Interleaved wiring physical topology based on optical switching dual-chip and its calculation method
By using the interleaved wiring physical topology of the optical switching dual-chip and topology engineering algorithms, the problems of traditional topologies being unable to adapt to arbitrary logical topologies and the high cost of Circulator solutions are solved, achieving efficient dynamic cluster adaptation and improved signal stability.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- YU GUANG YUEDONG HONG KONG CO LTD
- Filing Date
- 2026-03-30
- Publication Date
- 2026-07-31
AI Technical Summary
In existing technologies, OCS clusters based on traditional Uniform Wiring physical topologies cannot adapt to arbitrary logical topologies. The topology engineering problem is NP-Complete, and Google's Circulator solution increases hardware costs and introduces signal crosstalk.
It adopts an interleaved wiring physical topology based on optical switching dual chips, and solves the topology engineering problem in polynomial time through mirror-symmetric physical topology design and topology engineering algorithm, achieving compatibility with arbitrary logical topologies.
Solving topology engineering problems in polynomial time enhances the dynamic adaptability of clusters, saves hardware costs, eliminates signal crosstalk, and improves the quality and stability of optical signal transmission.
Smart Images
Figure CN122496736A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of cluster networking design technology, specifically relating to an interleaved wiring physical topology based on a dual-chip optical switch and its calculation method. Background Technology
[0002] In recent years, OCS (Optical Switching System) technology has been gradually applied to GPU cluster design due to its advantages of low cost per port, low generational evolution cost, and low power consumption. OCS is a passive device that can route optical signals without converting them into electrical signals. After physical connection between an OCS and different electrical switches, a transparent inter-switch link can be built based on a set of OCS configurations. A K×K OCS contains 2×K ports, which are divided into N-zones and S-zones; ports in different zones can communicate, but ports within the same zone cannot communicate.
[0003] Unlike electrical switches (EPS), OCS provides programmable bandwidth for electrical switching devices. Once the cluster is deployed, the connections constituting its physical topology remain largely unchanged. Therefore, the connection design between the OCS and directly connected devices is based on the OCS cluster. In practical applications, the OCS may be directly connected to a server or to an electrical switch. For convenience, the devices between the OCS and the OCS are described using an electrical switching group (Egroup). Within an Egroup, communication is direct via electrical signals, while between Egroups, optical signals are transmitted through the OCS to communicate with each other.
[0004] In OCS architecture GPU clusters, given the dynamic nature of OCS, the topology design can be divided into two core concepts: 1) Physical Topology, which describes the physical interconnection between electrical switching groups and the OCS; and 2) Logical Topology, which specifies the required number of links between electrical switching groups, typically determined by traffic characteristics. The physical topology is designed during cluster setup and remains largely fixed once determined. The logical topology, however, can be dynamically changed at a high frequency by configuring the OCS, thus providing links adapted to application traffic requirements based on the physical topology. Solving for the logical topology by configuring the OCS based on the physical topology is called Topology Engineering (ToE). Since the physical topology remains largely constant throughout the cluster's lifecycle, designing a suitable physical topology is crucial. A well-designed physical topology can support more logical topologies, making topology engineering calculations faster and enabling larger-scale operations.
[0005] In some research, physical topology design often employs a simple uniform wiring approach, where the uplink ports of electrical switches or Pods are uniformly connected to different OCSs. This approach follows the wiring logic of traditional electrical switched networks. However, these works neglect symmetry constraints, and it has been proven that physical topologies designed based on this method cannot support arbitrary logical topologies. Furthermore, the topology engineering problem is NP-complete under such physical topologies, making it impossible to solve the OCS configuration through topology engineering to meet the link requirements defined by the logical topology. In contrast, Google uses Circulator technology for physical topology design in clusters such as JupiterEvolving and TPUv4. While this approach can support arbitrary logical topologies, it's important to note that Circulator introduces an additional 1dB-1.4dB of insertion loss, increasing networking costs, and may also cause forward and reverse signal crosstalk. Summary of the Invention
[0006] This invention addresses the shortcomings of existing technologies. Specifically, networking schemes based on traditional Uniform Wiring physical topologies cannot adapt to any possible logical topologies. Furthermore, on such physical topologies, the topology engineering problem is an NP-complete problem, which limits the dynamic adaptability of the cluster. Google's Circulator-based physical networking scheme increases hardware costs, introduces additional optical signal insertion loss, and is prone to forward and reverse signal crosstalk. This invention provides an interleaved wiring physical topology structure and its calculation method based on a dual-chip optical switch. By integrating two independent optical switching chips in the OCS, and through a mirror-symmetric physical topology design, combined with the proposed topology engineering algorithm theory, the topology engineering problem can be solved in polynomial time, achieving compatibility with any logical topology.
[0007] To achieve the above objectives, the present invention adopts the following technical solution:
[0008] In a first aspect, the present invention provides an interleaved wiring physical topology based on a dual-chip optical switching system, comprising: T individual electrical switching groups and One OCS, K An even number of electrical switching groups are present; each group contains... K Each OCS integrates two independent optical switching chips; the physical topology corresponds to two mirrored sub-physical networks. The first sub-physical network includes the first optical switching chip of each OCS and the transmit port of the even-numbered port and the receive port of the odd-numbered port of each electrical switching group. The second sub-physical network includes the second optical switching chip of each OCS and the transmit port of the odd-numbered port and the receive port of the even-numbered port of each electrical switching group.
[0009] Optionally, the first t The second of the electrical switching groups k -1 port's transmit port and the 2nd port k The receiving port of port is connected to the receiving port of port 1 respectively. k The first optical switching chip of the OCS t The Nth port and the first t The first S port; the first t The second of the electrical switching groups k The sending port of the first port and the second port k The receive ports of port -1 are respectively connected to the first port. k The second optical switching chip of the OCS t The Nth port and the first t One S port; , .
[0010] Secondly, the present invention provides a topology engineering calculation method based on the interlaced wiring physical topology structure of the first aspect, comprising the following steps: According to the physical topology, the power switching group and the OCS are evenly connected to build a cluster. During cluster operation, the logical topology is determined by the traffic of services. The logical topology represents the link requirements between electrical switching groups, and the cluster provides programmable bandwidth for services based on changes in the logical topology. The logical topology is split into two sub-logical topologies that are transposes of each other. and Taking the first sub-logical topology and the first sub-physical network as input, construct the first sub-topology engineering problem and solve it to obtain the corresponding OCS configuration. Taking the second sub-logical topology and the second sub-physical network as input, construct the second sub-topology engineering problem and solve it to obtain the corresponding OCS configuration. Combine the two OCS configurations into the solution of the topology engineering problem.
[0011] Optionally, the logical topology is split into two sub-logical topologies that are transposes of each other, satisfying the following equation:
[0012] In the formula, The first logical topology defined The and the first Link requirements between individual electrical switching groups The first defined for sub-logic topology The and the first Link requirements between individual electrical switching groups.
[0013] Optionally, in the first sub-topology engineering problem, a set of electrical switching group numbers is defined. It covers the serial number identifiers of all electrical switching groups; it defines the OCS serial number set. This includes the serial number identifiers for all OCSs; it defines the OCS configuration matrix. This is used to describe the interconnection topology between electrical switching groups, where the following conditions are met: , ;when When, it indicates the first The first optical switching chip of the OCS device establishes a unidirectional link, and the link is then... i The second of the electrical switching groups k -1 port's transmit port is through the first k The first optical switching chip of the OCS is connected to the first j The second of the electrical switching groups k The receiving port of each port; The first sub-topology engineering problem satisfies the following constraints: Inter-switch flow capacity constraints:
[0014] In the formula, They represent the first , No. The set of ports contained in an electrical switching group This refers to the electrical connection between ports; OCS Configuration and Node Link Consistency Constraints:
[0015] In the formula, Indicates the first i The first electrical switching group m The OCS sequence number connected to the transmit port of each port; OCS port sender exclusive constraint:
[0016] OCS port receiver exclusive constraint:
[0017] The OCS configuration of the first subproblem is solved in polynomial time using a cost flow algorithm. .
[0018] Optionally, based on the characteristic that the first sub-physical network and the second sub-physical network are mirror images of each other, the OCS configuration of the first sub-problem is used. Solve the OCS configuration of the second subproblem ,when When, it indicates the first The second optical switching chip of the OCS device establishes a unidirectional link, which starts from the first... j Group 2 of the electrical switching unit k The Tx port of port k connects to the second optical switch chip of the k-th OCS. i Group 2 of the electrical switching unit k-1 Links to Rx ports.
[0019] The beneficial effects of this invention are: (1) In view of the shortcomings of traditional Uniform Wiring physical topology being unable to adapt to arbitrary logical topology, topology engineering solution being NP-COmplete, and limiting the dynamic adaptation capability of the cluster, this invention solves the topology engineering problem in polynomial time by using dual-chip OCS and interleaved wiring design, combined with integer matrix partitioning theory, to meet the link requirements of arbitrary logical topology definition, and greatly improve the dynamic adaptation capability of the cluster to different service traffic requirements. (2) Compared with Google’s Circulator-based networking scheme, the present invention does not rely on Circulator devices, which can save costs and completely eliminate the 1dB-1.4dB extra insertion loss caused by Circulator, while avoiding crosstalk between forward and reverse signals, effectively improving the quality and stability of optical signal transmission. Attached Figure Description
[0020] Figure 1 This is a schematic diagram of the physical topology of interleaved wiring based on a dual-chip optical switching system.
[0021] Figure 2 This is a flowchart of the topology engineering calculation method. Detailed Implementation
[0022] The invention will now be described in further detail with reference to the accompanying drawings.
[0023] Example 1 This embodiment proposes a physical topology for interleaved cabling based on a dual-chip optical switching system, focusing on the physical topology design based on dual-chip optical switching (OCS). For example... Figure 1 As shown, the OCS designed in this embodiment integrates two independent optical switching chips. The S-ports and N-ports of different optical switching chips cannot communicate across chips. Furthermore, the connection method between the OCS and the electrical switching equipment in this embodiment differs from existing technologies. The wiring between the two optical switching chips and the electrical switching equipment exhibits a "mirror-symmetric" characteristic. Specifically, the ports of the electrical switching group are grouped in pairs and wired as follows. If the cluster contains... T Each electrical switching group contains [number] electrical switching units. K Given ports (K is an even number), then... , : No. t The second of the electrical switching groups k- One port for sending (Tx, Figure 1 (Middle blue circle) and the 2nd k The receive port of each port (Rx, Figure 1 (circled in red), respectively connected to the first k The first optical switching chip of the OCS t The Nth port and the first t One S port.
[0024] No. t The second of the electrical switching groups k The sending port (Tx) of the first port and the second k- The receive port (Rx) of port 1 is connected to the first port respectively. k The second optical switching chip of the OCS t The Nth port and the first t One S port.
[0025] The physical topology design described above is the most important design feature that distinguishes this solution from other solutions. Using this physical topology design, the network can be logically decomposed into two mirror-image sub-physical networks: the first sub-physical network includes the first optical switching chip and the even-numbered transmit ports (Tx) and odd-numbered receive ports (Rx) of each electrical switching group; the second sub-physical network includes the second optical switching chip and the odd-numbered transmit ports (Tx) and even-numbered receive ports (Rx) of each electrical switching group. If a connection is built from the first optical switching chip in the first sub-physical network... i Group 2 of the electrical switching unit k- One port Tx port to the first j Group 2 of the electrical switching unit k A link on the Rx port of the first port can be synchronously constructed through the second optical switching chip in the second sub-physical network, starting from the first port. j Group 2 of the electrical switching unit k The Tx port of the first port to the first i Group 2 of the electrical switching unit k- A link with one Rx port.
[0026] Example 2 Based on the physical topology of Embodiment 1, this embodiment proposes a topology engineering calculation method, such as... Figure 2 As shown, the workflow can be divided into the following parts: The first stage is deployment. During deployment, the cluster equipment provider will define the network scale and requirements based on business needs, including the number of GPUs in the cluster, single-port bandwidth, switching chip capacity, OCS port specifications, etc. Then, according to the determined physical topology design, the electrical switching group and OCS are evenly connected to build the cluster. Once the deployment phase is complete, the physical topology design of the cluster remains largely stable unless there are needs for equipment upgrades or cluster expansion.
[0027] The second stage is the logical topology generation phase. The logical topology is determined during cluster operation based on service traffic requirements. The frequency of logical topology generation is determined by the equipment builder. Depending on the service requirements, in a data center (DCN) scenario, the logical topology may be updated daily, while in a GPU cluster scenario, it may be updated on a task-by-task basis. The physical topology design of this invention is compatible with different scenarios. The logical topology describes the link requirements between electrical switching groups, and the OCS cluster provides programmable bandwidth for services based on changes in the logical topology.
[0028] Then comes the topology engineering calculation stage. In this embodiment, thanks to... Figure 1 The design of physical topology can be achieved in polynomial time by solving the topology engineering problem (i.e., Figure 1 The connection relationship between the OCS N port and S port (in the OCS) satisfies the link requirements of any logical topology definition. Specifically, according to Figure 1 The physical topology design can be divided into two mirror-symmetric sub-physical topologies. The first sub-physical network includes the first optical switching chip in each OCS and the even-numbered transmit ports (Tx) and odd-numbered receive ports (Rx) of each electrical switching group. The second sub-physical network includes the second optical switching chip and the odd-numbered transmit ports (Tx) and even-numbered receive ports (Rx) of each electrical switching group. The mirror-symmetric nature of these two sub-physical topologies means that if a path from the first optical switching chip to the second optical switching chip is constructed in the first sub-physical network... i Group 2 of the electrical switching unit k- One port Tx port to the first j Group 2 of the electrical switching unit k A link on the Rx port of the first port can be synchronously constructed through the second optical switching chip in the second sub-physical network, starting from the first port. j Group 2 of the electrical switching unit k The Tx port of the first port to the first i Group 2 of the electrical switching unit k-A link with a single port (Rx port). Two opposing links in these two sub-physical topologies can be combined to define a bidirectional link. This means that if the logical topology is further decomposed, the topology engineering problem can be divided into two subproblems. Solving a single subproblem does not require concern with symmetry constraints. According to the theory of integer matrix partitioning, it can be proven that, ignoring symmetry constraints, this subproblem is solvable in polynomial time. Specifically, the logical topology C can be further decomposed into two transpose sub-logical topologies. and Where:
[0029] In the formula, The first logical topology defined The and the first Link requirements between individual electrical switching groups The first defined for sub-logic topology The and the first Link requirements between individual electrical switching groups.
[0030] The Eulerian circuit algorithm can be used to prove that this expression is always true. Then, the sub-logic topology is taken. Taking the first sub-physical network as input, construct a new sub-topology engineering problem using the following model.
[0031] For convenience, a set of electrical switching group numbers is defined: This includes serial number identifiers for all electrical switching groups; it defines the OCS serial number set: This includes the serial number identifiers for all OCSs; it defines the OCS configuration matrix. It is used to accurately describe the interconnection topology between electrical switching groups, wherein the following conditions are met: , ;in T Indicates the total number of electrical switching units, even numbers The total number of ports in a single electrical switching group. When, it indicates the first The first optical switching chip of the OCS device establishes a unidirectional link, and the link is then... i The second of the electrical switching groups k -1 port's Tx through the first k The first optical switching chip of the OCS is connected to the first j Group 2 of the electrical switching unit k The Rx of each port. This sub-topology engineering problem needs to satisfy the following constraints: 1. Inter-bank flow capacity constraints This constraint guarantees that the sum of link traffic on all ports within a single electrical switching group equals the link requirements between electrical switching groups defined in the corresponding logical topology. The mathematical expression is:
[0032] In the formula, They represent the first , No. The set of ports contained in an electrical switching group This refers to the electrical connection between ports. The first defined for sub-logic topology The and the first Link requirements between individual electrical switching groups.
[0033] 2. OCS Configuration and Node Link Consistency Constraints This constraint ensures a complete correspondence between OCS port-level configuration variables and underlying electrical switching group port link variables, achieving precise mapping from port configuration to node links. The mathematical expression is:
[0034] In the formula, Indicates the first i The first electrical switching group m The OCS sequence number of the sending port of each port is used to ensure that the OCS configuration associated with the same port is consistent with the node link status.
[0035] 3. OCS port transmitter exclusive constraint This constraint limits the transmitter of a single OCS port to connecting to a maximum of one target electrical switching group at any given time, thus avoiding port transmitter conflicts. The mathematical expression is:
[0036] 4. OCS port receiver exclusive constraint This constraint limits the receiver of a single OCS port to a maximum of one source-powered switching group at any given time, avoiding port receiver collisions. The mathematical expression is:
[0037] It should be noted that the above formula does not include symmetric constraints. Based on the integer matrix partitioning theorem, the OCS configuration of the first subproblem can be solved in polynomial time using a cost-flow algorithm. .
[0038] Based on the characteristic that subnetworks are "mirrored from each other", if we take the second physical topology and The OCS configuration of the second subproblem can be solved. and take At that time, the OCS configuration is a feasible solution to the second subproblem. Where, when When, it indicates the first The second optical switching chip of the OCS device establishes a unidirectional link, which starts from the first... j Group 2 of the electrical switching unit k The Tx port of the first port is accessed through the first... k The second optical switching chip of the OCS to the first i Group 2 of the electrical switching unit k- A link with a single Rx port. Based on the characteristic that subnetworks are "mirrored from each other," the OCS configurations of the two subproblems are merged into the solution to the original problem. The solution to the original problem then naturally satisfies the symmetry constraint and the integer programming model for topology engineering defined in the background section.
[0039] Therefore, the core of this solution lies in the physical topology design satisfying the characteristic of "mirror symmetry". Utilizing this characteristic, for any logical topology, the physical topology based on this invention can meet the logical topology requirements in polynomial time through topology engineering.
[0040] The above are merely preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principle of the present invention should be considered within the scope of protection of the present invention.
Claims
1. A physical topology for interleaved wiring based on a dual-chip optical switching system, characterized in that, include: T an even number of electrical switching groups and an even number of OCSs, K each electrical switching group containing K an even number of ports, each OCS integrating two independent optical switching chips; the corresponding physical topology network comprises two sub-physical networks which are mirror images of each other, the first sub-physical network comprising the first optical switching chip of each OCS and the transmitting ports of the even-numbered ports and the receiving ports of the odd-numbered ports of each electrical switching group, the second sub-physical network comprising the second optical switching chip of each OCS and the transmitting ports of the odd-numbered ports and the receiving ports of the even-numbered ports of each electrical switching group.
2. The interleaved wiring physical topology based on a dual-chip optical switch as described in claim 1, characterized in that: No. t The second of the electrical switching groups k -1 port's transmit port and the 2nd port k The receiving port of port is connected to the receiving port of port 1 respectively. k The first optical switching chip of the OCS t The Nth port and the first t The S-port; the first t The second of the electrical switching groups k The sending port of the second port and the second k The receive ports of port -1 are respectively connected to the first port. k The second optical switching chip of the OCS t The Nth port and the first t One S port; , .
3. A topology engineering calculation method, based on the interlaced wiring physical topology structure described in claim 2, characterized in that, Includes the following steps: According to the physical topology, the power switching group and the OCS are evenly connected to build a cluster. During cluster operation, the logical topology is determined by the traffic of services. The logical topology represents the link requirements between electrical switching groups, and the cluster provides programmable bandwidth for services based on changes in the logical topology. The logical topology is split into two sub-logical topologies that are transposes of each other. and Taking the first sub-logical topology and the first sub-physical network as input, construct the first sub-topology engineering problem and solve it to obtain the corresponding OCS configuration. Taking the second sub-logical topology and the second sub-physical network as input, construct the second sub-topology engineering problem and solve it to obtain the corresponding OCS configuration. Combine the two OCS configurations into the solution of the topology engineering problem.
4. The topology engineering calculation method as described in claim 3, characterized in that: The logical topology is split into two sub-logical topologies that are transposes of each other, satisfying the following equation: In the formula, The first logical topology defined The and the first Link requirements between individual electrical switching groups The first defined for sub-logic topology The and the first Link requirements between individual electrical switching groups.
5. The topology engineering calculation method as described in claim 4, characterized in that: In the first sub-topology engineering problem, a set of electrical switching group numbers is defined. It covers the serial number identifiers of all electrical switching groups; it defines the OCS serial number set. This includes the serial number identifiers for all OCSs; it defines the OCS configuration matrix. This is used to describe the interconnection topology between electrical switching groups, where the following conditions are met: , ;when When, it indicates the first The first optical switching chip of the OCS device establishes a unidirectional link, and the link is then... i The second of the electrical switching groups k -1 port's transmit port is through the first k The first optical switching chip of the OCS is connected to the first j The second of the electrical switching groups k The receiving port of each port; The first sub-topology engineering problem satisfies the following constraints: Inter-switch flow capacity constraints: In the formula, They represent the first , No. The set of ports contained in an electrical switching group This refers to the electrical connection between ports; OCS Configuration and Node Link Consistency Constraints: In the formula, Indicates the first i The first electrical switching group m The OCS sequence number connected to the transmit port of each port; OCS port sender exclusive constraint: OCS port receiver exclusive constraint: The OCS configuration of the first subproblem is solved in polynomial time using a cost flow algorithm. .
6. The topology engineering calculation method as described in claim 5, characterized in that: Based on the characteristic that the first and second sub-physical networks are mirror images of each other, and based on the OCS configuration of the first sub-problem. Solve the OCS configuration of the second subproblem ,when When, it indicates the first The second optical switching chip of the OCS device establishes a unidirectional link, which starts from the first... j Group 2 of the electrical switching unit k The Tx port of port k connects to the second optical switch chip of the k-th OCS. i Group 2 of the electrical switching unit k-1 Links to Rx ports.