seen that there is an optical path connection between any two switching units in different groups, which can provide an inter-group optical path switching channel for the corresponding switching units in the group.

[0069] In the embodiment of the present invention, whether according to In the optical-electric hybrid interconnection system, optical switches are arranged according to An optical switch is arranged in the optoelectronic hybrid interconnection system, and the mapping configuration of the opposite-side port of the optical switch can be flexibly switched according to the transmission requirements of different switching units in the group.

[0070] For the business characteristics of small north-south traffic and large east-west traffic in the intelligent computing center, as an aggregation layer switch, the optical switch can provide a more suitable transmission channel for intelligent computing services due to its characteristics of large bandwidth and low latency. Since intelligent computing services exhibit certain periodicity, burstiness, and discontinuity, when the number of optical path connections actually provided by the optical switch is less than the basic connection requirements (such as there is one optical path connection in any group of switching units), the mapping configuration of the connection optical path through the opposite-side ports of the optical switch can be reconfigured in time according to the service requirements, so as to find a suitable optical path connection for the service. In addition, when the possible number of optical path connections that the optical switch can provide is greater than the basic connection requirements between groups of switching units, the reconfigurability of the optical switch can be used to provide more connections than required between two groups of switching units to provide a larger transmission bandwidth.

[0071] The above-mentioned optoelectronic hybrid interconnection system for large-scale intelligent computing centers of the present invention is designed starting from the data interaction requirements of large model training services. A group of switching units includes the interconnection within the computing nodes and the interconnection between the servers within the group. The interconnection of computing nodes within the server is used to improve the data processing and communication efficiency within a single server; the interconnection between servers can ensure efficient data exchange between servers within the same group. The inter-group switching system is responsible for the data flow and coordination between different groups, and overall forms an efficient, scalable, and stable network architecture, thus supporting large-scale cluster computing and complex data processing requirements.

[0072] In addition, starting from the requirements of flexibility and construction cost, in terms of flexibility, the electrical switch of the intra-group switching system provides a flexible switching method for the computing nodes within the same group. At the same time, the inter-group optical switch provides a large-capacity and low-latency data transmission channel for the computing nodes across groups, and optical paths can be flexibly established through a service-oriented dynamic resource scheduling method to provide adaptable connection and switching capabilities for the service. In terms of construction cost, the invested optical switch has the ability to be smoothly upgraded. When the number of optical path connections provided by the actually invested optical switch is greater than the basic optical path connection number requirements, the present invention can provide a larger-capacity transmission channel between the intra-group switching units; when the number of optical path connections provided by the actually invested optical switch is less than the basic optical path connection number requirements, the flexible and switchable ability of the optical switch can be used to construct an adaptable optical path for the service. Compared with the network interconnection systems of other data centers, the present invention has a lower construction cost and greater advantages.

[0073] Those of ordinary skill in the art should understand that the various exemplary components, systems, and methods described in connection with the embodiments disclosed herein can be implemented in hardware, software, or a combination of both. Specifically, whether to implement it in hardware or software depends on the specific application and design constraints of the technical solution. Professional technicians can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of the present invention. When implemented in hardware, it can be, for example, an electronic circuit, an application-specific integrated circuit (ASIC), appropriate firmware, a plug-in, a functional card, and so on. When implemented in software, the elements of the present invention are programs or code segments used to perform the required tasks. The program or code segment can be stored in a machine-readable medium or transmitted through a data signal carried in a carrier wave on a transmission medium or a communication link.

[0074] It should be clear that the present invention is not limited to the specific configurations and processes described above and illustrated in the figures. For the sake of brevity, detailed descriptions of known methods are omitted here. In the above embodiments, several specific steps are described and illustrated as examples. However, the method process of the present invention is not limited to the specific steps described and illustrated. Those skilled in the art can make various changes, modifications, and additions, or change the order between steps after understanding the spirit of the present invention.

[0075] In the present invention, the features described and / or exemplified for one embodiment can be used in the same or a similar manner in one or more other embodiments, and / or combined with the features of other embodiments or replace the features of other embodiments.

[0076] The above are only the preferred embodiments of the present invention and are not used to limit the present invention. For those skilled in the art, various changes and modifications can be made to the embodiments of the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. An optoelectronic hybrid interconnection system for large-scale intelligent computing centers, characterized in that: The system comprises: an intra-group switching system and an inter-group switching system, wherein the intra-group switching system comprises a plurality of intra-group switching units, and the inter-group switching system comprises at least one optical switch; Each intra-group switching unit includes a first number of electrical switches and at least one server, each server having at least a first number of computing nodes corresponding to the first number of electrical switches, wherein the first number of computing nodes are interconnected, and each electrical switch is connected to a corresponding computing node in the first number of computing nodes through a downlink port; Each optical switch in the inter-group switching system includes multiple pairs of bilateral ports, and the multiple pairs of bilateral ports are used to provide multiple optical path connections between different pairs of opposite-side ports through mapping configurations between different pairs of opposite-side ports. The two opposite-side ports of each optical path connection of the same switch can be respectively connected to the uplink ports of the electrical switches of different intra-group switching units based on the connection relationship between the optical switch port and the electrical switch port. The mapping configuration enables inter-group optical path switching channels to be formed between different intra-group switching units.

2. The system according to claim 1, characterized in that The first number of computing nodes of the servers of the switching units in the group are interconnected via high-speed interconnection technology.

3. The system according to claim 1, characterized in that The intra-group switching system and the inter-group switching system are interconnected by using a single-fiber bidirectional optical module for data transmission.

4. The system according to claim 1, characterized in that The number of downlink ports of each electrical switch in the intra-group switching unit is greater than or equal to the number of the servers in the corresponding intra-group switching unit.

5. The system according to claim 1, characterized in that The electrical switches of the switching units in the group are corresponding to the computing nodes inside the server through a mapping relationship between the electrical switch numbers and the computing node numbers.

6. The system according to claim 1, characterized in that The connection relationship between the optical switch port and the electrical switch port includes: the same pair of bilateral ports in the optical switch are respectively connected to different uplink ports of the electrical switch in the same in-group switching unit; The mapping configuration includes establishing a one-to-one mapping relationship between opposite-side ports of different pairs of the dual-side ports.

7. The system according to claim 1, characterized in that The total number of pairs of bilateral ports of the optical switch of the inter-group switching system is greater than or equal to a second number, and at least one optical path connection is established between any two intra-group switching units, wherein the second number is N*(N-1) / 2, wherein N is the number of intra-group switching units in the intra-group switching system; The number of optical switches in the inter-group switching system satisfies the following formula: Wherein, O is the number of the optical switches, n o is the number of pairs of ports on both sides of the oth optical switch.

8. The system according to claim 1, characterized in that The total number of pairs of ports on both sides of the optical switch of the inter-group switching system is less than a second number, and an optical path connection is established between the selected intra-group switching units, wherein the second number is N*(N-1) / 2, wherein N is the number of intra-group switching units in the intra-group switching system; The number of optical switches in the inter-group switching system satisfies the following formula: Wherein, O is the number of the optical switches, n o is the number of pairs of ports on both sides of the oth optical switch.

9. The system according to claim 7 or 8, characterized in that: The mapping configuration of the opposite-side port of the optical switch of the inter-group switching system is switched according to the service transmission requirements of different intra-group switching units, thereby realizing the switching of the connection optical path.

10. The system according to claim 1, characterized in that The computing node is used to send and receive data information that needs to be interacted with; The electrical switch of the switching unit in the same group is used to receive data information sent from the corresponding computing node and forward it to the connected optical switch of the inter-group switching system, or to receive data information sent from the connected optical switch of the inter-group switching system and forward it to the corresponding computing node; The optical switch is used to receive data information from the electrical switch and forward the received data information between switching units in different groups.

Citation Information

Patent Citations

  • High performance computing interconnection network system and communication method

    CN106789750A

  • Machine learning-oriented distributed computing interconnection network system and communication method

    CN111193971A

  • Distributed optical switching interconnection integrated chip and interconnection system

    CN115663589A

  • Optical network system

    US20210176543A1