An awgr-based optical switching network architecture, scheduling method, medium, and device

Abstract

The application relates to the technical field of optical switching network and discloses an AWGR-based optical switching network architecture, a scheduling method, a medium and equipment, which comprise a graphic processor module, an optical module, an AWGR and a centralized central light source; the centralized central light source provides the optical module with a light source and uniformly schedules wavelength resources; the graphic processor modules are connected through the optical module and the AWGR to form the optical switching network architecture; when the centralized central light source provides the optical module with a light source and uniformly schedules wavelength resources, the communication wavelength is calculated in combination with the basic frequency of the AWGR, the device frequency offset constant, the frequency interval of adjacent communication channels and temperature drift and is provided to the connected optical module; the communication optical path between the graphic processor modules is established through the connected AWGR; and the optical switching is simultaneously performed through multiple AWGRs. The application is suitable for large-scale networks and has simple network architecture, can reduce the cost and ensure low latency.

Description

Technical Field

[0001] This invention relates to the field of optical switching network technology, and in particular to an optical switching network architecture, scheduling method, medium, and device based on AWGR. Background Technology

[0002] Graphics Processing Units (GPUs) possess computational advantages due to their unique parallel processing architecture, and are therefore widely used in intelligent computing centers for large-scale networks, such as for training Large Language Models (LLMs) and industrial automation control. In large-scale networks, different GPUs need to communicate with each other. As network scale continues to expand, the demand for communication traffic also increases. To address the shortcomings of electrical switches, such as low throughput and high energy consumption, optical switches have emerged as a replacement for electrical switches.

[0003] Common optical switching methods in intelligent computing centers include those based on Micro-Electro-Mechanical Systems (MEMS) and those based on Wavelength Selective Switches (WSS). MEMS-based switching establishes optical paths at the physical layer using MEMS, achieving transparency of transmission rates and dynamically changing the topology to adapt to different scenarios; however, this method suffers from high latency during dynamic reconfiguration, failing to meet the extremely low latency requirements of intelligent computing centers. WSS-based switching expands ports through cascading, enabling the selection and independent routing of specific wavelengths from multiple optical signals; however, WSS equipment is expensive, and the number of WSS ports is strictly limited due to technical constraints, making its application in large-scale networks difficult.

[0004] To meet the low-latency application requirements in large-scale intelligent computing centers, existing technologies have developed optical switching methods based on Arrayed Waveguide Grating Routers (AWGRs). AWGRs are passive devices that can route optical signals solely based on the wavelength of the input ports. By using tunable wavelength converters (TWCMs) to change the wavelength of the optical signals, extremely low-latency switching of optical signals is possible. Furthermore, AWGRs have no strict limitation on the number of ports, making them suitable for large-scale networks.

[0005] However, the AWGR-based optical switching method also has drawbacks, mainly including:

[0006] 1. Optical switching involves frequent photoelectric signal conversion, requiring a large number of optical modules. Current technologies typically use distributed optical modules, each of which needs an independent light source to generate optical signals, resulting in high costs. Furthermore, distributed optical modules can only generate optical signals of specific wavelengths, necessitating the use of TWCM, which further increases costs.

[0007] 2. This method is mostly implemented by cascading multiple AWGRs. Although this method can be applied to large-scale networks, the network complexity is high and it will affect network latency. Summary of the Invention

[0008] Therefore, the technical problem to be solved by the present invention is to overcome the shortcomings of the prior art and provide an optical switching network architecture, scheduling method, medium and equipment based on AWGR, which is suitable for large-scale networks and has a simple network architecture, and can reduce costs and ensure low latency.

[0009] To address the aforementioned technical problems, this invention provides an optical switching network architecture based on AWGR, comprising a graphics processor module, an optical module, an arrayed waveguide grating router, and a centralized central light source. The centralized central light source provides light to the optical modules and coordinates wavelength resources. The graphics processor modules are interconnected through the optical modules and the arrayed waveguide grating router to form an optical switching network architecture.

[0010] When the central light source provides light source for the optical module and coordinates the wavelength resources, it calculates the light source frequency of the communication wavelength by combining the base frequency of the array waveguide grating router, the device frequency offset constant, the frequency interval of adjacent communication channels, and the temperature drift. The allocated communication wavelength is obtained based on the light source frequency of the communication wavelength.

[0011] Furthermore, the system includes multiple optical modules, with each optical module connected to a graphics processor module, and each pair of input ports and output ports at the same location of all the optical modules connected to each pair of input ports and output ports of an arrayed waveguide grating router.

[0012] Furthermore, the number of optical modules is the same as the number of graphics processor modules, the number of input ports and output ports of the optical modules is the same as the number of arrayed waveguide grating routers, and the number of optical modules is the same as the number of input ports and output ports of the arrayed waveguide grating routers.

[0013] Furthermore, each pair of input ports and output ports at the same location of all the optical modules is connected to each pair of input ports and output ports of the arrayed waveguide grating router, specifically: the j-th pair of input ports and output ports of the i-th optical module is connected to the i-th pair of input ports and output ports of the j-th arrayed waveguide grating router, i=1,2,…,M, where M represents the number of optical modules, j=1,2,…,m, where m represents the number of pairs of input ports and output ports of the optical module.

[0014] Furthermore, the graphics processor modules are interconnected via the optical module and the arrayed waveguide grating router, specifically:

[0015] The optical module converts the electrical signal output by the graphics processor module into an optical signal and transmits it to the arrayed waveguide grating router. The arrayed waveguide grating router routes optical signals of different wavelengths to the corresponding output ports according to the different wavelengths of the received optical signals and transmits them to the optical module connected to the output port. The optical module converts the optical signal output by the arrayed waveguide grating router into an electrical signal and transmits it to the graphics processor module connected to the optical module, thereby realizing the switching of optical paths.

[0016] Furthermore, it also includes a network controller, which is connected to the central lumped light source. The network controller controls the central lumped light source to provide light sources for the optical modules and coordinates the wavelength resources by combining communication requests between the graphics processor modules.

[0017] This invention also provides a scheduling method for an AWGR-based optical switching network. When using the AWGR-based optical switching network architecture, the network scheduling method includes:

[0018] The central light source generates a light source with a wavelength of the communication wavelength and provides it to the optical module connected to the graphics processor module that issued the communication request. The communication optical path between the graphics processor modules is established through the arrayed waveguide grating router connected to the graphics processor module that issued the communication request, combined with time division multiplexing strategy, traffic awareness strategy and load balancing strategy.

[0019] The optical module divides the data to be transmitted by the graphics processor module connected to the optical module into multiple parts, and completes the data transmission by simultaneously performing optical switching through multiple arrayed waveguide grating routers connected to the optical module.

[0020] Furthermore, the method for calculating the communication wavelength is as follows:

[0021] ,

[0022] In the formula, Let represent the communication wavelength at the p-th arrayed waveguide grating router where the graphics processing unit (GPU) acting as the source node communicates with the GPU acting as the destination node, where c represents the speed of light and k1 is a preset coefficient. This represents the light source frequency used for communication between the graphics processing module (GPU) acting as the source node and the GPU acting as the destination node at the p-th arrayed waveguide grating router.

[0023] The calculation method is as follows:

[0024] ,

[0025] In the formula, This indicates the fundamental frequency of the arrayed waveguide grating router. This indicates the input port number of the p-th arrayed waveguide grating router, where the graphics processor module acts as the source node. This indicates the output port number of the p-th arrayed waveguide grating router, which is the destination node of the graphics processor module. Let be the device frequency offset constant of the arrayed waveguide grating router, mod be the modulo operation, and N represent the port size of the arrayed waveguide grating router. This represents the frequency spacing between adjacent communication channels of the arrayed waveguide grating router, and T is the current operating temperature of the arrayed waveguide grating router. Temperature calibration for devices in an arrayed waveguide grating router. The drift coefficients are the frequency and temperature drift coefficients. This indicates the frequency fine-tuning parameter.

[0026] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the aforementioned AWGR-based optical switching network scheduling method.

[0027] The present invention also provides an AWGR-based optical switching network scheduling device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the AWGR-based optical switching network scheduling method.

[0028] Compared with the prior art, the above-described technical solution of the present invention has the following advantages:

[0029] This invention constructs an optical switching network by connecting graphics processing unit (GPU) modules using optical modules and arrayed waveguide grating (AWGR) routers. Based on this, a lumped central light source provides light to the optical modules and coordinates wavelength resources. The lumped central light source calculates the communication wavelength by combining the AWGR's fundamental frequency, device frequency offset constant, frequency spacing of adjacent communication channels, and temperature drift, enabling rapid optical switching between GPU modules. The overall optical network architecture is simple, effectively reducing the complexity of large-scale networks, thereby achieving high-speed signal transmission and ensuring low latency. Using a lumped central light source for global coordination and scheduling effectively reduces light source costs compared to distributed optical modules; furthermore, it eliminates the need for high-cost equipment such as TWCMs, further reducing costs. Attached Figure Description

[0030] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings, wherein:

[0031] Figure 1 This is an example diagram of an optical switching network architecture based on AWGR in a preferred embodiment of the present invention.

[0032] Figure 2 This is an example diagram illustrating the routing principle of AWGR.

[0033] Figure 3 This is a diagram showing the switching delay data of the present invention in a simulation experiment.

[0034] Figure 4 This is a diagram showing the switching delay data of the MEMS-based optical switching method in the simulation experiment.

[0035] Figure 5 This is a comparison chart of the average switching latency of all GPUs in the simulation experiment of the present invention and the MEMS-based optical switching method. Detailed Implementation

[0036] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention.

[0037] This invention discloses an optical switching network architecture based on AWGR (Aperture Radial Waveguide Grating), including a graphics processor module, optical modules, an arrayed waveguide grating router, and a centralized central light source. The centralized central light source provides light to the optical modules and coordinates wavelength resources. The graphics processor modules are interconnected through the optical modules and the arrayed waveguide grating router to form the optical switching network architecture. Unlike traditional distributed optical module setups, the network architecture in this invention utilizes a centralized central light source to provide light to the optical modules, and on this basis, performs unified and coordinated scheduling of wavelength resources. A single central light source can effectively reduce costs.

[0038] An AWGR-based optical switching network architecture comprises multiple optical modules. Each optical module is connected to a graphics processor module. Each pair of input and output ports at the same location on all optical modules is connected to each pair of input and output ports of an arrayed waveguide grating router. The number of optical modules can be the same as or different from the number of graphics processors, as long as each graphics processor has one optical module to convert electrical signals into optical signals. The number of input and output port pairs of the optical modules can be the same as or different from the number of input and output port pairs of the arrayed waveguide grating routers, as long as each pair of input and output ports at the same location on all optical modules is connected to each pair of input and output ports of an arrayed waveguide grating router.

[0039] In this embodiment, Figure 1 Taking the network architecture shown as an example, the number of optical modules is the same as the number of graphics processor modules, and the number of input and output port pairs of the optical modules is the same as the number of arrayed waveguide grating routers. The graphics processor modules can be graphics processors or servers equipped with graphics processors, and the optical modules can be DR4s with 4 pairs of input and output ports. Figure 1 As shown, there are M graphics processing units (denoted as GPU1, GPU2, ..., GPU3) i GPU M-1 GPU M ) and M DR4s (denoted as DR41, DR42, ..., DR4) i , ..., DR4 M-1 DR4 M For example, GPU i Represents the i-th graphics processor, DR4 i Indicates the i-th DR4, GPU i and DR4 i Connected. Taking an AWGR with 64 pairs of input and output ports as an example, M=64 and the number of AWGRs is 4.

[0040] In an AWGR-based optical switching network architecture, the j-th pair of input and output ports of the i-th optical module is connected to the i-th pair of input and output ports of the j-th arrayed waveguide grating router. i = 1, 2, ..., M, where M represents the number of optical modules, and j = 1, 2, ..., m, where m represents the number of input and output port pairs of the optical module. In this embodiment, m = 4. Each pair of input and output ports at the same location on all DR4s is connected to each pair of input and output ports of an arrayed waveguide grating router, and these connections can be sequential, i.e., as shown below. Figure 1 As shown, DR4 i The j-th pair of input ports and output ports are connected to the i-th pair of input ports and output ports of the j-th arrayed waveguide grating router. Figure 1 The arrows in the diagram indicate the direction of optical signal transmission (j=1,2,3,4); they can also be connected out of order, as long as each pair of input and output ports at the same location on all DR4s is connected to all input and output ports of any arrayed waveguide grating router. The total transmission rate of each DR4 is T Gbps, then the transmission rate of each pair of input and output ports is (T / m) Gbps.

[0041] In an AWGR-based optical switching network architecture, the optical module converts the electrical signals output by the graphics processor module into optical signals (enabling them to be transmitted in the optical path) and transmits them to the arrayed waveguide grating router. The arrayed waveguide grating router routes the optical signals of different wavelengths to the corresponding output ports according to the different wavelengths of the received optical signals and transmits them to the optical modules connected to the output ports. The optical module converts the optical signals output by the arrayed waveguide grating router into electrical signals and transmits them to the graphics processor module connected to the current optical module (after which the graphics processor module can perform further processing), thereby realizing the switching of the optical path. Figure 2 The routing diagram for a 64×64 AWGR includes 64 input ports (denoted as input port 1, input port 2, ..., input port 64) and 64 output ports (denoted as output port 1, output port 2, ..., output port 64). Figure 2 The different line colors represent different wavelengths. When light signals of different wavelengths are input to input port 1, they will be routed to different output ports, enabling the display of 64 different wavelength light signals (denoted as λ1, λ2, ..., λ3). 64 The routing is as follows: there is a one-to-one correspondence between wavelength and output port, but each input port does not correspond to a wavelength, which is determined by the characteristics of AWGR cyclic wavelength routing.

[0042] The AWGR-based optical switching network architecture may also include a network controller, which is connected to the central light source. The network controller coordinates and schedules wavelength resources by combining communication requests between graphics processor modules and controlling the central light source.

[0043] This invention also discloses a scheduling method for optical switching networks based on AWGR. When using an AWGR-based optical switching network architecture, the routing algorithm strategies include time-division multiplexing, traffic-awareness, and load balancing. Unlike electrical switches, AWGR optical switches cannot buffer transmitted data; they are essentially direct physical links. Furthermore, within the same time slot, an AWGR optical path can only handle one service; otherwise, the link will be blocked. Therefore, a time-division multiplexing strategy is needed, setting different time slots to process each service sequentially. Since traffic in intelligent computing center scenarios is generally non-uniform, situations often arise where communication traffic is high between some nodes and low between others. Ports between high-traffic node pairs may remain occupied for extended periods. Therefore, a traffic-aware strategy is used to determine if the traffic between node pairs exceeds a certain threshold; if the traffic is too high, it is determined that traffic forwarding is necessary. Traffic load in intelligent computing center scenarios is generally unbalanced; therefore, a load balancing strategy is used to forward traffic between high-traffic-load node pairs through low-traffic-load nodes, making the traffic load across different nodes more balanced.

[0044] The network scheduling process includes the following steps:

[0045] S1: The graphics processing module generates the data to be transmitted and sends a communication request to the network controller (specifically, the SDN controller in this embodiment). The communication request is denoted as R(A,B), where A represents the graphics processing module as the source node and B represents the graphics processing module as the destination node.

[0046] S2: The SDN controller calculates the communication wavelength based on the communication request. The SDN controller controls the central light source to generate a light source with the communication wavelength and provides it to the optical modules connected to A and B. Through the arrayed waveguide grating router connected to the optical modules connected to A and B, a communication optical path between A and B is established by combining time division multiplexing strategy, traffic awareness strategy, and load balancing strategy.

[0047] S2-1: Number all graphics processor modules. The numbering method can be designed as needed based on the actual network size and structure. Alternatively, the same method as the AWGR port numbering can be used. The numbering is a positive integer, and the graphics processor module number corresponds to the AWGR port number.

[0048] S2-2: For each request (s, d) in the current time slot, where s represents the number of the graphics processing unit (GPU) module acting as the source node and d represents the number of the GPU module acting as the destination node, obtain the number of the input port corresponding to the AWGR connected to the GPU module (s) acting as the source node (denoted as...). The output port number corresponding to the AWGR connected to the graphics processor module (d) as the destination node (denoted as ) and the output port number corresponding to the AWGR (denoted as ). ).

[0049] S2-3: Determine the AND of all AWGR values. , Are the ports in the same location idle? If yes, proceed to S2-4; otherwise, proceed to S2-5.

[0050] S2-4: At this time, the link is idle. The central light source calculates the communication wavelength and uses the time division multiplexing strategy to process the communication service.

[0051] If two GPU nodes need to communicate, the central light source needs to provide light to the optical modules. The wavelength of the light source (i.e., the communication wavelength) is calculated as follows:

[0052] First, the frequency of the light source for calculating the communication wavelength is:

[0053] ,

[0054] In the formula, The light source frequency, in THz, represents the communication frequency between the graphics processing module (s) as the source node and the graphics processing module (d) as the destination node during the p-th AWGR communication. This indicates the base frequency of the AWGR, which is the lowest channel frequency and is related to the actual specifications of the AWGR. This indicates the input port number of the graphics processor module (s) acting as the source node in the p-th AWGR. This indicates the output port number of the graphics processor module (d) that serves as the destination node in the p-th AWGR; is the device frequency offset constant of AWGR, which is related to the actual specifications of AWGR; mod is the modulo operation, and N represents the port size of AWGR. In this embodiment, N=64. This indicates the frequency spacing between adjacent communication channels of the AWGR, which is related to the actual specifications of the AWGR; in this embodiment, it is set to 100GHz. Because the AWGR is sensitive to temperature changes, temperature drift adaptive compensation is set. For temperature drift adaptive compensation, T is the current operating temperature of the arrayed waveguide grating router, and is the device calibration temperature of the arrayed waveguide grating router. The drift coefficients for frequency and temperature are expressed in GHz. This represents the frequency fine-tuning parameter, used to improve accuracy; its value is adjusted according to the actual situation.

[0055] The specific communication wavelength is calculated based on its frequency:

[0056] ,

[0057] In the formula, This represents the communication wavelength between the graphics processing unit (S) as the source node and the graphics processing unit (D) as the destination node during the p-th AWGR communication. Represents the speed of light. As a preset coefficient, in this embodiment . The unit is nanometer (nm). After obtaining specific wavelength data from the central light source, it can provide a light source for the optical module, enabling communication between the source node and the destination node.

[0058] S2-5: At this point, the link is blocked. Use the traffic awareness strategy to determine whether the traffic between the graphics processing unit (GPU) module as the source node and the GPU module as the destination node is too high. If yes, execute S2-6; otherwise, execute S2-7.

[0059] S2-6: If the traffic between the source and destination graphics processing unit (GPU) modules is too high, a load balancing strategy is used to find a relay node for forwarding. The relay node must meet the following conditions: the path from the source GPU to the relay node (denoted as path s→M, where M represents the relay node) is available in the current time slot; the path from the relay node to the destination GPU (path M→d) is available in the next time slot; and the relay node has a low traffic load. If a relay node meeting these three conditions is found, proceed to S2-8; otherwise, proceed to S2-9.

[0060] S2-7: At this point, the link is blocked, and the request enters a waiting state until a time slot becomes available.

[0061] S2-8: Establish a path (path s→M) from the graphics processor module as the source node to the relay node, and add the path (path M→d) from the relay node to the graphics processor module as the destination node to the next time slot scheduling for traffic forwarding.

[0062] S2-9: No available relay nodes, traffic forwarding request is returned, proceed to S2-7.

[0063] S2-10: Communication request processing completed, update the status table of each port of AWGR.

[0064] S3: The optical module divides the data to be transmitted by the graphics processor module into multiple parts, and completes the data transmission by simultaneously performing optical switching through multiple arrayed waveguide grating routers.

[0065] The present invention also discloses a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements an AWGR-based optical switching network scheduling method.

[0066] The present invention also discloses an optical switching network scheduling device based on AWGR, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements an optical switching network scheduling method based on AWGR.

[0067] Compared with the prior art, the advantages of the present invention are:

[0068] 1. An optical switching network is constructed by connecting graphics processing unit (GPU) modules using optical modules and arrayed waveguide grating (AWGR) routers. Based on this, a centralized light source provides illumination to the optical modules and coordinates wavelength resources. The centralized light source calculates the communication wavelength by combining the AWGR's fundamental frequency, device frequency offset constant, frequency spacing of adjacent communication channels, and temperature drift, thus enabling fast optical switching between GPU modules. Compared to an optical switching architecture cascaded with multiple AWGR routers, using multiple parallel-configured large-port AWGR routers instead of a cascaded structure not only expands the network scale and reduces network complexity but also reduces the likelihood of multiple wavelengths competing for the same link, thereby ensuring high-speed signal transmission and achieving low latency.

[0069] 2. Using a centralized light source for global coordination and scheduling can directly provide the required wavelength light source to the optical modules. Using a centralized light source instead of traditional distributed optical modules can effectively reduce the cost of the light source. At the same time, using an arrayed waveguide grating router as the switching core and combining optical modules and a centralized light source for optical switching, the arrayed waveguide grating router has low power consumption and does not require the use of high-cost equipment such as TWCM, further reducing costs.

[0070] 3. Arrayed waveguide grating routers perform optical switching based on the principle of light interference, eliminating the need for switching time and enabling the switching of optical signals with extremely low latency. Furthermore, the optical signal is divided into multiple parts and transmitted in parallel through multiple parallel arrayed waveguide grating routers. Combined with the low optical switching latency, high-speed signal transmission can be achieved, ensuring the low latency and capacity requirements of the overall optical switching network.

[0071] 4. The network controller controls the entire optical switching network architecture and, in conjunction with the communication requests between graphics processor modules, performs global scheduling of wavelength resources. The routing algorithm is simpler and more efficient, further reducing optical switching latency.

[0072] To further demonstrate the beneficial effects of the present invention, using Figure 1 The network architecture was simulated using a network architecture. The number of GPUs and DR4s was 64 each, and the number of 64×64 AWGRs was 4. The network architecture was simulated and tested in a uniform flow mode. Network parameters were set as follows: wavelength allocation delay of the lumped central light source was 300-400µs, photoelectric conversion delay of DR4 was 250-350ns, and switching delay of AWGR was 10ns-15ns.

[0073] This invention is compared with the switching latency of MEMS-based optical switching methods. Switching latency refers to the time interval from when a GPU sends data, through the network architecture, to when the destination GPU begins receiving data.

[0074] In uniform flow mode, the switching delay data of this invention is as follows: Figure 3 As shown, Figure 3 The image shows the switching latency data for a subset of 64 GPUs. Also under uniform throughput mode, the switching cores were replaced with MEMS instead of AWGR; the switching latency data is as follows... Figure 4 As shown, Figure 4 The presentation also shows the switching latency data for a subset of GPUs (64 in total); although the latency caused by wavelength allocation is eliminated, the switching latency is still around 10ms. The average switching latency of the 64 GPUs in this invention is 0.31ms, while the average switching latency of the 64 GPUs using the MEMS-based optical switching method is 10.16ms. A comparison of the average switching latency of all GPUs in both methods is provided. Figure 5 As shown.

[0075] The switching delay of this invention mainly consists of the photoelectric conversion delay of DR4, the wavelength allocation delay of the lumped central light source, and the switching delay of AWGR, with the overall switching delay in the microsecond range. The switching delay of the MEMS-based optical switching method mainly consists of the photoelectric conversion delay of DR4 and the switching delay of MEMS, with the MEMS switching delay being much greater than the photoelectric conversion delay, resulting in an overall switching delay in the millisecond range. Simulation test data also shows that the switching delay of this invention is significantly less than that of the MEMS-based switching method, with a reduction of over 90%, demonstrating the advantages of this invention in terms of switching delay.

[0076] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0077] This application is described with reference to the flow of methods, apparatus (systems), and computer program products according to embodiments of this application. It should be understood that one or more flows of the flow can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in one or more flows of the flow.

[0078] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means that implement the functions specified in one or more processes.

[0079] These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, such that the instructions, which execute on the computer or other programmable apparatus, provide steps for implementing a function specified in one or more processes.

[0080] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. An optical switching network architecture based on AWGR, characterized in that: It includes a graphics processing unit (GPU) module, an optical module, an arrayed waveguide grating router, and a centralized light source. The centralized light source provides light to the optical modules and coordinates wavelength resources. The GPU modules are connected to each other through the optical modules and the arrayed waveguide grating router to form an optical switching network architecture. When the central light source provides light source for the optical module and coordinates the wavelength resources, it calculates the light source frequency of the communication wavelength by combining the base frequency of the array waveguide grating router, the device frequency offset constant, the frequency interval of adjacent communication channels, and the temperature drift, and obtains the allocated communication wavelength based on the light source frequency of the communication wavelength. The method for calculating the communication wavelength is as follows: ， In the formula, Let represent the communication wavelength at the p-th arrayed waveguide grating router where the graphics processing unit (GPU) acting as the source node communicates with the GPU acting as the destination node, where c represents the speed of light and k1 is a preset coefficient. This represents the light source frequency used for communication between the graphics processing module (GPU) acting as the source node and the GPU acting as the destination node at the p-th arrayed waveguide grating router. The calculation method is as follows: ， In the formula, This indicates the fundamental frequency of the arrayed waveguide grating router. This indicates the input port number of the graphics processing unit module acting as the source node in the p-th arrayed waveguide grating router. This indicates the output port number of the graphics processing unit module acting as the destination node in the p-th arrayed waveguide grating router. Here, is the device frequency offset constant of the arrayed waveguide grating router, mod is the modulo operation, and N represents the port size of the arrayed waveguide grating router. This represents the frequency spacing between adjacent communication channels of the arrayed waveguide grating router, and T is the current operating temperature of the arrayed waveguide grating router. Temperature calibration for devices in an arrayed waveguide grating router. The drift coefficients are the frequency and temperature drift coefficients. This indicates the frequency fine-tuning parameter.

2. The optical switching network architecture based on AWGR according to claim 1, characterized in that: It includes multiple optical modules, one of which is connected to a graphics processor module, and each pair of input ports and output ports at the same location of all the optical modules is connected to each pair of input ports and output ports of an arrayed waveguide grating router.

3. The optical switching network architecture based on AWGR according to claim 2, characterized in that: The number of optical modules is the same as the number of graphics processor modules. The number of input ports and output ports of the optical modules is the same as the number of arrayed waveguide grating routers.

4. The optical switching network architecture based on AWGR according to claim 2, characterized in that: Each pair of input ports and output ports at the same location of all the optical modules is connected to each pair of input ports and output ports of an arrayed waveguide grating router. Specifically, the j-th pair of input ports and output ports of the i-th optical module is connected to the i-th pair of input ports and output ports of the j-th arrayed waveguide grating router, where i=1,2,…,M, M represents the number of optical modules, j=1,2,…,m, and m represents the number of pairs of input ports and output ports of the optical module.

5. The optical switching network architecture based on AWGR according to claim 2, characterized in that: The graphics processor modules are connected to each other via the optical module and the arrayed waveguide grating router, specifically: The optical module converts the electrical signal output by the graphics processor module into an optical signal and transmits it to the arrayed waveguide grating router. The arrayed waveguide grating router routes optical signals of different wavelengths to the corresponding output ports according to the different wavelengths of the received optical signals and transmits them to the optical module connected to the output port. The optical module converts the optical signal output by the arrayed waveguide grating router into an electrical signal and transmits it to the graphics processor module connected to the optical module, thereby realizing the switching of optical paths.

6. The optical switching network architecture based on AWGR according to any one of claims 1-5, characterized in that: It also includes a network controller, which is connected to the central light source. The network controller controls the central light source to provide light sources for the optical modules and coordinates the wavelength resources by combining the communication requests between the graphics processor modules.

7. A scheduling method for optical switching networks based on AWGR, characterized in that: When using the AWGR-based optical switching network architecture as described in any one of claims 1-6, the network scheduling method includes: The central light source generates a light source with a wavelength of the communication wavelength and provides it to the optical module connected to the graphics processor module that issued the communication request. The communication optical path between the graphics processor modules is established through the arrayed waveguide grating router connected to the graphics processor module that issued the communication request, combined with time division multiplexing strategy, traffic awareness strategy and load balancing strategy. The optical module divides the data to be transmitted by the graphics processor module connected to the optical module into multiple parts, and completes the data transmission by simultaneously performing optical switching through multiple arrayed waveguide grating routers connected to the optical module.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the AWGR-based optical switching network scheduling method as described in claim 7.

9. A scheduling device for an optical switching network based on AWGR, characterized in that: It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the AWGR-based optical switching network scheduling method as described in claim 7.

Images