Scheduling control, interconnection method and system for processor unit cluster-oriented optical switching network

Abstract

This invention provides a scheduling control, interconnection method, and system for optical switching networks oriented towards processor unit clusters. The scheduling control method is executed by a network scheduling controller. The optical switching network includes multiple processor units, corresponding optoelectronic transceiver module arrays for each processor unit, and an AWGR cluster. The network scheduling controller stores the inherent wavelength-port routing mapping relationship of the AWGR and real-time load status information. The method includes: receiving a communication request from a source xPU; determining all candidate paths connecting the source xPU and the target xPU based on the request; selecting a scheduling path from the candidate paths based on the real-time load status information and a target optimization algorithm; generating and issuing a routing configuration command containing operating wavelength and configuration information based on the selected scheduling path, thereby realizing communication scheduling from the source xPU to the target xPU. This invention enables fine-grained scheduling of optical switching network resources, solving technical problems such as rigid resource allocation and low bandwidth utilization in existing optical switching schemes.

Description

Technical Field

[0001] This invention relates to the field of optical switching network technology, and in particular to an optical switching network scheduling control, interconnection method and system for processor unit clusters. Background Technology

[0002] With the rapid development of artificial intelligence and high-performance computing, deploying and coordinating large-scale processing units (x-Processing Units, xPUs) has become a core technology for building high-performance computing systems. These xPUs include various computing cores such as Graphics Processing Units (GPUs), Central Processing Units (CPUs), and Neural Network Processing Units (NPUs). The performance of the interconnect network formed by these processor units, especially the switching bandwidth and communication latency provided by the interconnect network, is becoming a key bottleneck restricting the overall computing power of high-performance computing systems. To address this challenge, the industry has developed various xPU interconnect technologies based on electrical signals. One type is interconnection based on general-purpose bus protocols, typically represented by the Peripheral Component Interconnect Express (PCIe) standard and the Compute Express Link (CXL) protocol, which has emerged in recent years. In this architecture, multiple xPUs connect to a centralized switching chip through their respective interfaces, and data flows need to be routed and forwarded through the central switching node. Another type is interconnect technology based on dedicated high-speed buses, represented by NVIDIA's NVLink and AMD's Infinity Fabric. This solution is designed specifically for high-performance computing scenarios. It establishes direct links between xPUs through a high-bandwidth, low-latency point-to-point electrical bus, and further aggregates all links by deploying dedicated active electrical switching chips (such as NVSwitch) to build a fully interconnected switching network. These active switching chips need to perform store-and-forward or pass-through switching on incoming data packets and perform complex routing decisions based on packet header information, thereby providing extremely high data exchange capabilities between multiple xPUs within a node.

[0003] Despite the significant success of the aforementioned electrical interconnect technologies, the physical characteristics of signal transmission in metallic media dictate that they will face a series of inherent physical bottlenecks when facing future systems with higher computing density and energy efficiency: ① Severe power consumption wall problem: To drive high-speed electrical links (such as hundreds of Gbps links) and compensate for signal attenuation, the core active electrical switching chips and transceiver circuits generate huge power consumption and heat, leading to a significant increase in the total cost of ownership; ② Electrical switching bandwidth density is limited by I / O (input / output) density bottleneck: The inherent electromagnetic interference and inter-channel crosstalk of high-frequency electrical signals require sufficient physical spacing between high-speed transmission lines, which fundamentally limits the I / O channel density that can be integrated per unit chip or package area; ③ Signal integrity challenge: The dual constraints of power consumption and density exacerbate the complexity of signal integrity design, making electrical interconnect system design dependent on expensive low-loss substrate materials and complex equalization circuits, such as decision-feedback equalization (DFE) and feed-forward equalization (FFE). These circuits (FFE) themselves also contribute significant power consumption and latency, greatly increasing research and manufacturing costs.

[0004] To fundamentally address the aforementioned issues, the industry has been exploring optical switching using Arrayed Waveguide Grating Routers (AWGRs). For example, Chinese invention patent application CN117151183A, entitled "A Chip Interconnect System for Deep Learning," discloses an interconnect system for CPU and GPU chips. This scheme utilizes the low-loss, high-bandwidth, and electromagnetic interference-resistant characteristics of optical signals to overcome the power consumption and signal integrity limitations of electrical schemes from a physical perspective. Furthermore, it significantly improves switching bandwidth through Wavelength Division Multiplexing (WDM) technology and the deployment of multiple parallel AWGRs. However, this design, which statically maps the physical layer optical path to the upper-layer application logic, is essentially a link-level implementation of circuit switching. Its control plane lacks real-time status feedback to the data plane. Therefore, when network congestion, link quality changes, or sudden changes in service flow patterns occur, the optical switching system cannot adaptively adjust, leading to low resource utilization and unexpected degradation of communication performance. This limits it to serving only as a coprocessor or accelerator for specific applications, rather than as a programmable, software-defined optical interconnect infrastructure to support future general-purpose heterogeneous computing.

[0005] Therefore, there is an urgent need in this field for a new optical network control and scheduling architecture for xPU interconnects to overcome the limitations of existing optical switching schemes that use predefined communication modes, thereby elevating the optical switching network from a static, pre-configured physical layer to a dynamic system with resource optimization capabilities, and unleashing the full potential of optical switching technology in the field of high-performance computing interconnects. Summary of the Invention

[0006] In view of this, embodiments of the present invention provide an optical switching network scheduling control, interconnection method and system for processor unit clusters. This scheme introduces a centralized network scheduling controller, which can dynamically calculate and establish optical paths that meet the target in the optical switching network composed of optoelectronic transceiver module arrays and parallel AWGR clusters according to real-time communication requests and network load status, thereby achieving traffic adaptation and efficient resource utilization.

[0007] One aspect of the present invention provides a scheduling and control method for an optical switching network oriented towards a processor unit cluster. This method is executed by a network scheduling controller, which is communicatively connected to the optical switching network. The optical switching network consists of multiple components, including multiple processor units, an array of optoelectronic transceiver modules at each processor unit, and an arrayed waveguide grating router (AWGR) cluster. The network scheduling controller stores the inherent wavelength-port routing mapping relationships of each AWGR in the AWGR cluster and the real-time load status information of the optical switching network. The method includes the following steps: Receive Request: Receive a communication request sent by the source processor unit; wherein, the communication request includes the identification information of the source processor unit and the identification information of the target processor unit to be interconnected; Candidate optical paths are determined based on the communication request, the preset component communication connection rules, and the wavelength-port routing mapping relationship. All candidate optical paths connecting the source processor unit and the target processor unit are determined. Each candidate optical path is defined by a source-end optoelectronic transceiver module, a target-end optoelectronic transceiver module, an AWGR, and a preset operating wavelength. Selecting the scheduling optical path: Based on real-time load status information and using a target optimization algorithm, the scheduling optical path that meets the communication request is selected from the candidate optical paths; Generate and issue configuration instructions: Generate routing configuration instructions based on the scheduling optical path. These instructions include the operating wavelength specified by the scheduling optical path and the configuration information of the optoelectronic transceiver module and AWGR in the scheduling optical path. Issue instructions to the optical switching network to establish a communication connection between the source processor unit and the target processor unit, thereby realizing communication scheduling control based on the communication connection.

[0008] In some embodiments of the present invention, the number of optoelectronic transceiver modules in the optoelectronic transceiver module array at each processor unit end is no more than the number of AWGRs in the AWGR cluster, and the number of input ports and output ports corresponding to each AWGR is no less than the number of processor units in the optical switching network. In an optical switching network, each subset consisting of an optoelectronic transceiver module and an AWGR forms an independent switching plane, providing parallel physical links for communication between processor units; Component communication connection rules include: For each candidate optical path connecting the source processor unit and the target processor unit, the identifier of the AWGR input port corresponds to the identifier of the source processor unit, the identifier of the AWGR output port corresponds to the identifier of the target processor unit, and the identifiers of the source-end optoelectronic transceiver module and the target-end optoelectronic transceiver module in the candidate optical path correspond to the identifier of the AWGR in the candidate optical path.

[0009] In some embodiments of the present invention, the load status information is a network transmission performance parameter used to indicate the load status of the optical switching network. The network transmission performance parameter includes one or more of throughput parameters, signal quality parameters, and delay parameters. The throughput parameter includes the used bandwidth or remaining bandwidth of each connectable optical path, the signal quality parameter indicates the transmission quality of each connectable optical path, and the delay parameter indicates the transmission delay of each connectable optical path.

[0010] In some embodiments of the present invention, the method further includes: collecting network load information from processor units and optoelectronic transceiver modules in the optical switching network by periodically querying or receiving event-triggered reports, and dynamically updating the stored real-time load status information.

[0011] In some embodiments of the present invention, the network scheduling controller stores network topology information; The method also includes the following fault handling steps: Analyze the stored real-time load status information. If the real-time load status information does not meet the set conditions, determine the corresponding component failure, or receive explicit component failure alarms uploaded by the processor unit or optoelectronic transceiver module to achieve fault detection. When a fault is detected in a specific component of the optical switching network, the network scheduling controller marks the faulty component as unavailable in the network topology information. This allows the faulty component marked as unavailable to be automatically excluded in the subsequent process of determining candidate optical paths, thus achieving logical isolation.

[0012] In some embodiments of the present invention, the method further includes a service recovery step: when it is determined that an active data transmission service is carried on the faulty component, the network scheduling controller automatically triggers a rerouting calculation for the service; the rerouting calculation is used to re-determine all candidate optical paths that do not contain the faulty component through the candidate optical path determination step, select a new and available scheduling optical path through the scheduling optical path selection step, and issue an updated routing configuration instruction to the optical switching network through the configuration instruction generation and issuance step, so as to switch the data transmission service to the new scheduling optical path without loss or with minimal interruption.

[0013] In some embodiments of the present invention, the communication request further includes a bandwidth requirement; the target optimization algorithm is used to select the optimal scheduled optical path that meets the bandwidth requirement from all candidate optical paths; When the remaining bandwidth of any single candidate optical path does not meet the bandwidth requirement, multiple candidate optical paths are selected together as the final parallel scheduling optical path, thereby satisfying the bandwidth requirement in the communication request through link aggregation.

[0014] Another aspect of the present invention provides a processor unit interconnection method, which is executed in an optical switching network. The optical switching network is communicatively connected to a network scheduling controller. The optical switching network consists of multiple components, including multiple processor units, an array of optoelectronic transceiver modules at each processor unit end, and an arrayed waveguide grating router (AWGR) cluster. Each optoelectronic transceiver module in the array has wavelength division multiplexing capability, including a transmit path containing a modulator array and a receive path containing a photodetector array. The modulator array and the photodetector array respectively contain modulators and photodetectors corresponding to each preset operating wavelength. The transmit path is used to simultaneously generate and modulate optical signals of one or more specific operating wavelengths, and the receive path is used to demultiplex and detect optical signals of multiple operating wavelengths. The method includes the following steps: Request and Configuration: The source processor unit sends a communication request to the network scheduling controller; and the source processor unit receives a routing configuration instruction corresponding to the communication request issued by the network scheduling controller; wherein, the routing configuration instruction includes the scheduling optical path information between the source processor unit and the target processor unit; each scheduling optical path is defined by a source-end optoelectronic transceiver module, a target-end optoelectronic transceiver module, an AWGR and a preset working wavelength, and the scheduling optical path information includes the specified working wavelength and the configuration information of the optoelectronic transceiver module and the AWGR; Electro-optical conversion: The source processor unit selects the source-end opto-transceiver module specified by the routing configuration instruction from the source-end opto-transceiver module array according to the routing configuration instruction; and the selected source-end opto-transceiver module activates the transmission path in the opto-transceiver module according to the specified operating wavelength, thereby modulating the electrical signal generated by the source processor unit onto the optical carrier of the specified operating wavelength to generate the transmitted optical signal. Optical signal routing: The selected source-end opto-transceiver module sends the transmitted optical signal to the AWGR specified by the routing configuration command; the AWGR, based on the wavelength of the transmitted optical signal and its inherent wavelength-port routing mapping relationship, routes the signal of each wavelength in the combined optical signal containing the transmitted optical signal to the target-end opto-transceiver module specified by the routing configuration command. Optical-to-electrical conversion: The target-end optoelectronic transceiver module performs wave demultiplexing on the multiplexed optical signal received by the receiving channel, separating it into optical signals corresponding to the specified operating wavelength, and converting the separated optical signals into transmission electrical signals to send to the target processor unit, thereby completing end-to-end communication.

[0015] In some embodiments of the present invention, the scheduling optical path information included in the routing configuration instruction is determined by the network scheduling controller executing the scheduling control method.

[0016] Another aspect of the present invention provides an optical switching network scheduling and control system for processor unit clusters, including a processor, a memory, and a computer program / instructions stored in the memory. The processor is used to execute the computer program / instructions, and when the computer program / instructions are executed, the system implements the steps of the method described in any of the above embodiments.

[0017] The optical switching network scheduling control, interconnection method, and system proposed in this invention for processor unit clusters, introduces an independent control plane (i.e., the network scheduling controller in this application) to pool and dynamically allocate the physical resources of the optical switching network. The method proposed in this application can flexibly calculate and configure data transmission paths that meet the objectives based on real-time service requirements and network physical status, effectively overcoming the problems of strong coupling between traditional optical switching schemes and applications, and rigid resource allocation.

[0018] Additional advantages, objects, and features of the invention will be set forth in part in the description which follows, and will also become apparent in part to those skilled in the art upon studying the description, or may be learned by practice of the invention. The objects and other advantages of the invention can be realized and obtained by means of the structures specifically pointed out in the description and drawings.

[0019] Those skilled in the art will understand that the objectives and advantages achievable with the present invention are not limited to those specifically described above, and that the above and other objectives achievable with the present invention will become clearer from the following detailed description. Attached Figure Description

[0020] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, are not intended to limit the scope of the invention. In the drawings: Figure 1 This is a schematic diagram of an optical switching system architecture based on an arrayed waveguide grating router in one embodiment of the present invention.

[0021] Figure 2 This is a schematic diagram of the structure of an optoelectronic transceiver module in one embodiment of the present invention.

[0022] Figure 3 This is a schematic diagram illustrating the components and functions of a network scheduling controller in one embodiment of the present invention.

[0023] Figure 4 This is a flowchart illustrating a scheduling and control method for an optical switching network oriented towards a processor unit cluster, according to an embodiment of the present invention.

[0024] Figure 5 This is a flowchart illustrating a processor unit interconnection method according to an embodiment of the present invention.

[0025] Figure 6 This is an example diagram of a multicast communication switching mechanism in one embodiment of the present invention.

[0026] Figure 7 This is an example diagram of optical switching in a multi-task concurrent communication scenario according to one embodiment of the present invention. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the embodiments and accompanying drawings. Here, the illustrative embodiments and descriptions of this invention are used to explain the invention, but are not intended to limit the invention.

[0028] It should also be noted that, in order to avoid obscuring the invention with unnecessary details, only the structures and / or processing steps closely related to the solution according to the invention are shown in the accompanying drawings, while other details that are not closely related to the invention are omitted.

[0029] It should be emphasized that the term "including / comprises" as used herein refers to the presence of a feature, element, step, or component, but does not exclude the presence or addition of one or more other features, elements, steps, or components.

[0030] It should also be noted that, unless otherwise specified, the term "connection" in this article can refer not only to a direct connection, but also to an indirect connection involving an intermediary.

[0031] In the following description, embodiments of the invention will be illustrated with reference to the accompanying drawings. In the drawings, the same reference numerals represent the same or similar parts, or the same or similar steps.

[0032] Existing processor unit electrical interconnection technologies suffer from numerous limitations in terms of power consumption, bandwidth, and signal integrity. To overcome these physical bottlenecks, this application designs an optical switching technology based on AWGR to achieve full interconnection between processor units. Furthermore, while existing technologies employ arrayed waveguide grating routers to interconnect processor units in optical switching networks, these solutions operate according to a fixed communication mode bound to upper-layer applications. For example, they perform optical switching between the source processor unit (hereinafter referred to as the source xPU) and the target processor unit (hereinafter referred to as the target xPU) based on a preset default path, lacking dynamic adaptability. Therefore, this application proposes a system architecture where a data plane (i.e., the optical switching network, Data Plane) and a control plane (also known as a network scheduling controller) are tightly coupled and work collaboratively. In this system architecture, the data plane is a high-bandwidth and scalable physical optical switching network constructed using arrayed waveguide grating routers, while the control plane is used for intelligent scheduling, management, and maintenance of the optical switching network to ensure efficient and reliable utilization of network resources. Furthermore, existing solutions generally neglect high-reliability design at the system level, lack effective fault self-healing mechanisms, and struggle to guarantee stable Quality of Service (QoS) requirements. Therefore, this application employs a network scheduling controller to monitor the load status information of the optical switching network in real time and manages faults through fault perception and rerouting. The optical switching network scheduling control method, interconnection method, and system proposed in this application for processor unit clusters can achieve traffic adaptation, fault self-healing, QoS guarantee, and flexible selection of scheduling paths (i.e., scheduling optical paths) according to application requirements, realizing high-performance interconnection between processor units. Moreover, the method and system proposed in this application can be directly applied to intelligent computing center scenarios running large-scale processor units.

[0033] Logically, this application provides a complete system-level solution to the interconnection problem between processor units by leveraging a system architecture consisting of a constructed optical switching network (data plane) and a network scheduling controller (control plane). The optical switching network uses passive arrayed waveguide grating routers as its switching core and can deploy multiple AWGRs in parallel, aiming to provide a linearly scalable, energy-efficient, fully interconnected physical resource pool for multiple processor units. The network scheduling controller is responsible for executing network control and management functions and works in conjunction with the physical optical switching network to achieve fine-grained scheduling of network resources and ensure reliable system operation. Figure 1 As shown, the optical switching network, as the physical foundation for realizing high-speed data exchange, mainly includes multiple processor units as network endpoints (the types of processor units in the optical switching network can be different; for example, the method and system proposed in this application are applicable to optical switching networks containing heterogeneous processor units such as GPUs, CPUs, and NPUs), an arrayed waveguide grating router cluster with a parallel deployment structure, and an array of optoelectronic transceiver modules at each processor unit end (i.e., the optoelectronic transceiver module array associated with each processor unit), and the optoelectronic transceiver module arrays associated with each processor unit are deployed in parallel. The number of processor units included in the optical switching network can be represented by N (N is a positive integer not less than 2), and the number of optoelectronic transceiver modules in the optoelectronic transceiver module array associated with each processor unit can be represented by M (M is a positive integer). The arrayed waveguide grating router cluster includes at least one arrayed waveguide grating router, and the AWGR contains multiple input ports and multiple output ports. In this invention, the processor unit refers to various computing or processing cores that require high-speed data exchange, such as heterogeneous cores such as GPUs, CPUs, and NPUs. To fully leverage the powerful physical layer communication capabilities of the optical switching network and make it an efficient and reliable switching system, this application also designs a network scheduling controller that can achieve traffic load balancing and high reliability of the optical switching network. The network scheduling controller communicates with the optical switching network in the data plane through a control bus to achieve close collaboration between the two.

[0034] More specifically, each processor unit in the optical switching network has a high-speed I / O interface for exchanging electrical signal data with its corresponding optoelectronic transceiver module array (i.e., the optoelectronic transceiver module array at each processor unit). Simultaneously, the driver or firmware running on the processor unit can interact with the network scheduling controller via the control bus, uploading communication requests and receiving and executing routing configuration commands. The optoelectronic transceiver module is middleware connecting the processor unit and the AWGR cluster; in this application, each optoelectronic transceiver module is a fully functional, multi-wavelength optoelectronic transceiver supporting wavelength division multiplexing (WDM). Figure 2As shown, each optoelectronic transceiver module includes a transmitter (Tx) and a receiver (Rx). The transmitting end, also known as the transmitting path, has the core function of converting parallel electrical signals from the xPU into optical signals carrying information. It contains a set of multi-wavelength light sources, a set of modulator arrays, and a multiplexer (MUX). The multi-wavelength light sources are used to generate multiple optical carriers of specific wavelengths. Each modulator in the modulator array corresponds to a specific wavelength channel. When the source xPU transmits electrical signal data, the driving circuit of the source-end opto-transceiver module (i.e., the opto-transceiver module at the source xPU end) selectively drives one or more modulators to work according to the routing configuration instructions, modulating the electrical signal data onto the optical carrier of the corresponding wavelength. Subsequently, the multiplexer can couple multiple optical signals from different modulators, carrying different information, into an output optical fiber. The receiving end, also known as the receiving path, has the core function of converting the received multi-wavelength optical signals (which can be referred to as multiplexed optical signals in this application) back into parallel electrical signals (which can be referred to as transmission electrical signals to distinguish them from the electrical signals generated by the source xPU). The receiving path contains a demultiplexer (DeMUX) and a set of photodetector arrays. When the target-end opto-transceiver module (i.e., the opto-transceiver module at the target xPU end) receives a WDM optical signal (i.e., a multiplexed optical signal) from the output port of the AWGR, the demultiplexer (e.g., based on a cascaded micro-ring filter or a stepped grating) first separates it into single-path optical signals of different wavelengths. Then, each photodetector in the photodetector array receives the single-path optical signal of the corresponding wavelength and converts it into a weak current signal. After being processed by circuits such as a transimpedance amplifier (TIA), it can be restored into a standard digital electrical signal and delivered to the target xPU.

[0035] As an example, the transmitting end of each optoelectronic transceiver module can integrate a distributed feedback (DFB) laser array, where each laser generates an optical signal of a specific and stable wavelength. The wavelength of the optical signal generated by the laser array can be expressed as ( , Alternatively, the following solutions can be used to replace the laser array: ① Tunable lasers: Each transmitter can use several tunable lasers that can quickly switch wavelengths over a wide spectral range. When communication with a specific target xPU is required, the tunable lasers quickly tune to a preset wavelength corresponding to the communication request according to the routing configuration instructions of the network scheduling controller. ② A centralized light source, in which one or more centralized high-power optical combs are deployed at the system level. This light source can generate a set of precisely spaced, dense wavelengths, which are then uniformly fed to all optoelectronic transceiver modules within the system through a passive optical distribution network (such as an optical splitter). This application does not specifically limit the optical carrier generating device.

[0036] As the core of the optical switching network, the AWGR cluster is crucial for realizing an all-optical interconnect topology. The arrayed waveguide grating router in the AWGR cluster is a passive optical device based on Planar Lightwave Circuit (PLC) technology. Its core function is to achieve deterministic cyclic wavelength routing. One AWGR can be deployed on one optical switching chip. The wavelength routing characteristics of the AWGR allow a direct, unobstructed optical path (i.e., optical channel) to be established between any pair of input and output ports by selecting a specific wavelength. This is the physical basis for achieving all-to-all communication in this invention. Specifically, when a wavelength division multiplexed optical signal containing multiple different wavelengths is injected from the i-th input port of any AWGR, the device performs spatial separation and routing based on the wavelength of the light: wavelength... The optical signal will be uniquely and passively routed to the k-th output port. The output port identifier k, the input port identifier i, and the wavelength identifier u follow a fixed, cyclic mathematical relationship, typically expressed as k = (i + u) mod N, where N is the total number of input (or output) ports of the AWGR. This means that for any source-side optoelectronic transceiver module (connected to input port i), as long as the wavelength of its transmitted optical signal is precisely controlled to k... (Each preset wavelength logically corresponds one-to-one with each target xPU in the optical switching system), thus passively determining its output from the k-th output port, thereby establishing a unique optical path. This characteristic enables the present invention to transform the routing function from the complex, high-power packet header parsing and dynamic routing decisions of traditional electrical switching into precise control of the light source wavelength at the network edge. Furthermore, the arrayed waveguide grating router itself consumes no power, its optical transmission delay is on the picosecond scale, and it is naturally immune to electromagnetic interference, making it an ideal device for building high-efficiency, high-density optical switching networks.

[0037] It should be noted that the wavelengths of the optical signals transmitted by the optoelectronic transceiver module and the AWGR in this application are all preset operating wavelengths. That is, the light source, modulator, multiplexer, demultiplexer and photodetector in the optoelectronic transceiver module are all configured according to multiple preset operating wavelengths, and the optical channel formed by the input port and output port of the AWGR is used to transmit the combined optical signal formed by multiple specific operating wavelength signals.

[0038] Physically, the network scheduling controller can be constructed using a high-performance Field-Programmable Gate Array (FPGA), which internally embeds the core control logic of this invention to balance performance and flexibility. The network scheduling controller can establish bidirectional links with key nodes in the optical switching network (including all processor units and optoelectronic transceiver modules) through an independent, out-of-band control bus network to collect network load status information in real time and issue routing configuration commands. The above-described use of an FPGA to construct the network scheduling controller and the coordination between the optical switching network and the network scheduling controller via the control bus are merely examples; this invention is not limited to these. For example, in scenarios involving large-scale deployment and cost and power consumption sensitivity, the proven and mature scheduling and fault management logic can also be embedded into a dedicated ASIC chip. ASIC solutions offer optimal performance, cost, and power consumption, but their design cycle is long and lacks flexibility for subsequent upgrades. For some systems with less stringent control latency requirements, the network scheduling function can also be implemented by high-performance software running on one or more dedicated CPU cores. This pure software solution has the shortest development cycle and the highest flexibility, but its control latency and throughput are generally inferior to hardware implementations.

[0039] Furthermore, this application uses the network scheduling controller as the control center of the entire optical switching network, which can monitor the optical switching network in real time and collect the status information and performance monitoring data of the entire network in real time, thereby realizing intelligent scheduling and high reliability assurance of physical layer resources. Figure 3As shown, the internal logic of the network scheduling controller can be divided into several collaborative functional modules (which can be represented as IP cores in hardware implementation), including a communication scheduling module, an interface control module, a status monitoring module, and a fault management module. The communication scheduling module, as the core computing engine for executing dynamic load balancing algorithms, obtains communication requests through the interface control module and, in conjunction with real-time load status information of the optical switching network read from the status database, executes the target optimization algorithm to determine the scheduling path. This scheduling path clarifies the optical switching plane required for data transmission and the specific wavelength optical channels within the AWGR, thereby achieving dynamic load balancing. The interface control module is responsible for handling control plane communication protocols with the xPU and external management systems (such as upper-layer applications). This module not only parses communication requests (such as API calls) from upper-layer applications and transforms them into internal scheduling tasks, but also provides feedback on the network load status to upper-layer applications. The status monitoring module collects key performance indicators (such as received optical power, bit error rate, and latency—network transmission performance parameters related to the load status of the optical switching network—from all optoelectronic transceiver modules and processor units through periodic queries or telemetry mechanisms, and stores this data in the status database, providing a basis for scheduling and fault management decisions. The fault management module utilizes a built-in "perception-decision-rerouting" closed-loop mechanism to automatically detect and isolate hardware faults, ensuring high system reliability. Specifically, the subset consisting of the optoelectronic transceiver module and the AWGR in this application can form an independent optical switching plane, providing parallel physical links for communication between processor units. Specifically, an AWGR and its communication connection to a source-end optoelectronic transceiver module and a target-end optoelectronic transceiver module can constitute an independent optical switching plane.

[0040] As an example, to achieve fine-grained management of the optical switching network, the network scheduling controller maintains two types of core static information: first, the overall network topology information, which records in detail the correspondence between each processor unit and the optoelectronic transceiver module array, as well as the communication connection relationship between each optoelectronic transceiver module and each specific port of each AWGR in the AWGR cluster; second, the wavelength routing mapping relationship, which records the inherent wavelength cyclic routing characteristics of each AWGR in the AWGR cluster, i.e., the deterministic mapping relationship between "AWGR input port - working wavelength - AWGR output port". In addition, the network scheduling controller also maintains dynamically updated load status information to indicate the real-time transmission performance of the network. This information may include, but is not limited to, throughput parameters (such as link bandwidth utilization), signal quality parameters (such as optical signal-to-noise ratio (OSNR) and bit error rate (BER), and delay parameters, etc., which are network transmission performance parameters related to the load status of the optical switching network. It should be noted that the "optical path" mentioned in this application specifically refers to an optical path between an input port and an output port activated by a specific working wavelength in an AWGR; while the "working wavelength" refers to the center wavelength of this optical path.

[0041] Figure 4 This is a flowchart illustrating the optical switching network scheduling and control method for the processor unit cluster proposed in this application. The method is executed by the network scheduling controller, and as follows... Figure 4 As shown, the method includes steps S110 to S130. Furthermore, for simplicity, the following description uses the example of a network scheduling controller and an optical switching network establishing a communication connection via a control bus.

[0042] Step S110: Receive a communication request sent by the source xPU. The communication request includes at least the identification information of the source xPU and the identification information of the target xPU that will communicate with the source xPU. Based on the number of target xPUs, this method can support multiple communication modes such as unicast, multicast, and broadcast. The communication mode is determined based on the number of target xPUs in the communication request: Unicast communication, i.e., point-to-point directional data transmission, has only one target xPU. Its implementation depends on the precise selection of the optical signal wavelength and the inherent passive wavelength routing characteristics of the AWGR device (i.e., the routing mapping relationship between the inherent operating wavelength of the AWGR and the ports (including input ports and output ports)). Multicast communication corresponds to multiple target xPUs. Its implementation mechanism is that the optoelectronic transceiver module uses wavelength division multiplexing technology to modulate and activate multiple wavelength channels in parallel. Broadcast communication (the target xPUs are all other xPUs except the source xPU) is a special type of multicast communication, and its implementation principle is similar to multicast.

[0043] In some embodiments of the present invention, the communication request may further include bandwidth requirements and quality of service (QoS) requirements, facilitating more refined resource allocation by the network scheduling controller. The bandwidth requirement defines the transmission rate required for the communication task, while the QoS requirement can define specific performance thresholds, such as the minimum acceptable optical signal-to-noise ratio (OSNR) or the maximum tolerable bit error rate (BER), serving as important constraints for subsequent path selection.

[0044] Step S120: Using a target optimization algorithm, based on load status information and communication requests, dynamically determine the scheduling path between the source xPU and the target xPU, as well as the preset operating wavelength corresponding to the communication request. Specifically, this step first identifies all candidate optical paths (i.e., candidate paths) connecting the source xPU and the target xPU based on the network's physical connection topology (i.e., the entire network topology information), preset component communication connection rules (also known as candidate optical path selection strategies), and the inherent wavelength-port routing mapping relationship of the AWGR. Then, using the target optimization algorithm and combined with real-time load status information, select one or more paths from the candidate path set as the final scheduling path corresponding to the communication request.

[0045] In this application, a potential optical path can be formed when the optoelectronic transceiver module corresponding to a source xPU, the optoelectronic transceiver module corresponding to a target xPU, and an AWGR connecting the source and target optoelectronic transceiver modules are determined. The specific operating wavelength used by this optical path is uniquely determined by the cyclic routing characteristics of the AWGR and the AWGR port identifier connected to the source / target xPU optoelectronic transceiver modules.

[0046] In some embodiments of the present invention, to simplify path calculation and achieve structured resource allocation, the physical topology of the optical switching network can be configured as follows: ① The number M of optoelectronic transceiver modules in the optoelectronic transceiver module array corresponding to each processor unit is no more than the number of AWGRs in the AWGR cluster; and the number of input ports of each AWGR is no less than the total number N of xPUs in the network, and the number of output ports of each AWGR is no less than the total number N of xPUs in the network. This ensures that each AWGR can form an independent optical switching plane with full connectivity potential. ② To achieve bandwidth capacity expansion, the optical switching network deploys optoelectronic transceiver module arrays and AWGR clusters in parallel. Under this premise, the present invention defines a fixed set of component communication connection rules, which are the basis for determining the candidate optical path set. The rules include: identifying the corresponding component / port that can perform data transmission; for each candidate optical path connecting the source processor unit and the target processor unit, the identifier of the AWGR input port corresponds to the identifier of the source processor unit xPU, the identifier of the AWGR output port corresponds to the identifier of the target processor unit xPU, and the identifiers of the source-end opto-transceiver module and the target-end opto-transceiver module in the candidate optical path correspond to the identifier of the AWGR in the candidate optical path (the AWGR, source-end opto-transceiver module and target-end opto-transceiver module corresponding to the identifier form an optical switching plane).

[0047] For example, this application can design each AWGR in the optical switching network to have a unique identifier, and each input port and each output port on each AWGR to have a unique identifier, and the input port identifier and output port identifier are the same between AWGRs with different identifiers. Furthermore, in this application, each processor unit has a unique identifier, each optoelectronic transceiver module associated with each processor unit has a unique identifier, and the identifiers of the optoelectronic transceiver modules are the same between optoelectronic transceiver module arrays corresponding to different processor units. The identifier can be an identifier sequence number; if the identifier sequence numbers are the same, they are considered to correspond to the identifiers. Therefore, the component communication connection rules can be as follows: each AWGR with a different identifier sequence number communicates with multiple optoelectronic transceiver modules corresponding to each xPU. The source xPU always communicates through the input port of the AWGR with the same identifier sequence number connected to it, and the target xPU always communicates through the output port of the AWGR with the same identifier sequence number connected to it. Simultaneously, the k-th optoelectronic transceiver module of the source xPU and the k-th optoelectronic transceiver module of the target xPU are both fixedly connected to the k-th AWGR in the AWGR cluster.

[0048] As an example, in this application, all processor units, all optoelectronic transceiver modules in each optoelectronic transceiver module array, and all AWGRs in the AWGR cluster are deployed in parallel, and these parallel-deployed components are all assigned unique identifiers. The i-th xPU is denoted as xPU-i; the k-th AWGR is denoted as AWGR-k; and the k-th optoelectronic transceiver module on xPU-i is denoted as module-ik. The input / output ports of the AWGR are also identified according to the xPU to which they are connected. In this case, the above-mentioned component communication connection rules can be more precisely described as follows: when xPU-i needs to send data to xPU-j, if the k-th AWGR is chosen, the data flow must follow the following fixed path: xPU-i -> source module-ik -> input port i of AWGR-k -> output port j of AWGR-k -> target module-jk -> xPU-j.

[0049] Therefore, the above component communication connection rules can be summarized as follows: the AWGR input port identifier used by the candidate path must correspond to the identifier of the source xPU, and the output port identifier must correspond to the identifier of the target xPU; simultaneously, the identifiers of the source / target optoelectronic transceiver modules involved in the optical path must correspond to the identifier of the AWGR used. Under this structured connection rule, for a given source / target xPU pair, the candidate path can be determined by selecting the identifier k of the AWGR.

[0050] Furthermore, the objective optimization algorithm, after determining all candidate paths, executes a dynamic selection decision mechanism. It selects a scheduling optical path from the candidate optical paths based on one-dimensional or multi-dimensional indicators from real-time load status information to maximize / minimize the objective function. The objective function can be designed based on cost and / or the transmission performance of the optical switching network. For example, in this application, the algorithm is based on a hierarchical dynamic scheduling mechanism with a quantified cost function. Its core idea is to calculate a comprehensive cost for each potential communication path based on the real-time acquired load status information and select a scheduling path based on the cost (e.g., selecting the candidate optical path with the optimal cost as the scheduling path). The objective optimization algorithm can also aim at optimizing the overall transmission performance of the optical switching network (i.e., a comprehensive goal of maximizing optical switching network throughput, optimizing resource utilization, and ensuring communication quality), or at achieving traffic balance in the optical switching network, or at maximizing the transmission rate of the optical switching network. Multiple objectives can be set to execute the objective optimization algorithm; this invention is not limited to these. If a scheduling path is selected between the source xPU and the target xPU, the scheduling path can be the candidate path with the lowest overall cost, or any candidate path with a cost less than a set threshold. That is, this application does not specifically limit the method of selecting a scheduling path from the candidate paths based on the overall cost.

[0051] As an example, the objective optimization algorithm is implemented using a cost function that aims to optimize the overall transmission performance of the optical switching network. This function weights and sums multiple performance metrics from the load status information, such as load balancing, port resource utilization, and transmission delay, thereby flexibly balancing the three key performance dimensions of throughput, signal quality, and latency. Specifically, the network scheduling controller collects real-time performance data from each component through a status monitoring module, such as bandwidth utilization of each optical path, the number of activated wavelengths at each port, and end-to-end transmission delay. This raw data is then aggregated and processed to calculate the parameters of each dimension in the cost function. Based on this, for a communication request from xPU-i to xPU-j, the path cost function for the candidate path determined based on the k-th optical switching plane (which is constructed based on the k-th AWGR) can be defined as: ; in, , and These are the plane load factor, port congestion factor, and delay penalty factor, respectively. , and The configurable weight coefficients for the three factors respectively, and satisfy the following conditions: + =1. By adjusting the weighting coefficients, priorities can be set for different optimization objectives. These factors can be calculated by the communication scheduling module based on real-time load status information in the status database. , and The specific calculation method is as follows: ① Plane Loading Factor This factor belongs to the throughput dimension and is used to measure the overall load level of the k-th optical switching plane. Its absolute load is obtained by summing the allocated bandwidth of all active optical channels on this optical switching plane. And based on the theoretical maximum total bandwidth of the plane. Normalization is performed to obtain the plane loading factor, i.e. When this factor is given a higher weight, the scheduling algorithm will prioritize optical switching planes with lower loads to achieve macroscopic traffic balance.

[0052] ②Port congestion factor This factor belongs to the signal quality dimension and is used to quantify the wavelength resource occupancy of the source xPU on the k-th optical switching plane. The network scheduling controller maintains and updates in real time the number of wavelengths currently activated (i.e., used for other multicast / broadcast tasks) on the i-th input port of the k-th optical switching plane. The normalized congestion factor can be obtained by comparing it with the maximum number of wavelengths available at that port (N-1). This factor quantifies the congestion level of a port. When this factor is given a higher weight, the algorithm will tend to select ports with lower wavelength multiplexing for communication, which helps minimize potential optical crosstalk between wavelength division multiplexing channels and ensure signal quality. This factor can assist the network scheduling controller in making fine-grained micro-level decisions when multiple planes have similar macroscopic loads.

[0053] ③ Delayed penalty factor This factor belongs to the delay dimension and reflects the real-time transmission delay of a specific optical path. The network scheduling controller obtains and maintains the real-time delay from xPU-i to xPU-j through the k-th optical switching plane by periodically probing end-to-end or from telemetry data. By using a preset maximum acceptable delay. By comparing these parameters, a normalized latency factor is obtained. For latency-sensitive tasks (such as barrier operations in synchronous computation or perception decision-making in autonomous driving), this can be significantly improved. The weighting of latency makes it the dominant factor in path selection, thereby enabling differentiated scheduling based on quality of service.

[0054] The cost function factors and weighted summation methods described above are merely examples. Optimization objectives can also incorporate parameters from other dimensions (such as packet loss rate), and the objective optimization algorithm itself can employ other types, such as heuristic algorithms or reinforcement learning-based scheduling models. Furthermore, scheduling mechanisms can be combined with redundancy strategies. For instance, under normal operating conditions, load balancing algorithms can distribute services across all available optical switching planes, achieving full resource utilization and avoiding the idle waste of backup resources in traditional primary / backup solutions.

[0055] In some embodiments of the present invention, step S120 may determine at least one optical switching plane based on the communication request. Before selecting a scheduling path, the method further includes: when the bandwidth requirement included in the communication request exceeds the transmission rate of a single optical channel, the network scheduling controller will initiate a link aggregation mechanism. Specifically, the network scheduling controller first calculates the number of optical channels that need to be used in parallel (equivalent to the number of optical switching planes required) based on the total bandwidth requirement for data transmission from the source xPU to the target xPU and the available rate of a single optical channel in the candidate optical path set. Subsequently, when executing the target optimization algorithm, a sufficient number of scheduling paths that meet the target are selected from the candidate path set. Finally, the source xPU will slice the data stream to be transmitted and transmit the data slices in parallel through these selected scheduling optical paths. For example, a 200Gbps transmission task, in a system with a single channel rate of 100Gbps, will be allocated to two parallel scheduling optical paths, thereby achieving an aggregated bandwidth of 2×100Gbps.

[0056] Furthermore, if the network scheduling controller receives multiple communication requests from the same source xPU or target xPU simultaneously within a short period (i.e., multi-task concurrency), the scheduling algorithm will calculate and allocate an optimal scheduling path for each independent communication request from the set of available candidate paths at that time. Typically, to achieve load balancing, these concurrent tasks will be distributed across different optical switching planes.

[0057] In summary, whether to meet high bandwidth requirements (link aggregation) or to handle multi-task concurrency, the network scheduling controller can distribute data streams to multiple different optical switching planes for parallel transmission. In unicast scenarios where bandwidth requirements can be met by a single channel, to improve transmission robustness or utilize idle resources, the scheduling algorithm can also selectively copy or split a single data stream onto multiple optical switching planes for parallel transmission.

[0058] The quantization scheduling mechanism based on the objective optimization algorithm in step S120 enables the dynamic load balancing function of the network scheduling controller to be configurable and adaptive, thereby optimizing it according to the needs of the application scenario. As an example, the complete execution flow of step S120 can be summarized as follows: Parsing the request and identifying the candidate path set: The optical network control device first determines all available candidate optical paths based on the source / target xPU identifier (i, j) in the communication request and in conjunction with preset component communication connection rules. In a structured topology, this is equivalent to determining all available optical switching plane identifiers k, where k∈{1, ..., M}.

[0059] Parallel computing path cost: Utilizing the computing capabilities of the chip (such as the parallel processing of an FPGA), and based on real-time load status information, calculate the corresponding path cost C(i, j, k) for each candidate path simultaneously.

[0060] Based on the strategy, the network scheduling controller selects the final scheduling optical path from the candidate optical paths whose costs have been calculated, according to the specific communication requirements (such as whether link aggregation is required) and the preset selection strategy. For example, in a unicast task that requires only one scheduling optical path, the network scheduling controller will select the candidate path that minimizes the cost function C(i, j, k), and its corresponding switching plane identifier k_optimal can be expressed as: k_optimal = argmin {C(i,j, k)} for k in {1, ..., M}.

[0061] Step S130: Generate routing configuration instructions based on the scheduling path selected by the communication request, and send them to the optical switching network through the control bus, thereby realizing communication scheduling from the source xPU to the target xPU.

[0062] More specifically, the generation and distribution process of routing configuration instructions can be broken down as follows: The communication scheduling module distributes a routing decision containing the final scheduling path information to the interface control module. This routing decision clarifies the identifiers of the source / target xPU, the source / target opto-transceiver module, the AWGR, and the operating wavelength uniquely determined by these components. Subsequently, the interface control module generates routing configuration instructions conforming to the control plane communication protocol based on the routing decision and distributes them via the control bus to the specified source processing unit and the corresponding opto-transceiver module (referring to the source opto-transceiver module) in the optical switching network. Finally, the driver on the source xPU parses the instruction and completes the specific configuration of the corresponding opto-transceiver module and AWGR via the board-level control bus. Therefore, the routing configuration instruction can include the operating wavelength specified for the scheduling optical path and the configuration information of the opto-transceiver module and AWGR in the scheduling optical path. The configuration information of the opto-transceiver module and AWGR is the identification information of the opto-transceiver module and AWGR.

[0063] It should be noted that although a bidirectional communication connection can be established between the source processor unit and the target processor unit through the method mentioned in this application, the processor unit interconnection in this application is only a terminology expression, and its actual meaning is to establish a communication connection from the source processor unit to the target processor unit.

[0064] Once a communication task is successfully scheduled (e.g., the network scheduling controller determines successful scheduling after receiving a scheduling success notification from the optical switching network), the network scheduling controller immediately updates and maintains the status of affected network resources (such as the bandwidth occupancy of the selected optical path) in its internal status database to provide a reference for the next decision cycle. Furthermore, to ensure the continued effectiveness of the decision-making basis, the scheduling control method proposed in this application also includes a periodically executed status monitoring step S140: periodically collecting network-wide load status information and updating the status database. This step is continuously executed by the status monitoring module, which collects real-time performance indicators from all processor units and optoelectronic transceiver modules in the optical switching network by periodically querying or receiving telemetry data, and uses this latest information to update the load status database stored internally by the network scheduling controller, providing accurate data support for the next scheduling decision cycle.

[0065] In some embodiments of the present invention, to ensure the high reliability of the optical switching system in complex production environments, the network scheduling controller and the data plane collaboratively construct an efficient fault self-healing mechanism. This function is mainly executed by the fault management module within the network scheduling controller. The fault self-healing workflow follows a complete "perception-decision-rerouting" closed-loop logic, with the specific steps as follows: Fault Detection: The status monitoring module can achieve real-time fault detection through two methods. The first is by continuously analyzing performance indicators (i.e., network transmission performance parameters related to the load status of the optical switching network) collected by the status monitoring module. When it detects response timeouts, a sharp deterioration in the bit error rate (BER), or an abnormal decrease in received optical power (ROP) (i.e., when the real-time load status information does not meet set conditions), it proactively determines component failure. The set conditions can be configured (e.g., setting the condition that the real-time load status information exceeds a threshold). The second method is by receiving explicit alarms uploaded by the xPU driver (abnormal events such as communication timeouts or data verification errors reported by the xPU driver layer) to identify faulty components in the optical switching network, thus achieving fault detection. Fault Decision and Isolation: Upon detecting a fault in the optical switching network, an isolation command is generated and sent to the optical switching network to logically isolate the faulty component. The isolation command includes the identification information of the faulty component and a flag for the isolation command (used by the xPU or optoelectronic transceiver module to determine that the received command is an isolation command). That is, once an anomaly is detected, the fault management module can immediately initiate a diagnostic process, analyze and diagnose by injecting diagnostic optical packets to distinguish and locate the scope of the fault. For example, it can determine whether the fault is a port-level fault affecting a single link or a plane-level fault affecting the entire switching plane. After the fault is accurately located, the fault management module can immediately update the internally maintained system resource status table, marking the confirmed faulty component as unavailable, thereby logically completing fault isolation and preventing new communication tasks from being assigned to the faulty resource. Fault Rerouting (Service Recovery): If it is determined that the faulty component is carrying transmission data, a new working scheduling path will be determined based on the communication request corresponding to the transmission data. That is, for the original active communication flow on the isolated resource, the fault management module will trigger a forced rerouting and resubmit the metadata of these affected communication tasks to the communication scheduling module. Then, the communication scheduling module will call the target optimization algorithm to recalculate a new scheduling path for these interrupted tasks among all available and healthy physical resources, and issue the updated routing configuration command to the relevant xPU to restore data transmission on the new path.

[0066] As an example, during the rerouting process, after the network scheduling controller identifies the faulty component, it can determine whether the faulty component is carrying transmission data based on the real-time load status information. If the faulty component is carrying transmission data, the transmission process of the data is restricted, and a new scheduling path needs to be determined to complete the communication task: if the source xPU corresponding to the transmission data does not receive the instruction to complete the communication task within a certain period of time, it will regenerate the communication request corresponding to the transmission data and send it to the network scheduling controller to re-determine the scheduling path; or, after determining that the faulty component is carrying transmission data, the network scheduling controller will trace the communication request corresponding to the transmission data uploaded by the source xPU, regenerate the routing configuration instruction, and send it to the optical switching network so that the source xPU can resend the electrical signal to the corresponding optoelectronic transceiver module. In this application, "the faulty component is carrying transmission data" refers to the fact that the scheduling path to which the faulty component belongs carries transmission data.

[0067] Steps S110-S130 describe the collaborative decision-making process of the network scheduling controller. The routing configuration instructions generated by this process guide the optical switching network to complete a specific communication scheduling from the source xPU to the target xPU. The communication scheduling process from the source xPU to the target xPU can be summarized as follows: The source xPU sends the electrical signal to be transmitted (the first electrical signal) to a specific source-end opto-transceiver module according to the instructions; the opto-transceiver module configures its transmitting end according to the instructions, modulates the first electrical signal onto an optical carrier of a specified operating wavelength, and generates a transmitted optical signal; the optical signal is passively routed to the specific opto-transceiver module connected to the target xPU via a specified AWGR; the target-end opto-transceiver module detects the optical signal at its receiving end, and after restoring it to an electrical signal (the second electrical signal), it finally transmits it to the target xPU.

[0068] In some embodiments of the present invention, this application proposes an interconnection method for processor units, which is performed by an optical switching network (data plane). For example... Figure 1 As shown, each processor unit in the optical switching network is connected to a corresponding optoelectronic transceiver module in the associated optoelectronic transceiver module array, and each optoelectronic transceiver module is connected to a passive AWGR via optical fiber. Figure 5 As shown, the interconnection method includes the following steps: Step S210: The source xPU receives and parses the routing configuration instruction, and sends an electrical signal according to the instruction. Specifically, the source xPU first sends a communication request to the network scheduling controller and receives the routing configuration instruction calculated for its communication request from the optical network scheduling controller. This instruction specifies the scheduling optical path to be used in this communication (including the optoelectronic transceiver module and AWGR and its corresponding operating wavelength). Subsequently, according to the instruction, the source xPU sends the data to be transmitted (the first electrical signal) through its high-speed I / O interface to the source optoelectronic transceiver module specified in the instruction.

[0069] Step S220: The source-end optoelectronic transceiver module performs electro-optical conversion. The optoelectronic transceiver module corresponding to the source xPU in the scheduling path receives the routing configuration command from the network scheduling controller and the first electrical signal from the source xPU. Based on the preset wavelength in the routing configuration command, it activates the corresponding modulator and uses the activated modulator to modulate the received electrical signal onto an optical carrier of a specific wavelength to generate an optical signal carrying information.

[0070] As an example, in multicast or broadcast communication scenarios, routing configuration instructions may include multiple operating wavelengths. In this case, the source-side optoelectronic transceiver module will activate multiple modulators in parallel and modulate the same electrical signal data onto multiple different optical carriers. Its internal multiplexer will couple these multiple optical signals into a single WDM optical signal as the final output optical signal.

[0071] Step S230: Passive optical routing is performed by the arrayed waveguide grating router device AWGR. According to the scheduling path information in the routing configuration instruction, the optoelectronic transceiver module at the source end specified in the scheduling path sends the transmitted optical signal to the AWGR specified in the scheduling path. The AWGR will passively and deterministically route one or more wavelengths contained in the transmitted optical signal to one or more corresponding output ports according to the inherent wavelength-port routing mapping relationship. These output ports are connected to the optoelectronic transceiver module corresponding to the target xPU.

[0072] Step S240: The target-side optoelectronic transceiver module performs optoelectronic conversion. One or more target-side optoelectronic transceiver modules receive the optical signal from the AWGR in their receiving path. Their internal demultiplexer first separates the optical signal of a specific wavelength from the received multiplexed optical signal, and then the corresponding photodetector recovers the separated optical signal into an electrical signal (the second electrical signal), which is finally delivered to the target xPU, thus completing a full communication cycle.

[0073] In some embodiments of this invention, the optical switching system comprises N xPUs, each associated with M independent optoelectronic transceiver modules via their I / O interfaces. Each processor unit's associated optoelectronic transceiver module array contains the same number of optoelectronic transceiver modules; therefore, the optical switching network can contain N×M optoelectronic transceiver modules. The optical switching core is crucial for achieving passive data switching and consists of multiple parallel optical switching planes. The core device of each optical switching plane is an N×N port arrayed waveguide grating router. This design of multiple parallel optical switching planes is the core architecture for modular bandwidth expansion in this invention, ensuring that the M optical switching planes can operate in parallel both physically and logically, jointly carrying the data traffic in the system. Furthermore, the input port corresponding to the optical channel of the AWGR in the scheduling path corresponds to the source xPU, and the output port corresponding to the optical channel in the scheduling path corresponds to the target xPU. Additionally, the optoelectronic transceiver modules corresponding to the source xPU and the target xPU in the scheduling path correspond to the AWGR in the scheduling path.

[0074] As an example, it is assumed that all routing decisions have been made by the upper-layer network scheduling controller, and routing configuration instructions have been generated. The working principle of the physical layer optical switching system (data plane) and the specific workflow for performing communication tasks under the guidance of the network scheduling controller will be described in detail below.

[0075] ① Assuming a fully interconnected optical switching system configured with N=8 xPUs and M=4 AWGR parallel optical switching planes, and a unicast communication needs to be established from xPU-2 to xPU-6, and the network scheduling controller, after optimization calculation, decides to execute this communication task on the third optical switching plane (i.e., the plane composed of AWGR-3), the physical workflow is as follows: (1) Command Reception and Module Configuration: The source processor unit first receives a routing configuration command from the network scheduling controller. This command specifies that the source xPU for this communication is xPU-2, and uses the third optoelectronic transceiver module (hereinafter referred to as module 2.3) corresponding to the source xPU, the target xPU is xPU-6, and uses the third optoelectronic transceiver module (hereinafter referred to as module 6.3) and the third optical switching plane corresponding to the target xPU, and specifies the corresponding operating wavelength. Subsequently, xPU-2 configures module 2.3 through its board-level control bus.

[0076] (2) Photoelectric conversion: According to the configuration, the transmission path (including modulator and light source) of module 2.3 is activated. Its internal light source provides the operating wavelength specified by the command (the operating wavelength required from input port 2 to output port 6 according to the predefined wavelength-port routing mapping relationship; in this application, this operating wavelength can be referred to as λ). 2-6 The module's modulator then modulates the serial electrical signal data stream (e.g., at a rate of 100 Gbps) from the xPU-2 onto a matching optical carrier. 2-6 On the optical carrier.

[0077] (3) Passive optical switching: λ carrying data 2-6 The optical signal is injected from module 2.3 into input port 2 of the third arrayed waveguide grating router (AWGR-3) via optical fiber. Based on the inherent physical characteristics of the AWGR, this λ... 2-6 The optical signal will be passively and deterministically routed to output port 6 of AWGR-3.

[0078] (4) Photoelectric conversion: The optical signal output from the 6th output port of AWGR-3 enters the third photoelectric transceiver module associated with xPU-6 (referred to as module 6.3). At the receiving end of module 6.3, the demultiplexer converts the wavelength λ 2-6 The optical signal is directed to the corresponding photodetector, which converts it into an electrical signal data stream (e.g., at a rate of 100 Gbps).

[0079] (5) Data delivery: The restored electrical signal is amplified and shaped by circuits such as the transimpedance amplifier (TIA) and then delivered to the I / O interface of xPU-6, thus completing an end-to-end communication from xPU-2 to xPU-6. The establishment of the entire data plane communication path and data exchange are completed by pre-configured and passive optical devices, without the need for real-time parsing and routing of data packets, thus possessing the characteristics of extremely low latency and non-blocking physical layer.

[0080] ② Assuming a fully interconnected optical switching system configured with N=8 xPUs and M=4 AWGR parallel optical switching planes, xPU-3 needs to send the same data simultaneously to xPU-2, xPU-5, and xPU-6 (i.e., a multicast communication scenario), and the network scheduling controller decides that this communication task should be executed on the second optical switching plane, the physical workflow is as follows: Figure 6 As shown: (1) Command Reception and Multi-Wavelength Configuration: The source processor unit xPU-3 receives a routing configuration command from the network scheduling controller. This command contains a list of target xPUs {xPU-2, xPU-5, xPU-6}, specifies that the second optoelectronic transceiver module (module 3.2) associated with the source xPU and the second optoelectronic transceiver module associated with the target xPU will be used in this communication, and specifies the three operating wavelengths required to reach these three target xPUs respectively. Subsequently, xPU-3 configures module 3.2.

[0081] (2) Multi-wavelength parallel modulation: According to the configuration, the transmission path of module 3.2 is activated and three wavelengths with specified operating wavelengths are driven in parallel (using the second optical switching plane, the wavelengths required from input port 3 to output ports 2 / 5 / 6 are λ respectively). 3-2 , λ 3-5 , λ 3-6 Matching light source and modulator. The same serial electrical signal data stream from xPU-3 is simultaneously modulated onto three optical carriers of different wavelengths.

[0082] (3) Wavelength division multiplexing and transmission: The three modulated optical signals of different wavelengths are combined by a multiplexer (MUX) inside the transmitter of module 3.2 to form a WDM signal. The WDM signal is output from the optical fiber port of module 3.2 and injected into the third input port of AWGR-2.

[0083] (4) Passive wavelength routing and optical switching: When the WDM optical signal enters the AWGR-2, the device, based on its inherent wavelength routing characteristics, spatially separates the different wavelength components in the wavelength division multiplexing signal and routes them to different output ports: λ 3-2 The optical signal is routed to output port 2, λ 3-5 To output port 5, λ 3-6To output port 6. This process is completely passive and requires no external control or configuration updates.

[0084] (5) Distributed parallel reception: The second optoelectronic transceiver modules (i.e., modules 2.2, 5.2 and 6.2) associated with xPU-2, xPU-5 and xPU-6 respectively receive their respective optical wavelengths and corresponding data through their corresponding AWGR output ports, thereby completing a one-to-many multicast communication.

[0085] ③ Assuming the fully interconnected optical switching system is configured with N xPUs and M AWGRs, when the source xPU needs to broadcast communication to all other N-1 destination xPUs, its physical implementation principle is exactly the same as multicast. The transmission path of the source-end optoelectronic transceiver module only needs to activate and modulate all N-1 working wavelengths corresponding to all N-1 target xPUs at the same time.

[0086] In some embodiments of the present invention, the system architecture proposed in this application can efficiently support multi-task concurrent communication scenarios, that is, to simultaneously perform communication scheduling tasks corresponding to multiple communication requests. For example... Figure 7 As shown, in an optical switching system configured with N=4 and M=4, assume the following three independent, concurrent communication tasks exist simultaneously: Task A (high-bandwidth unicast): xPU-1 transmits high-bandwidth data to xPU-2, requiring 200 Gbps bandwidth (assuming the available bandwidth rate of a single-plane, single-wavelength channel is 100 Gbps); Task B (normal unicast): xPU-1 transmits data to xPU-3; Task C (multicast): xPU-1 multicasts data to xPU-3 and xPU-4. The network scheduling controller can make the following efficient parallel scheduling decisions: (1) Bandwidth aggregation for Task A: To meet the 200 Gbps bandwidth requirement, the data stream of Task A is sliced ​​and allocated in parallel to the first and second optical switching planes. Modules 1.1 and 1.2 associated with xPU-1 will utilize the first and second optical switching planes corresponding to AWGR-1 and AWGR-2 respectively, while using the wavelength λ corresponding to the target xPU-2. 1-2 Data fragments are transmitted to xPU-2, thereby achieving an aggregate bandwidth of 2×100 Gbps.

[0087] (2) Concurrent scheduling of task B: Task B is concurrently allocated to the third optical switching plane corresponding to AWGR-3. Module 1.3 associated with xPU-1 will be connected to AWGR-3, utilizing the operating wavelength λ required from source xPU-1 to destination xPU-3. 1-3 Data is transmitted to xPU-3.

[0088] (3) Concurrent scheduling of task C: Task C is concurrently allocated to the fourth optical switching plane corresponding to AWGR-4. Module 1.4 associated with xPU-1 will pass through AWGR-4, simultaneously utilizing two wavelengths λ with target addresses xPU-3 and xPU-4. 1-3 and λ 1-4 Perform multicast operations to xPU-3 and xPU-4.

[0089] In this scenario, three communication tasks of different natures (high-bandwidth aggregation, ordinary unicast, and multicast) are executed in parallel on four physically isolated planes without interfering with each other. This capability significantly improves the system's total throughput and concurrent processing capacity, effectively avoiding resource contention and headend blocking problems that may occur in a single switching matrix. The multi-plane parallel task scheduling strategy proposed in this invention elevates the optical switching network from a simple switching unit into a reconfigurable communication processing matrix that supports large-scale concurrency. This allows the optical switching network to dynamically allocate physical plane resources for bandwidth aggregation or task concurrency based on real-time traffic demands, which is one of the core advantages of this invention.

[0090] This application does not specifically limit the physical integration and packaging form of the optoelectronic transceiver module. The system architecture of this invention is compatible with various cutting-edge optoelectronic integration technologies, such as board-level pluggable optical interconnects, co-packaged optics, 2.5D silicon interposer integration, and 3D heterogeneous integration. When using board-level pluggable optical interconnects, the xPU and the optoelectronic transceiver module are located on independent server boards and connected to a central switching board integrating multiple optical switching planes via high-density fiber optic connectors or an optical backplane. When using co-packaged optics, the xPU package and the optical engine containing optoelectronic transceiver functions are mounted adjacently on the same high-density printed circuit board as independent components, connected by short-distance electrical traces on the substrate. When using 2.5D silicon interposer integration, the xPU die and the optoelectronic transceiver module die, serving as photonic I / O dies, are mounted side-by-side with high precision on the same silicon interposer, interconnected by ultra-high-density wiring within the interposer. When employing 3D heterogeneous integration, the logic layer containing xPU transistors and the photonic layer containing photonic devices (modulators, photodetectors, etc.) can be vertically bonded together using hybrid wafer-to-wafer or die-to-wafer bonding techniques, achieving extremely high integration density and interconnect efficiency. These approaches represent different stages of evolution from board-level to on-chip integration, and the control and data plane collaborative architecture of this invention is applicable to all of them, thereby ensuring the long-term technological evolution potential of the solution.

[0091] This application proposes a scheduling and control method and an interconnection method for optical switching networks targeting processor clusters. The aim is to provide a system-level solution for large-scale processor clusters that combines superior performance, high reliability, and intelligent flow control capabilities through deep co-design of the data plane and control plane. This invention combines wavelength division multiplexing (WDM) and space division multiplexing (SDM) technologies to construct a powerful physical data plane and introduces a highly flexible and intelligent control plane, enabling dynamic optimization of network resources and fault self-healing. Specifically, the method proposed in this application has the following significant advantages: ① A decoupled system architecture is adopted, comprising an optical switching network (data plane) and a network scheduling controller (control plane). The optical switching network includes at least M parallel optical switching planes with N×N port arrayed waveguide grating routers as the switching core. This application replaces active electrical switching chips with passive photonic devices, breaking through the inherent bottleneck of electrical interconnection from a physical principle perspective. Utilizing the low-loss transmission characteristics of optical signals in waveguides and the passive switching characteristics of AWGRs, the power consumption and latency of high-speed interconnect links can be significantly reduced. By using wavelength division multiplexing technology in a single optical fiber and combining it with multi-plane parallel spatial multiplexing, an order-of-magnitude increase in interconnect bandwidth density can be achieved, overcoming the density limitations of traditional electrical switching I / O interfaces. At the same time, optical signals are naturally immune to electromagnetic interference, fundamentally solving the signal integrity problem in high-speed electrical design. The network scheduling controller is coupled and integrated into the optical switching network and configured to generate routing decisions to control data transmission in the optical switching plane and manage the operating status of the optical switching network. It can realize intelligent management of physical resources, elevating the static optical physical layer into a dynamic, programmable network system.

[0092] ② Cooperative communication technical solution. The network scheduling controller receives one or more communication requests and, based on the target optimization algorithm and real-time network status, calculates the optimal scheduling path (including the switching plane and operating wavelength) for each request, generating a routing configuration instruction. The source xPU, according to this instruction, modulates the data onto the specified operating wavelength through its associated optoelectronic transceiver module and routes the data to one or more target xPUs via the selected optical switching plane. This cooperative communication solution demonstrates significant benefits in terms of performance, energy efficiency, reliability, and scalability, ensuring optimal resource allocation.

[0093] ③ Native support for diverse communication modes. Through the system's multi-wavelength parallel modulation and multi-plane parallel switching mechanism, various high-efficiency communication modes (such as single-cycle hardware broadcasting and large-scale task concurrency) can be efficiently supported at the hardware level, providing a high-performance physical foundation for upper-layer applications. When performing multicast or broadcast communication with multiple target xPUs, the source xPU achieves this by modulating data in parallel onto multiple optical wavelengths corresponding to the multiple target xPUs; when performing high-bandwidth communication or multi-task concurrency, the network scheduling controller allocates communication tasks to multiple different optical switching planes for parallel execution. Therefore, compared to the mechanism in traditional electrical networks that relies on multiple packet copying and serial forwarding, the method proposed in this application can greatly improve communication efficiency.

[0094] ④ Dynamic load balancing is achieved using a multi-objective optimization algorithm. The control plane, through its multi-objective optimization algorithm, can comprehensively consider multiple dimensions such as the real-time load of each optical switching plane, the wavelength resource occupancy of the source ports, and end-to-end latency, dynamically calculating the optimal path. This capability ensures that physical resources are fully and evenly utilized, achieves traffic adaptation, and can provide differentiated services based on the QoS requirements of different tasks.

[0095] ⑤ Implement fault self-healing functionality. The network scheduling controller possesses fault diagnosis and dynamic rerouting capabilities. When the network scheduling controller detects a fault in any optical switching plane or its associated link, it logically isolates the faulty component and automatically reroutes the communication traffic originally carried on the faulty component to one or more healthy optical switching planes. This ensures service continuity and high system reliability in the event of hardware failure. For example, based on the fault self-healing capability of a distributed redundancy model, the optical switching system can automatically and quickly complete fault isolation and service recovery when faced with hardware failure.

[0096] Corresponding to the above method, the present invention also provides an optical switching network scheduling and control system for processor unit clusters. The system includes a computer device, which includes a processor and a memory. The memory stores computer programs / instructions. The processor is used to execute the computer programs / instructions stored in the memory. When the computer programs / instructions are executed by the processor, the system implements the steps of the method described above.

[0097] Those skilled in the art will understand that the exemplary components, systems, and methods described in conjunction with the embodiments disclosed herein can be implemented in hardware, software, or a combination of both. Whether implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this invention. When implemented in hardware, it can be, for example, electronic circuits, application-specific integrated circuits (ASICs), appropriate firmware, plug-ins, function cards, etc. When implemented in software, the elements of this invention are programs or code segments used to perform the desired tasks. The programs or code segments can be stored in a machine-readable medium or transmitted over a transmission medium or communication link via data signals carried in a carrier wave.

[0098] It should be clarified that the present invention is not limited to the specific configurations and processes described above and shown 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 shown as examples. However, the method process of the present invention is not limited to the specific steps described and shown. Those skilled in the art can make various changes, modifications, and additions, or change the order of steps, after understanding the spirit of the present invention.

[0099] In this invention, features described and / or illustrated for one embodiment may be used in the same or similar manner in one or more other embodiments, and / or combined with or in place of features of other embodiments.

[0100] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, various modifications and variations of the embodiments of the present invention are possible. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A scheduling and control method for optical switching networks oriented towards processor unit clusters, characterized in that, The method is executed by a network scheduling controller, which is communicatively connected to the optical switching network. The optical switching network consists of multiple components, including multiple processor units, optoelectronic transceiver module arrays at each processor unit, and an arrayed waveguide grating router (AWGR) cluster. The network scheduling controller stores the inherent wavelength-port routing mapping relationships of each AWGR in the AWGR cluster and the real-time load status information of the optical switching network. The method includes the following steps: Receive request: Receive a communication request sent by the source processor unit; wherein the communication request includes the identification information of the source processor unit and the identification information of the target processor unit to be interconnected; Candidate optical paths are determined: based on the communication request, the preset component communication connection rules, and the wavelength-port routing mapping relationship, all candidate optical paths connecting the source processor unit and the target processor unit are determined; wherein, each candidate optical path is defined by a source-end optoelectronic transceiver module, a target-end optoelectronic transceiver module, an AWGR, and a preset operating wavelength; Selecting a scheduling optical path: Based on the real-time load status information and using a target optimization algorithm, a scheduling optical path that meets the communication request is selected from the candidate optical paths; Generate and issue configuration instructions: Generate routing configuration instructions based on the scheduling optical path. The instructions include the operating wavelength specified by the scheduling optical path and the configuration information of the optoelectronic transceiver module and AWGR in the scheduling optical path. Issue the instructions to the optical switching network to establish a communication connection between the source processor unit and the target processor unit, thereby realizing communication scheduling control based on the communication connection.

2. The method according to claim 1, characterized in that, The number of optoelectronic transceiver modules in the optoelectronic transceiver module array at each processor unit end shall not exceed the number of AWGRs in the AWGR cluster, and the number of input ports and output ports corresponding to each AWGR shall not be less than the number of processor units in the optical switching network. In the optical switching network, each subset consisting of an optoelectronic transceiver module and an AWGR forms an independent optical switching plane, providing parallel physical links for communication between processor units; The component communication connection rules include: For each candidate optical path connecting the source processor unit and the target processor unit, the identifier of the AWGR input port corresponds to the identifier of the source processor unit, the identifier of the AWGR output port corresponds to the identifier of the target processor unit, and the identifiers of the source-end optoelectronic transceiver module and the target-end optoelectronic transceiver module in the candidate optical path correspond to the identifier of the AWGR in the candidate optical path.

3. The method according to claim 1, characterized in that, The load status information is a network transmission performance parameter used to indicate the load status of the optical switching network. The network transmission performance parameter includes one or more of throughput parameters, signal quality parameters, and delay parameters. The throughput parameter includes the used or remaining bandwidth of each connectable optical path, the signal quality parameter indicates the transmission quality of each connectable optical path, and the delay parameter indicates the transmission delay of each connectable optical path.

4. The method according to claim 1, characterized in that, The method further includes: collecting network load information from processor units and optoelectronic transceiver modules in the optical switching network by periodically querying or receiving event-triggered reports, and dynamically updating the stored real-time load status information.

5. The method according to claim 4, characterized in that, The network scheduling controller stores network topology information; The method also includes the following fault handling steps: Analyze the stored real-time load status information, and if the real-time load status information does not meet the set conditions, determine the corresponding component failure, or receive explicit component failure alarms uploaded by the processor unit or optoelectronic transceiver module to achieve fault detection. When a fault is detected in a specific component of the optical switching network, the network scheduling controller marks the faulty component as unavailable in the network topology information. This allows the faulty component marked as unavailable to be automatically excluded in the subsequent process of determining candidate optical paths, thus achieving logical isolation.

6. The method according to claim 5, characterized in that, The method further includes a service recovery step: when it is determined that an active data transmission service is carried on the faulty component, the network scheduling controller automatically triggers a rerouting calculation for the service; the rerouting calculation is used to re-determine all candidate optical paths that do not contain the faulty component through the candidate optical path determination step, select a new and available scheduling optical path through the scheduling optical path selection step, and issue an updated routing configuration instruction to the optical switching network through the configuration instruction generation and issuance step, so as to switch the data transmission service to the new scheduling optical path without loss or with minimal interruption.

7. The method according to claim 3, characterized in that, The communication request also includes a bandwidth requirement; the target optimization algorithm is used to select the optimal scheduled optical path that meets the bandwidth requirement from all candidate optical paths; When the remaining bandwidth of any single candidate optical path does not meet the bandwidth requirement, multiple candidate optical paths are selected together as the final parallel scheduling optical path, thereby satisfying the bandwidth requirement in the communication request through link aggregation.

8. A processor unit interconnection method, characterized in that, The method is executed in an optical switching network that is communicatively connected to a network scheduling controller. The optical switching network consists of multiple components, including multiple processor units, an array of optoelectronic transceiver modules at each processor unit, and an arrayed waveguide grating router (AWGR) cluster. Each optoelectronic transceiver module in the optoelectronic transceiver module array has wavelength division multiplexing capability, including a transmit path containing a modulator array and a receive path containing a photodetector array. The modulator array and photodetector array respectively contain modulators and photodetectors corresponding to each preset operating wavelength. The transmit path is used to simultaneously generate and modulate optical signals of one or more specific operating wavelengths, and the receive path is used to demultiplex and detect optical signals of multiple operating wavelengths. The method includes the following steps: Request and Configuration: The source processor unit sends a communication request to the network scheduling controller; and the source processor unit receives a routing configuration instruction corresponding to the communication request issued by the network scheduling controller; wherein, the routing configuration instruction includes scheduling optical path information between the source processor unit and the target processor unit; each scheduling optical path is defined by a source-end optoelectronic transceiver module, a target-end optoelectronic transceiver module, an AWGR, and a preset operating wavelength, and the scheduling optical path information includes the specified operating wavelength and the configuration information of the optoelectronic transceiver module and the AWGR; Electro-optical conversion: The source processor unit selects the source-end opto-transceiver module specified by the routing configuration instruction from the source-end opto-transceiver module array according to the routing configuration instruction; and the selected source-end opto-transceiver module activates the transmission path in the opto-transceiver module according to the specified operating wavelength, thereby modulating the electrical signal sent by the source processor unit onto the optical carrier of the specified operating wavelength to generate a transmitted optical signal; Optical signal routing: The selected source-end opto-transceiver module sends the transmitted optical signal to the AWGR specified by the routing configuration instruction; the AWGR, based on the wavelength of the transmitted optical signal and its inherent wavelength-port routing mapping relationship, routes the signal of each wavelength in the combined optical signal containing the transmitted optical signal to the target-end opto-transceiver module specified by the routing configuration instruction. Optical-to-electrical conversion: The target-end optoelectronic transceiver module performs wave demultiplexing on the multiplexed optical signal received by the receiving channel, separating it into optical signals corresponding to the specified operating wavelength, and converting the separated optical signals into transmission electrical signals and sending them to the target processor unit, thereby completing end-to-end communication.

9. The method according to claim 8, characterized in that, The scheduling optical path information contained in the routing configuration instruction is determined by the network scheduling controller executing the scheduling control method.

10. A scheduling and control system for an optical switching network oriented to a processor unit cluster, comprising a processor, a memory, and computer programs / instructions stored in the memory, characterized in that, The processor is configured to execute the computer program / instructions, and when the computer program / instructions are executed, the system implements the steps of the method as described in any one of claims 1 to 7.

Images