A spatio-temporal graph convolution-based optical algorithm fusion network dynamic scheduling method and related device

Abstract

The application relates to the technical field of network resource scheduling, in particular to a light-computing fusion network dynamic scheduling method based on a space-time graph convolution and related devices. The method comprises the following steps: constructing a unified resource graph of a light-computing fusion network; performing space-time feature extraction and fusion on the unified resource graph through a pre-trained space-time graph convolution network model to obtain a predicted state of each node in a next time window; establishing a multi-dimensional joint optimization objective function; determining a probability score of each node; performing pruning according to the probability score of each node to obtain a pruned candidate subgraph; adopting a heuristic search algorithm to obtain an optimal scheduling path according to the multi-dimensional joint optimization objective function; and generating light layer and electric layer collaborative scheduling instructions according to the optimal scheduling path. The application can realize intelligent, efficient and real-time dynamic scheduling of the light-computing fusion network.

Description

Technical Field

[0001] This application relates to the field of network resource scheduling technology, and in particular to a dynamic scheduling method and related apparatus for optical computing fusion networks based on spatiotemporal graph convolution. Background Technology

[0002] With the rapid development of data-intensive applications such as large-scale AI model training and distributed computing, the demand for computing power is growing exponentially. The traditional independent management model of optical networks and computing resources can no longer meet the stringent requirements of emerging services for low latency, high bandwidth, and high energy efficiency. Against this backdrop, the Optical-Computing Integrated Network (OCI) has emerged, aiming to achieve joint optimization of computing tasks and network transmission through optoelectronic collaboration and computing-network integration.

[0003] However, existing dynamic scheduling technologies for optical-computing converged networks still have the following shortcomings: passive and lagging scheduling modes, insufficient utilization of spatiotemporal correlation, disconnect between optoelectronic collaborative scheduling, and separation of computing and network resource management. Therefore, there is an urgent need for a new technology solution that can realize intelligent, efficient, and real-time dynamic scheduling of optical-computing converged networks. Summary of the Invention

[0004] To address the aforementioned issues, this application provides a dynamic scheduling method and related apparatus for optical computing fusion networks based on spatiotemporal graph convolution, which can solve at least one of the above-mentioned technical problems.

[0005] According to one aspect of the embodiments of this application, a dynamic scheduling method for optical computing fusion networks based on spatiotemporal graph convolution is proposed, the method comprising: A unified resource map of the optical computing convergence network is constructed, which is obtained by collecting multi-dimensional state data of each node of the optical computing convergence network in multiple historical time windows. The unified resource graph is subjected to spatiotemporal feature extraction and fusion by a pre-trained spatiotemporal graph convolutional network model to obtain the predicted state of each node in the next time window. The predicted state includes predicted latency distribution, predicted traffic load and predicted resource demand. Establish a multidimensional joint optimization objective function related to the predicted latency distribution, the predicted traffic load, and the predicted resource demand; Determine the probability score of each node, the probability score being used to characterize the confidence of a node as an intermediate node in the optimal scheduling path; The unified resource graph is pruned based on the probability scores of each node to obtain a pruned candidate subgraph. The optimal scheduling path is obtained by solving the multidimensional joint optimization objective function on the candidate subgraph using a heuristic search algorithm. Based on the optimal scheduling path, a collaborative scheduling instruction for the optical and electrical layers of the optical computing fusion network is generated to perform collaborative scheduling of optoelectronic resources.

[0006] In the above scheme, the construction of a unified resource map for the optical computing fusion network includes: Obtain the topology information of the optical computing fusion network, the topology information including the location of each node and the physical connection relationship between each node; Construct the original graph structure based on the position of each node and the physical connection relationship between each node; Within each historical time window, multidimensional status data of each node is collected. The multidimensional status data includes optical path status parameters, computing resource parameters, and service traffic data within the historical time window. The multidimensional state data of each node collected in each of the historical time windows are fused with the original graph structure to obtain the unified resource graph.

[0007] In the above scheme, the spatiotemporal graph convolutional network model includes a spatial feature extraction module and a temporal feature extraction module. The process of extracting and fusing spatiotemporal features of the unified resource graph using the pre-trained spatiotemporal graph convolutional network model to obtain the predicted state of each node within the next time window includes: The spatial feature extraction module extracts the adjacency matrix composed of each node in the unified resource graph and the node feature matrix composed of the multidimensional state data of each node. For each node, the multidimensional state data of the node itself and the multidimensional state data of each neighboring node adjacent to the node are aggregated to obtain the aggregated spatial features of the node. The time feature extraction module extracts spatial features from the multidimensional state data collected in each historical time window along the time dimension to obtain the time features of each node used to characterize the dynamic law of the optical computing fusion network's evolution over time. For each node, the aggregated spatial features and the temporal features of the node are fused to obtain the predicted state of the node in the next time window.

[0008] In the above scheme, a multi-dimensional joint optimization objective function related to the predicted latency distribution, the predicted traffic load, and the predicted resource demand is established, including: The end-to-end transmission delay value is determined based on the predicted delay distribution, and a delay weight corresponding to the end-to-end transmission delay value is constructed. The bandwidth resource utilization efficiency value is determined based on the predicted traffic load, and a bandwidth weight corresponding to the bandwidth resource utilization efficiency value is constructed. The computing power-task matching degree is determined based on the predicted resource demand, and the computing power weight corresponding to the computing power-task matching degree is constructed. Determine the total energy consumption of all nodes and links on the target path, and construct the energy consumption weight corresponding to the total energy consumption. The target path is a path composed of any number of connected nodes. The multidimensional joint optimization objective function is constructed based on the end-to-end transmission delay value, the delay weight, the bandwidth resource utilization efficiency value, the bandwidth weight, the computing power-task matching degree, the computing power weight, the total energy consumption, and the energy consumption weight.

[0009] In the above scheme, the step of pruning the unified resource graph based on the probability scores of each node to obtain the pruned candidate subgraph includes: For each node, if the probability score of the node is lower than a preset probability threshold, the node is removed from the unified resource graph. Each node that was not removed is taken as a retained node, and the candidate subgraph after pruning is determined based on each of the retained nodes. Wherein, the difference between the number of retained nodes and the preset node retention threshold is less than the preset difference, and the preset node retention threshold is the product of the probability threshold and the total number of nodes in the unified resource graph.

[0010] In the above scheme, the step of employing a heuristic search algorithm to obtain the optimal scheduling path on the candidate subgraph based on the multidimensional joint optimization objective function includes: For each business requirement, determine the source node and destination node in the business requirement, and generate multiple candidate paths based on the source node and the destination node; Based on the end-to-end transmission delay value, the delay weight, the bandwidth resource utilization efficiency value, the bandwidth weight, the computing power-task matching degree, the computing power weight, the total energy consumption, and the energy consumption weight of the multidimensional joint optimization objective function, a unified constraint condition is constructed for each of the candidate paths. Construct the cost constraints for the candidate paths in the heuristic search algorithm; The heuristic search algorithm is used to search for the optimal scheduling path in each candidate path according to the unified constraint and the cost constraint. The optimal scheduling path satisfies the unified constraint and the cost constraint.

[0011] In the above scheme, generating the optical and electrical layer collaborative scheduling instructions for the optical-computing fusion network based on the optimal scheduling path includes: If the service requirement belongs to the service type of high-capacity long-distance transmission, an optical layer scheduling instruction for the end-to-end optical channel is generated to transmit the service requirement through the end-to-end optical channel in the optical layer according to the optical layer scheduling instruction. The service type of high-capacity long-distance transmission is used to indicate that the physical distance between the source node and the destination node in the service requirement is greater than a preset distance threshold or the bandwidth requirement value in the service requirement is greater than a first preset bandwidth requirement threshold. If the service requirement belongs to the service type that requires flexible scheduling and fine-grained switching, an electrical layer scheduling instruction for the electrical packet forwarding path is generated so that data can be transmitted through the electrical packet forwarding path of the electrical layer according to the electrical layer scheduling instruction. The service type that requires flexible scheduling and fine-grained switching is used to indicate that the bandwidth requirement value in the service requirement is less than the second preset bandwidth requirement threshold. If the business requirement belongs to the business type of mixed business scenario, a collaborative scheduling instruction for optical channel and electrical packet forwarding is generated. The delay weight, bandwidth weight, computing power weight and energy consumption weight in the multi-dimensional joint optimization objective function are dynamically adjusted according to the collaborative scheduling instruction. The optical layer carrying ratio and the electrical layer carrying ratio are determined according to the dynamically adjusted delay weight, bandwidth weight, computing power weight and energy consumption weight, and data transmission is performed according to the optical layer carrying ratio and the electrical layer carrying ratio. The business type belonging to the mixed business scenario is used to characterize the business requirement, which includes at least two types of business requirements, including latency-sensitive, computation-intensive, high-traffic, and balanced.

[0012] According to one aspect of the embodiments of this application, a dynamic scheduling device for optical computing fusion networks based on spatiotemporal graph convolution is proposed, the device comprising: The construction unit is used to construct a unified resource map of the optical computing convergence network. The unified resource map is obtained by collecting multi-dimensional state data of each node of the optical computing convergence network in multiple historical time windows. The feature extraction unit is used to extract and fuse spatiotemporal features of the unified resource graph through a pre-trained spatiotemporal graph convolutional network model to obtain the predicted state of each node in the next time window. The predicted state includes predicted latency distribution, predicted traffic load, and predicted resource demand. The function establishment unit is used to establish a multi-dimensional joint optimization objective function related to the predicted latency distribution, the predicted traffic load, and the predicted resource demand. A determining unit is used to determine the probability score of each of the nodes, the probability score being used to characterize the confidence of a node as an intermediate node in the optimal scheduling path; The pruning unit is used to prune the unified resource graph according to the probability scores of each node to obtain the pruned candidate subgraph. The solution unit is used to obtain the optimal scheduling path on the candidate subgraph by employing a heuristic search algorithm based on the multidimensional joint optimization objective function; The scheduling unit is used to generate a coordinated scheduling instruction between the optical and electrical layers of the optical computing fusion network based on the optimal scheduling path, so as to perform coordinated scheduling of optoelectronic resources according to the coordinated scheduling instruction between the optical and electrical layers.

[0013] According to one aspect of the embodiments of this application, an electronic device is provided, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the dynamic scheduling method for optical computing fusion networks based on spatiotemporal graph convolution as described above.

[0014] According to one aspect of the embodiments of this application, a computer program product is provided, the computer program product including a computer program, the computer program being read and executed by a processor of an electronic device, causing the electronic device to perform the dynamic scheduling method for optical computing fusion networks based on spatiotemporal graph convolution as described above.

[0015] The beneficial effects of this application are as follows: This application uses a pre-trained spatiotemporal graph convolutional network model to extract and fuse spatiotemporal features of a unified resource graph, obtaining the predicted state of each node in the next time window, thereby constructing a multi-dimensional joint optimization objective function. Simultaneously, the same resource graph is pruned using the probability scores of each node. A heuristic search algorithm combined with the multi-dimensional joint optimization objective function is used to solve for the optimal scheduling path on the candidate subgraph. The optimal scheduling path generates optical and electrical layer collaborative scheduling instructions for the optical-computing fusion network, enabling collaborative scheduling of optoelectronic resources based on these instructions. This application enables intelligent, efficient, and real-time dynamic scheduling of optical-computing fusion networks. Attached Figure Description

[0016] Figure 1 This is a system architecture diagram of the optical computing fusion network dynamic scheduling method based on spatiotemporal graph convolution provided in the embodiments of this application. Figure 2 A flowchart illustrating the dynamic scheduling method for optical computing fusion networks based on spatiotemporal graph convolution provided in this application embodiment; Figure 3 An overall logical diagram of dynamic scheduling of optical computing fusion network based on spatiotemporal graph convolution provided in this application embodiment; Figure 4 A logic diagram for generating optical and electrical layer collaborative scheduling instructions provided in this application embodiment; Figure 5 The overall logic timing diagram provided for the embodiments of this application; Figure 6 A block diagram of a dynamic scheduling device for optical computing fusion network based on spatiotemporal graph convolution provided in an embodiment of this application; Figure 7 This is a schematic diagram of the structure of a terminal provided in an embodiment of this application; Figure 8 This is a schematic diagram of the structure of a server provided in an embodiment of this application. Detailed Implementation

[0017] To enable those skilled in the art to better understand the solutions of this application, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0018] It should be noted that while some processes described in the specification, claims, and accompanying drawings include multiple steps appearing in a specific order, it should be clearly understood that these steps may not be performed in the order they appear herein, or may be performed in parallel. The step numbers are merely used to distinguish different steps and do not themselves represent any execution order. Furthermore, descriptions such as "first," "second," or "objective" in this document are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. "Multiple" in this document refers to at least two.

[0019] It is worth noting that in the specific embodiments of this application, multi-dimensional state data of nodes are involved, such as optical path state parameters, computing resource parameters, and service traffic data. When the above embodiments of this application are applied to specific products or technologies, permission or consent from the target object is required, and the collection, use, and processing of related data must comply with relevant laws, regulations, and standards. For example, when an embodiment of this application needs to obtain multi-dimensional state data of nodes, such as optical path state parameters, computing resource parameters, and service traffic data, separate permission or consent from the target object can be obtained through pop-up windows or redirection to a confirmation page. After obtaining separate permission or consent from the target object, the multi-dimensional state data of the nodes used to enable the embodiments of this application to operate normally, such as optical path state parameters, computing resource parameters, and service traffic data, can then be obtained.

[0020] Please see Figure 1 , Figure 1 This is a system architecture diagram of the optical computing fusion network dynamic scheduling method based on spatiotemporal graph convolution provided in this application embodiment. It includes a terminal 140, an Internet connection 130, a gateway 120, a server 110, etc.

[0021] Terminal 140 can take various forms, including desktop computers, laptops, PDAs (personal digital assistants), mobile phones, vehicle terminals, and dedicated terminals. Furthermore, it can be a single device or a collection of multiple devices. For example, multiple desktop computers can be interconnected via a local area network, sharing a single monitor to work collaboratively, forming a single terminal 140. Terminal 140 can communicate with the Internet 130 via wired or wireless means to exchange data.

[0022] Server 110 refers to a computer system capable of providing certain services to terminal 140. Compared to ordinary terminal 140, server 110 has higher requirements in terms of stability, security, and performance. Server 110 can be a single high-performance computer in a network platform, a cluster of multiple high-performance computers, a portion of a single high-performance computer (e.g., a virtual machine), or a combination of portions of multiple high-performance computers (e.g., virtual machines). Server 110 can also communicate with the Internet 130 via wired or wireless means to exchange data.

[0023] Gateway 120, also known as an internetwork connector or protocol converter, is a computer system or device that acts as a translator, enabling network interconnection at the transport layer. It bridges the gap between two systems using different communication protocols, data formats, languages, or even completely different architectures. Gateways can also provide filtering and security functions. Messages sent from terminal 140 to server 110 are forwarded to the corresponding server 110 via gateway 120. Messages sent from server 110 to terminal 140 are also forwarded to the corresponding terminal 140 via gateway 120.

[0024] The following provides a detailed description of the specific implementation methods of the embodiments of this application: Please see Figure 2 , Figure 2 This is a flowchart illustrating the dynamic scheduling method for optical computing fusion networks based on spatiotemporal graph convolution provided in this application embodiment. The dynamic scheduling method for optical computing fusion networks based on spatiotemporal graph convolution can be implemented by server 110 and / or terminal 140. Figure 2 The illustrated dynamic scheduling method for optical computing fusion networks based on spatiotemporal graph convolution includes: Step 210: Construct a unified resource map of the optical computing convergence network. The unified resource map is obtained by collecting multi-dimensional state data of each node of the optical computing convergence network in multiple historical time windows. Step 220: Extract and fuse spatiotemporal features of the unified resource graph using a pre-trained spatiotemporal graph convolutional network model to obtain the predicted state of each node in the next time window. The predicted state includes the predicted latency distribution, the predicted traffic load, and the predicted resource demand. Step 230: Establish a multi-dimensional joint optimization objective function related to the predicted latency distribution, the predicted traffic load, and the predicted resource demand; Step 240: Determine the probability score of each node, the probability score being used to characterize the confidence of a node as an intermediate node in the optimal scheduling path; Step 250: Prune the unified resource graph according to the probability scores of each node to obtain the pruned candidate subgraph; Step 260: Using a heuristic search algorithm, the optimal scheduling path is obtained by solving the multidimensional joint optimization objective function on the candidate subgraph; Step 270: Generate a collaborative scheduling instruction between the optical and electrical layers of the optical computing fusion network based on the optimal scheduling path, so as to perform collaborative scheduling of optoelectronic resources according to the collaborative scheduling instruction between the optical and electrical layers.

[0025] First, the optical-computing converged network in this application refers to a network architecture that deeply integrates optical transmission networks with computing resources (including computing nodes such as CPUs, GPUs, and NPUs). In this architecture, the optical layer is responsible for providing high-bandwidth, low-latency data transmission, while the computing layer (electrical layer) is responsible for providing computing task processing capabilities. The two achieve collaborative scheduling and joint optimization of resources through a unified control plane.

[0026] The pre-trained Spatio-Temporal Graph Convolutional Network (ST-GCN) proposed in this application is a deep learning model capable of simultaneously modeling the spatial dependencies and temporal dynamic evolution of graph-structured data. In this application, ST-GCN is used to extract the spatio-temporal features of the optical-computing fusion network on graph snapshot sequences (graph snapshot sequences in a unified resource graph), thereby predicting future resource demands and network states. Here, the spatio-temporal graph convolutional network can be trained using training samples composed of multiple graph snapshot sequences. In the training samples, normal and abnormal graph snapshot sequences are constructed, and a validation set is built to train a robust and generalizable spatio-temporal graph convolutional network.

[0027] In step 210, the topology information of the optical-computing converged network and the physical connections between nodes are acquired. The topology information includes the location and representation of each node, which is either an optical switching node or a computing node. Simultaneously, multi-dimensional status data of each node is collected, including optical path status parameters (including optical signal-to-noise ratio, wavelength occupancy rate, and spectrum fragmentation rate), computing resource parameters (including CPU / GPU utilization, memory occupancy rate, and task queue length), and historical service traffic data.

[0028] Using optical switching nodes and computing nodes as graph nodes, and physical fiber optic links between nodes as edges, an original graph structure G = (V, E) is constructed, where V represents the set of nodes and E represents the set of links. Each node is assigned a feature vector x_v, which contains the aforementioned multidimensional state data of that node. Continuous real-time state data (multidimensional state data) is sampled at fixed time intervals, i.e., the multidimensional state data of each node is collected through multiple historical time windows, forming a graph snapshot sequence {G_1, G_2, …, G_T}, where G_1 is the multidimensional state data of each node collected in the first historical time window, G_2 is the multidimensional state data of each node collected in the second historical time window, and G_T is the multidimensional state data of each node collected in the Tth historical time window.

[0029] In step 220, the spatial feature extraction module employs a graph convolutional network (GCN). Through convolution operations between the adjacency matrix (obtained from the original graph structure G) and the node feature matrix (obtained from the multidimensional state data of each node in the original graph structure G), it aggregates the feature information of the node itself (its own multidimensional state data) and the feature information of its neighboring nodes (the multidimensional state data of the nodes adjacent to the node). This extracts the aggregated spatial features and spatial dependencies between the nodes of the optical computing fusion network at each time step. The spatial dependencies include the traffic correlation, resource coupling degree, and link capacity constraints between adjacent nodes.

[0030] The time feature extraction module uses gated recurrent units (GRU) or one-dimensional temporal convolutional networks (TCN) to model the spatial feature sequence (multidimensional state data) of continuous time along the time dimension, capture the dynamic law of network state evolution over time, and obtain time features, which include periodic fluctuations in traffic, sudden congestion trends, and resource consumption decay patterns.

[0031] By fusing spatial and temporal features, the predicted state of each node within the next time window is output, including predicted latency distribution, predicted traffic load, and predicted resource requirements. The predicted latency distribution is obtained based on optical path state parameters, the predicted traffic load is obtained based on service traffic data, and the predicted resource requirements are obtained based on calculated resource parameters.

[0032] The ST-GCN model internally achieves deep fusion of spatial and temporal features by stacking multiple spatiotemporal convolutional blocks. Each spatiotemporal convolutional block sequentially performs spatial graph convolution (GCN) and temporal convolution (TCN). Spatial GCN is used to aggregate the neighbor information of a node in the fiber optic topology (feature information of the nodes adjacent to the node), while temporal TCN is used to capture the state evolution pattern along the time axis. The two operations are performed alternately until the time dimension is compressed to 1. At this point, each node corresponds to a spatiotemporal fusion representation vector. An example of dimensionality change (taking T=100, N=64, F=6 as an example, where T is the number of time dimensions, N is the number of spatial dimensions, and F is the number of feature dimensions) is shown in the table below:

[0033] Obtain the fused feature vector z for each node. v ∈R 256 Finally, a fully connected mapping layer is used to convert the fused feature vector of each node into three specific predicted values.

[0034]

[0035] Among them: W OUTIt is an n × 256 weight matrix; OUT It is an n × 1 bias vector; the three outputs correspond to the predicted traffic load, the predicted resource demand, and the predicted latency distribution, respectively.

[0036] The fusion feature mapping relationship is shown in the table:

[0037] In step 230, based on the predicted state obtained in step 220, a multi-dimensional joint optimization objective function is established that comprehensively considers transmission delay (related to the predicted delay distribution), bandwidth resource utilization efficiency (related to the predicted traffic load), computing power-task matching degree (related to the predicted resource demand), and energy consumption efficiency. The formula for the multi-dimensional joint optimization objective function min F is expressed as follows:

[0038] The end-to-end transmission delay function is defined as:

[0039] The propagation delay of the optical signal on the target link, The predicted value of node queuing delay (obtained from the predicted delay distribution output by ST-GCN).

[0040] For bandwidth resource utilization efficiency:

[0041] To predict the future capacity required for the target link (obtained from predicted traffic load). This represents the total available capacity. 1 This refers to bandwidth wastage rate.

[0042] Computing power-task matching degree:

[0043] This represents the current available computing power (CPU / GPU floating-point operations) of the target node. The required computing power is obtained by predicting resource demand. This is used in the objective function. This is used to punish those with insufficient computing power.

[0044] This represents the total energy consumption of all nodes and links along the path:

[0045] in This refers to the dynamic power consumption of optical switching nodes or computing nodes. This represents the power consumption of the erbium-doped fiber amplifier (EDFA).

[0046] Weighting coefficient For dynamically adjustable parameters, satisfying The system uses the following preset configuration table, depending on the different business types.

[0047]

[0048] In step 240, the scheduling path is optimized based on the graph convolutional network classifier, which uses the original graph structure (i.e., the unified resource graph) constructed in step 210. Based on this, where N is the total number of nodes and M is the total number of links, each node is generated through graph convolution operations. probability fraction , This indicates the confidence level of the node as an intermediate node in the optimal scheduling path.

[0049] (1) Input source: The hidden layer representation of the nodes in the intermediate layer of the ST-GCN model received by the graph convolutional network classifier in step 220. The hidden layer representation of this node It has integrated the spatial features of the node itself and its 2-3 hop neighbors, as well as the temporal features of multiple past moments, i.e., the hidden layer representation of the node. It integrates the aggregated spatial and temporal characteristics of each node.

[0050] (2) Calculation method: hv is subjected to a linear transformation and then activated by the Sigmoid function:

[0051] in and These are the trainable parameters for a graph convolutional network classifier.

[0052] In some embodiments, determining the probability score of each of the nodes includes: For each node, determine the hidden layer representation of the node output by the spatiotemporal graph convolutional network model for that node; The probability score of the node is calculated based on the hidden layer representation of the node and the Sigmoid function.

[0053] In step 250, a threshold pruning strategy is adopted: all... The nodes, among which The preset probability threshold (in this embodiment) (Values ​​between). The set of nodes retained after pruning. satisfy: That is, the number of nodes to be retained and the preset node retention threshold ( The difference between the values ​​is less than a preset difference, such as 2 or 3.

[0054] The original search space complexity is The order of magnitude; the search space complexity after pruning is:

[0055] With typical network scale For example, when At that time, the search space from Compress to The reduction rate exceeded 98%.

[0056] In step 260, on the pruned candidate subgraph, a heuristic search algorithm is used to solve for the optimal scheduling path and resource allocation scheme that satisfies the multidimensional joint optimization objective function. For example, A... The (A-Star) heuristic search algorithm finds the optimal scheduling path on the candidate subgraph G′.

[0057] (1) Formal definition of the problem: Given the pruned candidate subgraph G′=(V′, E′), where V′ represents the retained nodes and E′ represents the links formed between the retained nodes, the source node s, the destination node d, and the business requirements (… , ), the predicted state in step 220, and multiple weight vectors w( , , , Find candidate paths. satisfy: (Uniform constraints) (2) A Core formula of the algorithm: For each node v, define a cost estimation function f(v) = g(v) + h(v), which is also the cost constraint, where: g(v): The actual cumulative cost from the source node s to the current node v, which is obtained by summing all single-step ΔF on the path; h(v): The estimated residual cost from the current node v to the destination node d (heuristic function). The relaxation estimation method is used to ignore queuing delay and resource contention, and it is guaranteed that h(v)≤F(v→d) (preset admissibility condition).

[0058] The single-step expansion cost ΔF(v→u) is defined as follows, excluding the computing power term, because only the computing power at the endpoint is used to determine whether it is sufficient to complete the task.

[0059] ΔF(v→u)= In the formula, This is the difference between the end-to-end transmission delay of the current node and the end-to-end transmission delay of the destination node. This is the difference between the bandwidth resource utilization efficiency value of the current node and the bandwidth resource utilization efficiency value of the destination node. The difference between the total energy consumption of the current node and the total energy consumption of the destination node. (3) Iteration termination condition: When A The algorithm terminates when it retrieves the destination node d from the open list. The candidate path obtained at this point is the optimal path that satisfies all constraints. .

[0060] (4) Feasibility calculation explanation: Since |V′| ≈ 12~18 after pruning, the order of magnitude of the candidate paths is approximately 10. 5 A The algorithm can complete the solution within 0.1~1ms, meeting the frame-level scheduling latency requirements.

[0061] In step 270, as Figure 4 As shown, for high-volume services requiring large-capacity, long-distance transmission, end-to-end optical channels (wavelength-level scheduling) are prioritized, and dedicated wavelength resources are allocated for data transmission; the bandwidth requirement value in the service request parameters and the physical distance between the source and destination nodes are used for quantitative determination.

[0062] The 50Gbps threshold (first preset bandwidth requirement threshold): Currently, the typical capacity of a single wavelength in a commonly used backbone network is 100Gbps / 200Gbps, and 50Gbps is approximately 50% of that of a single wavelength. If the service bandwidth requirement exceeds this value, the cost of multi-port aggregation at the electrical layer is high, and allocating the entire wavelength is more economical and efficient (reducing port occupancy and forwarding hop count). For services below this value, carrying them through electrical layer packet switching offers greater flexibility.

[0063] 200km threshold (preset distance threshold): In the Optical Transport Network (OTN) standard, 200km corresponds to approximately 1ms of fiber propagation delay. For services exceeding this distance, the cumulative delay of end-to-end electrical layer hop-by-hop forwarding (approximately 50~200μs per hop) may exceed the service SLA requirements, making the low-latency advantage of direct optical layer hop-by-hop more significant.

[0064] For small-granular services that require flexible scheduling and fine-grained switching, configure electrical packet forwarding paths (packet-level scheduling) to fully utilize the high flexibility and low overhead characteristics of electrical layer switching; make comprehensive judgments based on the bandwidth requirements, service interaction modes, and burstiness in the service requests.

[0065] 1Gbps threshold (second preset bandwidth requirement threshold): 1Gbps corresponds to GE interface or 1 / 10 of 10GE. Services below this value are more flexible when carried by electrical layer packet switching (paths can be adjusted at any time, and bandwidth utilization can be improved by statistical multiplexing). Allocating optical layer wavelengths would result in significant resource fragmentation.

[0066] A peak-to-average power ratio (PAPR) of ≥ 3:1 indicates that the traffic exhibits significant bursts, making it suitable for the statistical multiplexing characteristics of electrical layer packet switching (multiple bursty services can share the same electrical layer path, utilizing their idle periods to stagger peak times). If fixed wavelength channels are allocated through the optical layer, the bandwidth of the wavelength will be significantly wasted during bursty low-peak periods.

[0067] For mixed service scenarios, the priority weights of optical channels and electrical packet forwarding are dynamically adjusted to achieve adaptive allocation and collaborative utilization of optical and electrical layer resources.

[0068] The coordinated control commands are sent to the optical cross-connect equipment and electrical packet switching equipment through the southbound interface to complete the scheduling and execution of the service flow.

[0069] The photoelectric collaborative scheduling in step 270 adopts the following dynamic adjustment process for mixed service scenarios: (a) Determination of Mixed Business Scenarios At the start of each scheduling cycle, the system retrieves the category tags of all services in the current queue of tasks to be scheduled, and counts the quantity and bandwidth requirements of each type of service in the queue. If there are ≥2 different service requirement types (latency-sensitive, compute-intensive, high-traffic, and balanced) in the queue, it is determined to be a mixed service scenario, triggering a dynamic adjustment process.

[0070] (ii) Calculation of the mixed weight vector The bandwidth-weighted average method is used to merge the weighting coefficients in the above configuration table according to the bandwidth ratio of various services:

[0071]

[0072]

[0073]

[0074] The same calculation applies to different business requirement types. , , The comprehensive weight vector in the mixed scenario is obtained. , , , ].

[0075] (III) Dynamic adjustment of the optical layer / electrical layer load-bearing ratio The services carried by the optical layer include not only "high-volume" (corresponding to β), but also "latency-sensitive" (corresponding to α, with low propagation latency in the optical layer) and "computation-intensive" (corresponding to γ). Therefore, the proportion of services carried by the optical layer should be driven by these three weights.

[0076]

[0077] The coefficients 0.6, 0.25, and 0.15 correspond to the degree of influence of β, α, and γ on the optical layer resource requirements, respectively. β has the highest influence (bandwidth level is the primary criterion for judging whether it is suitable for the optical layer), followed by α (latency-sensitive types also need low latency in the optical layer, but the data volume is usually not large), and then γ (computation-intensive types have a large data volume but a higher tolerance for latency than autonomous driving). 10% is the basic reservation ratio of optical layer resources. The energy consumption weight is not included in the calculation of the optical / electrical layer allocation because the energy consumption difference between the optical and electrical layers is a fixed physical characteristic. This factor has already been reflected in the energy consumption cost in the formula of the above multidimensional joint optimization objective function, and there is no need to repeatedly include it in the optical / electrical layer allocation process.

[0078] The above determination and calculation are repeated in each scheduling cycle (e.g., every 100 microseconds). When the service composition ratio changes, the system automatically updates the weight vector and the optical / electrical layer bearer ratio in the next scheduling cycle without manual intervention.

[0079] The system collects actual network status feedback information after execution at the frame-level (microsecond level), including actual traffic load, resource utilization, scheduling latency, and task completion rate. The feedback information is then compared with the prediction results to calculate the prediction deviation.

[0080] The predicted bias is used to incrementally learn and update the spatiotemporal graph convolutional network model, enabling online adaptive fine-tuning of the model and allowing it to continuously track dynamic changes in the network state. Simultaneously, the updated state information is distributed among adjacent nodes, forming a real-time control closed loop with frame-level response.

[0081] This embodiment takes the optical interconnect network of a kilo-card-level GPU cluster in a smart computing center as its application background, combined with Figure 3 and Figure 5 Describe the specific implementation process of this application.

[0082] 1. Network environment configuration: A converged optical computing network consisting of 64 compute nodes (each node equipped with 8 GPUs, totaling 512 GPUs) was configured, interconnected via a hybrid optoelectronic switching plane. The optical switching plane employs a wavelength selective switching (WSS) architecture, supporting dynamic configuration of 32 available wavelengths; the electrical switching plane uses an Ethernet packet switching architecture. Each node is configured with a data acquisition agent to collect status data such as GPU utilization, memory usage, inter-node communication traffic, and link optical signal-to-noise ratio at 10-microsecond intervals.

[0083] 2. Resource Map Construction The resource graph construction module receives status data reported by the data acquisition agent and constructs a unified resource graph. Sixty-four computing nodes are used as graph nodes, and edges are established between nodes with physical fiber optic connections to form an initial graph structure. For each time t, a multi-dimensional feature vector x_v(t) = [GPU utilization, memory usage, transmit queue length, receive queue length, optical signal-to-noise ratio, temperature] is extracted for each node, forming a graph snapshot sequence {G_1, G_2, …, G_100} (with a time window length of 100 sampling periods).

[0084] 3. Spatiotemporal Feature Extraction The graph snapshot sequence is input into the pre-trained ST-GCN model. The graph convolutional layer of this model adopts Chebyshev Graph Convolution (Chebyshev GCN), with the kernel order K set to 3. The spatial coupling features between each node and its 1-hop, 2-hop, and 3-hop neighbors are extracted through the Laplacian transformation of the adjacency matrix. The temporal convolutional layer adopts one-dimensional dilated causal convolution, with the dilation rate d set to [1, 2, 4, 8], and the receptive field covers 32 time steps.

[0085] The model outputs the predicted state of each node within the next 5 time windows (i.e., the next 50 microseconds): predicting the trend of GPU utilization, predicting the amount of communication traffic, predicting the probability of link congestion, and predicting the transmission latency.

[0086] 4. Optimize decision-making and path finding The optimization decision module receives the prediction results output by ST-GCN and constructs a multi-dimensional joint optimization objective function. For the AllReduce communication stream of the training task to be scheduled, the weight coefficients are set as α=0.4 (latency priority), β=0.2 (bandwidth efficiency), γ=0.3 (computing power matching), and δ=0.1 (energy consumption) to reflect the core requirements of the training task for low latency and high throughput.

[0087] The graph classifier takes the current graph structure as input and scores each node. For the i-th node, the classifier outputs a probability p_i, representing the confidence that the node is selected into the optimal scheduling path. The classifier prunes nodes with probability scores below a threshold θ (θ=0.3 in this embodiment), reducing the original 64-node graph to an average of 12-18 candidate subgraphs. On the candidate subgraphs, a Deep Q-Network (DQN) reinforcement learning algorithm is used to search for the optimal scheduling path. The search space is reduced from a combinatorial explosion size (~10^21 paths) to a controllable range (~10^5), and the algorithm converges within 1 millisecond.

[0088] 5. Optoelectronic Coordinated Scheduling Execution The optimization decision module determines the optimal scheduling scheme: for high-volume data channels spanning more than 5 nodes and with a data volume exceeding 50 GB, a dedicated wavelength channel is allocated for optical layer transmission (wavelength-level scheduling); for low-volume control information between nodes (less than 1 MB), electrical packet switching transmission is adopted (packet-level scheduling); for medium-volume data (1 GB to 50 GB), the optical layer and electrical layer transmission modes are dynamically switched according to the real-time link status.

[0089] The optoelectronic co-execution module generates control commands and sends them to the optical cross-connect device and the electrical packet switch through the OpenConfig southbound interface, completing path reconstruction within 2 milliseconds.

[0090] 6. Feedback Updates and Results Verification After the scheduling is completed, the data acquisition agent of each node collects actual status data in the next sampling period, including actual traffic volume, actual completion latency, and actual resource occupancy. The feedback update module calculates the prediction deviation Δ = |predicted value|. The actual value | When Δ exceeds the preset tolerance threshold (the threshold is set to 15% in this embodiment), the incremental learning and update process of the model is triggered. The ST-GCN model is lightly fine-tuned using the data of the most recent 50 sampling periods, and each fine-tuning takes about 5 milliseconds.

[0091] Experimental results show that, compared with the traditional centralized scheduling method based on OpenFlow, this embodiment reduces the average task completion latency by 47.6%, increases bandwidth utilization by 23.8%, improves energy efficiency by 18.3%, and reduces packet loss rate by 62.4% in scenarios with sudden traffic surges.

[0092] Compared with the prior art, this application has the following beneficial effects: (1) High-precision proactive predictive scheduling. This application captures the topological coupling relationship of the optical computing fusion network in the spatial dimension and the dynamic evolution law of traffic in the temporal dimension through a spatiotemporal graph convolutional network, thereby achieving advanced and accurate prediction of resource demand. Compared with traditional passive scheduling based on threshold alarms, it can effectively prevent congestion and improve the scheduling decision from passive response to proactive planning.

[0093] (2) Microsecond-level real-time response capability. This application collects status information at the frame level and forms a real-time feedback closed loop. Combined with a distributed status transmission mechanism, it breaks the bottleneck of the traditional vertical interaction mode of "node reporting - central decision-making - instruction issuance", significantly reducing scheduling latency from the periodic level to the frame level response, and significantly improving the ability to quickly adapt to sudden traffic.

[0094] (3) Coordinated utilization of optoelectronic resources. This application uses a unified optoelectronic coordinated scheduling strategy to dynamically allocate optical channels and electrical packet forwarding resources according to service characteristics, so as to realize the complementary coordination of ultra-high bandwidth transmission at the optical layer and flexible switching at the electrical layer, effectively improving the efficiency of bandwidth resource utilization and reducing transmission latency.

[0095] (4) Multidimensional global optimization. This application constructs a multidimensional joint optimization objective function that comprehensively considers transmission delay, bandwidth utilization, computing power matching degree and energy consumption efficiency. The scheduling decision is no longer limited to a single objective, but pursues the global optimal solution of integrated computing network resources. It is applicable to various application scenarios such as training stream scheduling, real-time big data transmission, and edge optical computing network.

[0096] Please see Figure 6 , Figure 6 This is a schematic diagram of the structure of a dynamic scheduling device for optical computing fusion networks based on spatiotemporal graph convolution provided in an embodiment of this application. This dynamic scheduling device for optical computing fusion networks based on spatiotemporal graph convolution is applied to computer equipment, and may include: The construction unit 601 is used to construct a unified resource map of the optical computing convergence network. The unified resource map is obtained by collecting multi-dimensional state data of each node of the optical computing convergence network in multiple historical time windows. The feature extraction unit 602 is used to extract and fuse spatiotemporal features of the unified resource graph through a pre-trained spatiotemporal graph convolutional network model to obtain the predicted state of each node in the next time window. The predicted state includes the predicted latency distribution, the predicted traffic load, and the predicted resource demand. The function establishment unit 603 is used to establish a multi-dimensional joint optimization objective function related to the predicted delay distribution, the predicted traffic load, and the predicted resource demand. The determining unit 604 is used to determine the probability score of each of the nodes, the probability score being used to characterize the confidence of a node as an intermediate node in the optimal scheduling path; The pruning unit 605 is used to prune the unified resource graph according to the probability scores of each node to obtain a pruned candidate subgraph. The solution unit 606 is used to obtain the optimal scheduling path on the candidate subgraph by using a heuristic search algorithm based on the multidimensional joint optimization objective function; The scheduling unit 607 is used to generate a collaborative scheduling instruction between the optical and electrical layers of the optical computing fusion network according to the optimal scheduling path, so as to perform collaborative scheduling of optoelectronic resources according to the collaborative scheduling instruction between the optical and electrical layers.

[0097] Reference Figure 7 , Figure 7 To implement the structural block diagram of a portion of the terminal 140 in this application embodiment, the terminal 140 includes: a radio frequency (RF) circuit 710, a memory 715, an input unit 730, a display unit 740, a sensor 750, an audio circuit 760, a wireless fidelity (WiFi) module 770, a processor 780, and a power supply 790, among other components. Those skilled in the art will understand that... Figure 7 The terminal 140 structure shown does not constitute a limitation on a mobile phone or computer, and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0098] The RF circuit 710 can be used to receive and transmit signals during information transmission or calls. In particular, it receives downlink information from the base station and processes it with the processor 780; in addition, it transmits uplink data to the base station.

[0099] The memory 715 can be used to store software programs and modules. The processor 780 executes various functional applications of the terminal and dynamic scheduling processing of the optical computing fusion network based on spatiotemporal graph convolution by running the software programs and modules stored in the memory 715.

[0100] The input unit 730 can be used to receive input numeric or character information, and to generate key signal inputs related to the terminal's settings and function control. Specifically, the input unit 730 may include a touch panel 731 and other input devices 732.

[0101] The display unit 740 can be used to display input or provided information, as well as various menus of the terminal. The display unit 740 may include a display panel 741.

[0102] Audio circuitry 760, speaker 761, and microphone 762 provide an audio interface.

[0103] In this embodiment, the processor 780 included in the terminal 140 can execute the dynamic scheduling method of optical computing fusion network based on spatiotemporal graph convolution in the previous embodiment.

[0104] The terminal 140 in this application embodiment includes, but is not limited to, mobile phones, computers, intelligent voice interaction devices, smart home appliances, vehicle terminals, and aircraft. This application embodiment can be applied to various scenarios, including but not limited to cloud technology, artificial intelligence, smart transportation, and assisted driving.

[0105] Figure 8 This is a partial structural block diagram of a server 110 implementing an embodiment of this application. The server 110 can vary significantly due to different configurations or performance characteristics, and may include one or more central processing units (CPUs) 822 (e.g., one or more processors) and memory 832, and one or more storage media 830 (e.g., one or more mass storage devices) for storing application programs 842 or data 844. The memory 832 and storage media 830 can be temporary or persistent storage. The program stored in the storage media 830 may include one or more modules (not shown in the diagram), each module including a series of instruction operations on the server 110. Furthermore, the CPU 822 may be configured to communicate with the storage media 830 and execute the series of instruction operations in the storage media 830 on the server 110.

[0106] Server 110 may also include one or more power supplies 826, one or more wired or wireless network interfaces 850, one or more input / output interfaces 858, and / or one or more operating systems 841, such as Windows Server™, Mac OS X™, Unix™, Linux™, FreeBSD™, etc.

[0107] The central processing unit 822 in server 110 can be used to execute the dynamic scheduling method for optical computing fusion networks based on spatiotemporal graph convolution according to the embodiments of this application.

[0108] This application also provides a computer-readable storage medium for storing program code for executing the dynamic scheduling method for optical computing fusion networks based on spatiotemporal graph convolution of the foregoing embodiments.

[0109] This application also provides a computer program product, which includes a computer program. A processor of a computer device reads and executes the computer program, causing the computer device to perform the aforementioned dynamic scheduling method for optical computing fusion networks based on spatiotemporal graph convolution.

[0110] Furthermore, the terms “comprising” and “including”, and any variations thereof, are intended to cover non-exclusive inclusion, such that a process, method, system, product, or apparatus that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or units that are not explicitly listed or that are inherent to such process, method, product, or apparatus.

[0111] It should be understood that in this application, "at least one (item)" means one or more, and "more than" means two or more. "And / or" is used to describe the relationship between related objects, indicating that three relationships can exist. For example, "A and / or B" can represent three cases: only A exists, only B exists, and both A and B exist simultaneously, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one (item) of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one (item) of a, b, or c can represent: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.

[0112] It should be understood that in the description of the embodiments of this application, "multiple" means two or more, "greater than", "less than", "exceeding" etc. are understood to exclude the number itself, and "above", "below", "within" etc. are understood to include the number itself.

[0113] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces, or indirect coupling or communication connection between apparatuses or units, and may be electrical, mechanical, or other forms.

[0114] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of the embodiments of this application, depending on actual needs.

[0115] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0116] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0117] It should also be understood that the various implementation methods provided in this application can be combined arbitrarily to achieve different technical effects.

[0118] In the embodiments of this application, the terms "module" or "unit" refer to a computer program or part of a computer program that has a predetermined function and works with other related parts to achieve a predetermined goal, and can be implemented wholly or partially using software, hardware (such as processing circuitry or memory), or a combination thereof. Similarly, a processor (or multiple processors or memory) can be used to implement one or more modules or units. Furthermore, each module or unit can be part of an overall module or unit that includes the functionality of that module or unit.

[0119] The above is a detailed description of the embodiments of this application. However, this application is not limited to the above embodiments. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of this application. All such equivalent modifications or substitutions are included within the scope defined by the claims of this application.

Claims

1. A dynamic scheduling method for optical computing fusion networks based on spatiotemporal graph convolution, characterized in that, The method includes: A unified resource map of the optical computing convergence network is constructed, which is obtained by collecting multi-dimensional state data of each node of the optical computing convergence network in multiple historical time windows. The unified resource graph is subjected to spatiotemporal feature extraction and fusion by a pre-trained spatiotemporal graph convolutional network model to obtain the predicted state of each node in the next time window. The predicted state includes predicted latency distribution, predicted traffic load and predicted resource demand. Establish a multidimensional joint optimization objective function related to the predicted latency distribution, the predicted traffic load, and the predicted resource demand; Determine the probability score of each node, the probability score being used to characterize the confidence of a node as an intermediate node in the optimal scheduling path; The unified resource graph is pruned based on the probability scores of each node to obtain a pruned candidate subgraph. The optimal scheduling path is obtained by solving the multidimensional joint optimization objective function on the candidate subgraph using a heuristic search algorithm. Based on the optimal scheduling path, a collaborative scheduling instruction for the optical and electrical layers of the optical computing fusion network is generated to perform collaborative scheduling of optoelectronic resources.

2. The dynamic scheduling method for optical computing fusion networks based on spatiotemporal graph convolution according to claim 1, characterized in that, The unified resource map for constructing the optical computing fusion network includes: Obtain the topology information of the optical computing fusion network, the topology information including the location of each node and the physical connection relationship between each node; Construct the original graph structure based on the position of each node and the physical connection relationship between each node; Within each historical time window, multidimensional status data of each node is collected. The multidimensional status data includes optical path status parameters, computing resource parameters, and service traffic data within the historical time window. The multidimensional state data of each node collected in each of the historical time windows are fused with the original graph structure to obtain the unified resource graph.

3. The dynamic scheduling method for optical computing fusion networks based on spatiotemporal graph convolution according to claim 1, characterized in that, The spatiotemporal graph convolutional network model includes a spatial feature extraction module and a temporal feature extraction module. The pre-trained spatiotemporal graph convolutional network model extracts and fuses spatiotemporal features of the unified resource graph to obtain the predicted state of each node within the next time window, including: The spatial feature extraction module extracts the adjacency matrix composed of each node in the unified resource graph and the node feature matrix composed of the multidimensional state data of each node. For each node, the multidimensional state data of the node itself and the multidimensional state data of each neighboring node adjacent to the node are aggregated to obtain the aggregated spatial features of the node. The time feature extraction module extracts spatial features from the multidimensional state data collected in each historical time window along the time dimension to obtain the time features of each node used to characterize the dynamic law of the optical computing fusion network's evolution over time. For each node, the aggregated spatial features and the temporal features of the node are fused to obtain the predicted state of the node in the next time window.

4. The dynamic scheduling method for optical computing fusion networks based on spatiotemporal graph convolution according to claim 1, characterized in that, Establish a multi-dimensional joint optimization objective function related to the predicted latency distribution, the predicted traffic load, and the predicted resource demand, including: The end-to-end transmission delay value is determined based on the predicted delay distribution, and a delay weight corresponding to the end-to-end transmission delay value is constructed. The bandwidth resource utilization efficiency value is determined based on the predicted traffic load, and a bandwidth weight corresponding to the bandwidth resource utilization efficiency value is constructed. The computing power-task matching degree is determined based on the predicted resource demand, and the computing power weight corresponding to the computing power-task matching degree is constructed. Determine the total energy consumption of all nodes and links on the target path, and construct the energy consumption weight corresponding to the total energy consumption. The target path is a path composed of any number of connected nodes. The multidimensional joint optimization objective function is constructed based on the end-to-end transmission delay value, the delay weight, the bandwidth resource utilization efficiency value, the bandwidth weight, the computing power-task matching degree, the computing power weight, the total energy consumption, and the energy consumption weight.

5. The dynamic scheduling method for optical computing fusion networks based on spatiotemporal graph convolution according to claim 4, characterized in that, The step of pruning the unified resource graph based on the probability scores of each node to obtain pruned candidate subgraphs includes: For each node, if the probability score of the node is lower than a preset probability threshold, the node is removed from the unified resource graph. Each node that was not removed is taken as a retained node, and the candidate subgraph after pruning is determined based on each of the retained nodes. Wherein, the difference between the number of retained nodes and the preset node retention threshold is less than the preset difference, and the preset node retention threshold is the product of the probability threshold and the total number of nodes in the unified resource graph.

6. The dynamic scheduling method for optical computing fusion networks based on spatiotemporal graph convolution according to claim 5, characterized in that, The step of employing a heuristic search algorithm to obtain the optimal scheduling path on the candidate subgraph based on the multidimensional joint optimization objective function includes: For each business requirement, determine the source node and destination node in the business requirement, and generate multiple candidate paths based on the source node and the destination node; Based on the end-to-end transmission delay value, the delay weight, the bandwidth resource utilization efficiency value, the bandwidth weight, the computing power-task matching degree, the computing power weight, the total energy consumption, and the energy consumption weight of the multidimensional joint optimization objective function, a unified constraint condition is constructed for each of the candidate paths. Construct the cost constraints for the candidate paths in the heuristic search algorithm; The heuristic search algorithm is used to search for the optimal scheduling path in each candidate path according to the unified constraint and the cost constraint. The optimal scheduling path satisfies the unified constraint and the cost constraint.

7. The dynamic scheduling method for optical computing fusion networks based on spatiotemporal graph convolution according to claim 6, characterized in that, The step of generating the optical and electrical layer collaborative scheduling instructions for the optical computing fusion network based on the optimal scheduling path includes: If the service requirement belongs to the service type of high-capacity long-distance transmission, an optical layer scheduling instruction for the end-to-end optical channel is generated to transmit the service requirement through the end-to-end optical channel in the optical layer according to the optical layer scheduling instruction. The service type of high-capacity long-distance transmission is used to indicate that the physical distance between the source node and the destination node in the service requirement is greater than a preset distance threshold or the bandwidth requirement value in the service requirement is greater than a first preset bandwidth requirement threshold. If the service requirement belongs to the service type that requires flexible scheduling and fine-grained switching, an electrical layer scheduling instruction for the electrical packet forwarding path is generated so that data can be transmitted through the electrical packet forwarding path of the electrical layer according to the electrical layer scheduling instruction. The service type that requires flexible scheduling and fine-grained switching is used to indicate that the bandwidth requirement value in the service requirement is less than the second preset bandwidth requirement threshold. If the business requirement belongs to the business type of mixed business scenario, a collaborative scheduling instruction for optical channel and electrical packet forwarding is generated. The delay weight, bandwidth weight, computing power weight and energy consumption weight in the multi-dimensional joint optimization objective function are dynamically adjusted according to the collaborative scheduling instruction. The optical layer carrying ratio and the electrical layer carrying ratio are determined according to the dynamically adjusted delay weight, bandwidth weight, computing power weight and energy consumption weight, and data transmission is performed according to the optical layer carrying ratio and the electrical layer carrying ratio. The business type belonging to the mixed business scenario is used to characterize the business requirement, which includes at least two types of business requirements, including latency-sensitive, computation-intensive, high-traffic, and balanced.

8. A dynamic scheduling device for optical computing fusion networks based on spatiotemporal graph convolution, characterized in that, The device includes: The construction unit is used to construct a unified resource map of the optical computing convergence network. The unified resource map is obtained by collecting multi-dimensional state data of each node of the optical computing convergence network in multiple historical time windows. The feature extraction unit is used to extract and fuse spatiotemporal features of the unified resource graph through a pre-trained spatiotemporal graph convolutional network model to obtain the predicted state of each node in the next time window. The predicted state includes predicted latency distribution, predicted traffic load, and predicted resource demand. The function establishment unit is used to establish a multi-dimensional joint optimization objective function related to the predicted latency distribution, the predicted traffic load, and the predicted resource demand. A determining unit is used to determine the probability score of each of the nodes, the probability score being used to characterize the confidence of a node as an intermediate node in the optimal scheduling path; The pruning unit is used to prune the unified resource graph according to the probability scores of each node to obtain the pruned candidate subgraph. The solution unit is used to obtain the optimal scheduling path on the candidate subgraph by employing a heuristic search algorithm based on the multidimensional joint optimization objective function; The scheduling unit is used to generate a coordinated scheduling instruction between the optical and electrical layers of the optical computing fusion network based on the optimal scheduling path, so as to perform coordinated scheduling of optoelectronic resources according to the coordinated scheduling instruction between the optical and electrical layers.

9. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the dynamic scheduling method for optical computing fusion networks based on spatiotemporal graph convolution as described in any one of claims 1 to 7.

10. A computer program product, the computer program product comprising a computer program, characterized in that, The computer program is read and executed by the processor of the electronic device, causing the electronic device to perform the dynamic scheduling method for optical computing fusion network based on spatiotemporal graph convolution as described in any one of claims 1 to 7.

Images