A unified modeling and tensor representation method and system for optical-computation multidimensional resources

Abstract

This invention provides a unified modeling and tensor representation method and system for multi-dimensional optical-computing resources, comprising: obtaining standardized feature vectors of heterogeneous computing resources and optical network resources through resource detection, normalization processing, and quantization encoding; constructing a resource tensor containing seven dimensions: computing power, memory, storage, wavelength, spectrum, path, and time, and defining rules for subtraction operations for resource occupancy, addition operations for resource release, and multiplication operations for resource reachability; mapping real-time resource states and task requirements to tensor representations using a state mapping algorithm, and realizing resource state transformation through tensor operations; employing a tensor space search algorithm guided by causal analysis, combining historical experience to align decision boundaries, and generating optimal resource allocation schemes; and simultaneously generating resource bottleneck and conflict early warning reports through historical state tensor sequence analysis. This invention improves resource utilization efficiency in complex environments.

Description

Technical Field

[0001] This invention relates to the fields of computing resource management and optical network resource optimization, and in particular to a unified modeling and tensor representation method and system for optical-computing multidimensional resources, applicable to resource management in a computing-network converged environment where high-performance computing, cloud computing, distributed computing, and optical networks are deployed together. Background Technology

[0002] Computational and network resource scheduling is a key technology area in modern computing systems, especially in distributed computing environments such as computing power networks. With the rapid development of computationally intensive applications such as artificial intelligence and big data analytics, and the accelerated construction of a nationwide integrated computing power network, higher demands are being placed on the collaborative management of computing and network resources.

[0003] In traditional technologies, resource management typically employs a layered approach, such as a "compute first, network second" or "network first, compute second" strategy. In the "compute first, network second" strategy, the system first allocates computing resources (such as CPU, GPU, memory, etc.) and then considers network connectivity; while in the "network first, compute second" strategy, the system prioritizes configuring network paths and bandwidth, and then allocates computing resources based on the network topology. For example, existing resource scheduling systems often implement computing resource management and network resource management as two independent subsystems, exchanging limited information between them through a simple interface.

[0004] Current advanced methods attempt to coordinate the management of computing and network resources, but the problem of inconsistent resource representation still exists. These methods typically use different data structures to represent computing resources (such as vectors or matrices to represent computing power, memory, etc.) and network resources (such as graph structures to represent network topology, bandwidth, etc.), and then use complex algorithms to establish a mapping relationship between the two different representations. This heterogeneous representation leads to a complex resource scheduling decision-making process, making it difficult to achieve global optimization, and is prone to resource fragmentation and conflicts under resource constraints, severely limiting the performance and service capabilities of computing-network convergence.

[0005] This fragmented resource management approach has significant technical drawbacks: First, the lack of a unified resource representation model makes end-to-end global resource optimization difficult, resulting in low resource utilization. Second, the separate resource management mechanism makes the system unable to quickly respond to dynamically changing resource demands, especially when dealing with tasks that have strict requirements for computing power and bandwidth, making it difficult to achieve joint optimization of computing nodes and optical network paths. In addition, the independence of decisions at different resource levels makes it difficult to predict resource bottlenecks, increasing system complexity and maintenance costs. Summary of the Invention

[0006] The purpose of this invention is to provide a unified modeling and tensor representation method and system for optical-computing multidimensional resources. By introducing a multidimensional resource tensor representation model, the originally separate computing resources (such as computing power, memory, and storage) and optical network resources (such as wavelength, spectrum, path, and time) are mapped to a unified mathematical space, and a complete tensor operation rule system is defined to achieve unified management and optimization of heterogeneous resources, thus solving the problems of inconsistent resource representation and fragmented management in traditional technologies.

[0007] To achieve the above objectives, this invention provides a unified modeling and tensor representation method for optical-computational multidimensional resources, comprising:

[0008] The characteristic parameters of heterogeneous computing resources and optical network resources are obtained, and standardized multidimensional resource feature vectors are obtained through resource detection, normalization processing and quantization encoding.

[0009] Based on the standardized multidimensional resource feature vector, a seven-dimensional resource tensor containing computing power dimension, memory dimension, storage dimension, wavelength dimension, spectrum dimension, path dimension and time dimension is obtained through dimension mapping and coordinate space construction. Tensor subtraction operation rules for resource occupation, tensor addition operation rules for resource release and tensor multiplication operation rules for resource reachability are defined to obtain a tensor operation system.

[0010] According to the tensor operation system, the real-time resource state is mapped to a resource state tensor and the task resource requirement is mapped to a resource requirement tensor through the state mapping algorithm. The resource state transformation is realized through tensor operation to obtain the resource state tensor and the corresponding transformation function.

[0011] Based on the resource state tensor, the resource demand tensor, and the resource state tensor and its corresponding transformation function, the resource allocation tensor and the corresponding joint scheduling decision scheme are obtained through tensor space search and optimization. Furthermore, resource bottleneck reports and resource conflict early warning reports are obtained through historical resource state tensor sequence analysis.

[0012] Furthermore, the acquisition of characteristic parameters of heterogeneous computing resources and optical network resources, through resource detection, normalization processing, and quantization encoding, yields a standardized multi-dimensional resource feature vector, including:

[0013] Based on the hardware specifications of heterogeneous computing nodes, raw parameter data of computing power, memory capacity, cache size and storage bandwidth of each computing node are collected to obtain the raw feature set of computing resources.

[0014] Based on the optical network topology and device status information, the number of available wavelengths, spectrum range, time slot allocation, and hop count and delay parameters of the optional routing paths are collected to obtain the original feature set of optical network resources;

[0015] Based on the original feature set of computing resources and the original feature set of optical network resources, resource parameters of different dimensions and orders of magnitude are converted into standardized values ​​within a unified numerical range through a normalization function, thereby obtaining a set of normalized feature values.

[0016] Based on the set of normalized feature values, the continuous normalized feature values ​​are mapped to discrete quantization levels through a quantization function and digitally encoded to obtain the standardized multidimensional resource feature vector.

[0017] Furthermore, based on the standardized multidimensional resource feature vector, a seven-dimensional resource tensor is obtained through dimensional mapping and coordinate space construction, comprising computing power, memory, storage, wavelength, spectrum, path, and time dimensions. Tensor subtraction rules for resource occupancy, tensor addition rules for resource release, and tensor multiplication rules for resource reachability are defined, resulting in a tensor operation system, including:

[0018] Based on the standardized multidimensional resource feature vector, the tensor dimensions, including computing power dimension, memory dimension, storage dimension, wavelength dimension, spectrum dimension, path dimension and time dimension, are determined by the dimension mapping rules, and a tensor dimension definition set is obtained.

[0019] Based on the tensor dimension definition set, a seven-dimensional coordinate space is obtained by mapping spatial coordinates to construct corresponding coordinate spaces for computing power dimension, memory dimension, storage dimension, wavelength dimension, spectrum dimension, path dimension and time dimension.

[0020] Based on the seven-dimensional coordinate space, the physical meaning and value rules of each element in the tensor are determined by the element assignment function, thus obtaining the tensor element value definition rules;

[0021] Based on the seven-dimensional coordinate space and the tensor element value definition rules, the standardized multi-dimensional resource feature vector is filled into the corresponding coordinate positions through the tensor initialization algorithm to obtain the seven-dimensional resource tensor.

[0022] Based on the seven-dimensional resource tensor, the tensor subtraction operation is defined to represent the remaining resource state after resources are allocated or occupied, and the tensor subtraction operation rule is obtained.

[0023] Based on the seven-dimensional resource tensor, the tensor addition operation is defined to represent the increased resource state after the resource is released or recycled, and the tensor addition operation rule is obtained.

[0024] Based on the seven-dimensional resource tensor, the tensor multiplication operation rules are obtained by defining tensor multiplication operations to represent the connectivity and reachability between different resource nodes.

[0025] Based on the tensor subtraction operation rules, the tensor addition operation rules, and the tensor multiplication operation rules, the non-negativity constraints, capacity upper limit constraints, and path continuity constraints that tensor operations must satisfy are defined through constraint condition functions, thus obtaining the tensor operation constraint set.

[0026] Based on the tensor subtraction operation rules, the tensor addition operation rules, the tensor multiplication operation rules, and the tensor operation constraint set, a complete tensor operation library containing subtraction operation functions, addition operation functions, multiplication operation functions, and constraint verification functions is implemented through software programming, thus obtaining the tensor operation system.

[0027] Furthermore, according to the tensor operation system, the real-time resource state is mapped to a resource state tensor and the task resource requirement is mapped to a resource requirement tensor through a state mapping algorithm, and the resource state transformation is achieved through tensor operations to obtain the resource state tensor and the corresponding transformation function, including:

[0028] Based on the actual resource status data and the seven-dimensional coordinate space of the seven-dimensional resource tensor, the CPU utilization, memory usage, storage input / output rate, wavelength occupancy, spectrum allocation status, path selection information, and time slice usage of the computing nodes in the current system are mapped to specific element values ​​in the seven-dimensional coordinate space through a state mapping algorithm, thereby obtaining the resource status tensor.

[0029] Based on the resource requirement description of the task or application, the resource requirement tensor is obtained by quantifying and mapping the task's requirements for computing power, memory capacity, storage bandwidth, number of wavelengths, spectrum resources, path length and time window to the seven-dimensional coordinate space through a requirement mapping algorithm.

[0030] Based on the resource state tensor, the resource demand tensor, and the tensor subtraction operation rules, the new state after resource allocation is calculated by tensor subtraction operation, and the state tensor after resource allocation is obtained by checking whether each element in the new state satisfies the non-negativity constraint through the tensor operation constraint set.

[0031] Based on the resource state tensor, the resource to be released tensor, and the tensor addition operation rules, the new state after resource release is calculated by tensor addition operation, and the tensor operation constraint set is used to check whether each element in the new state exceeds the capacity limit constraint, thus obtaining the state tensor after resource release.

[0032] Based on the state tensor after resource allocation and the state tensor after resource release, the output state tensor after the input state tensor is defined by the state transition function, thus obtaining the resource state tensor and the corresponding transition function.

[0033] Furthermore, before obtaining the resource allocation tensor and the corresponding joint scheduling decision scheme through tensor space search and optimization based on the resource state tensor, the resource demand tensor, and the resource state tensor and its corresponding transformation function, the process further includes:

[0034] Based on the resource state tensor and its corresponding transition function, and the state transition sequence recorded in the system evolution history, the resource state tensor and operation sequence are combined and represented as a time-series decision sequence through a sequence coding algorithm to obtain the resource state transition sequence model.

[0035] Based on the tensor subtraction, tensor addition, and tensor multiplication operations in the tensor operation system, a set of allowed resource operation tags is defined through a tag set construction algorithm to obtain a tag vocabulary;

[0036] Based on the resource state transition sequence model and the labeled vocabulary, a probability distribution for selecting the next operation in the current state is constructed through a conditional probability model, and the tensor operation constraint set is used as a hard constraint to obtain a constraint sequence transition model.

[0037] Based on the constrained sequence transition model, the resource state tensor, and the resource demand tensor, an operation sequence that satisfies the resource demand tensor and maximizes the state transition probability is generated in the operation space of the labeled vocabulary by a sequence generation algorithm, thereby obtaining a resource allocation operation sequence plan.

[0038] Based on the resource allocation operation sequence plan and the resource state tensor and its corresponding transformation function, the intermediate state tensor corresponding to each step of the operation is calculated sequentially through the sequence-tensor mapping function to obtain the intermediate state tensor sequence and the final resource allocation state tensor.

[0039] Further, the step of obtaining the resource allocation tensor and the corresponding joint scheduling decision scheme based on the resource state tensor, the resource demand tensor, and the resource state tensor and its corresponding transformation function through tensor space search and optimization includes:

[0040] Based on the resource demand tensor and the resource state tensor, the matching degree between task requirements and available system resources in the dimensions of computing power, memory, storage, wavelength, spectrum, path, and time is calculated using the tensor matching degree calculation function to obtain a set of resource matching degree indicators.

[0041] Based on the resource matching degree index set and the system optimization objective, weights are assigned to each dimension of the matching degree index through weight allocation, and a multi-objective optimization function containing a matching degree weighted summation term, a delay penalty term, and a cost penalty term is constructed to obtain the optimization objective function;

[0042] Based on the optimization objective function and the resource state tensor, the causal dependencies between the resource state variables of each dimension in the resource state tensor and the weighted summation term, delay penalty term, and cost penalty term in the optimization objective function are analyzed using a causal discovery algorithm. A causal relationship graph is constructed, and an intervention operator is defined to represent the operation of setting the numerical values ​​of the resource dimension variables in the causal relationship graph. By calculating the difference between the first function value of the optimization objective function before the intervention operator is applied and the second function value of the optimization objective function after the intervention operator is applied, a causal effect matrix is ​​obtained.

[0043] Based on the causal effect matrix and the optimization objective function, the resource dimension variable with the largest causal effect value is selected as the priority intervention object through the causal intervention optimization algorithm. A search strategy based on causal effect ranking in the seven-dimensional coordinate space is designed to obtain a search strategy based on causal relationship.

[0044] Based on the causal search strategy, the resource state tensor, and the resource demand tensor, the resource allocation tensor is obtained by using a tensor space search algorithm to find the resource allocation tensor that maximizes the value of the optimization objective function under the premise of satisfying the resource non-negativity constraint.

[0045] Based on the resource allocation tensor, the element values ​​of each dimension of the allocation tensor in the tensor space are converted into CPU core allocation instructions, memory block allocation instructions, storage channel allocation instructions, wavelength allocation instructions, spectrum slot allocation instructions, routing path selection instructions, and time slice scheduling instructions through the decision mapping function, thereby obtaining the joint scheduling decision scheme.

[0046] Further, the step of finding the resource allocation tensor that maximizes the value of the optimization objective function based on the causal search strategy, the resource state tensor, and the resource demand tensor using a tensor space search algorithm under the premise of satisfying the resource non-negativity constraint, and obtaining the resource allocation tensor, includes:

[0047] Based on the causal effect ranking results in the causal relationship-based search strategy and the resource state tensor, intervention operations are sequentially performed on multiple resource dimension variables with the largest causal effect values ​​through causal intervention tensor space search, generating multiple sets of candidate intervention operations and corresponding resource allocation tensor sets, and obtaining a set of candidate resource allocation schemes;

[0048] Based on the set of candidate resource allocation schemes and the set of historically optimal resource allocation tensors stored in historical scheduling experience, the Euclidean distance between the candidate scheme and the historically optimal scheme is calculated using a decision boundary alignment algorithm. The historically optimal scheme with the smallest distance is selected as the reference scheme, and the allocation values ​​of each dimension of the candidate scheme are adjusted so that the allocation values ​​of each dimension of the candidate scheme are close to the corresponding allocation values ​​of the reference scheme, thus obtaining the resource allocation tensor.

[0049] Furthermore, the resource bottleneck report and resource conflict early warning report obtained through historical resource state tensor sequence analysis include:

[0050] Based on the resource status data recorded during system operation, a multi-scale sampling strategy is used to extract historical resource status tensors at three time granularities: second, minute, and hour, to obtain a multi-granularity historical information set.

[0051] Based on the multi-granularity historical information set, a three-level coding network is used to encode second-level historical information into primary features, minute-level historical information is fused with primary features and then encoded into intermediate features, and hour-level historical information is fused with intermediate features and then encoded into advanced features, resulting in a progressive historical coding representation.

[0052] Based on the seven-dimensional coordinate space of the seven-dimensional resource tensor and the historical resource usage pattern, a representative query anchor point is defined in each of the computing power dimension, memory dimension, storage dimension, wavelength dimension, spectrum dimension, path dimension and time dimension through the key information extraction algorithm to obtain a set of query anchor points;

[0053] Based on the set of query anchors and the progressive historical encoding representation, the attention weight between each query anchor and the historical encoding feature is calculated through a parallel attention mechanism. The historical encoding features are weighted and summed to obtain the context vector of each query anchor, thus obtaining the anchoring inference result.

[0054] Based on the anchored inference results and the resource state tensor, the ratio of used resources to total available resources is calculated in each dimension using an enhanced utilization analysis function and combined with historical trend data to obtain a resource utilization tensor containing utilization values ​​for each of the seven dimensions. The utilization values ​​of each dimension are extracted from the resource utilization tensor, and the resource dimensions with utilization values ​​exceeding a preset threshold are selected as bottleneck points using a bottleneck identification algorithm to obtain a resource bottleneck report containing bottleneck dimension identification, bottleneck degree, and historical trend analysis.

[0055] Based on the progressive historical encoding representation and the anchored inference results, a short-term prediction model is constructed through a multi-timescale prediction network to predict the resource status in the next five minutes, a medium-term prediction model to predict the resource status in the next hour, and a long-term prediction model to predict the resource status in the next twenty-four hours. The future resource status prediction is generated by weighted fusion, and the predicted resource status tensor set is obtained.

[0056] Based on the predicted resource state tensor set and the known future task requirement tensor set, a conflict detection function is used to check whether any dimension element of the predicted resource state minus the future task requirement is negative. If it exists, it is marked as a resource allocation conflict, and a resource conflict early warning report is obtained, which includes the time of conflict occurrence, the dimension of conflicting resources, and the severity of conflict.

[0057] Furthermore, based on the progressive historical encoding representation and the anchored inference results, a short-term prediction model is constructed using a multi-timescale prediction network to predict the resource status for the next five minutes, a medium-term prediction model to predict the resource status for the next hour, and a long-term prediction model to predict the resource status for the next twenty-four hours. These predictions are then weighted and fused to generate a future resource status prediction, resulting in a predicted resource status tensor set, including:

[0058] Based on the progressive historical encoding representation and the anchored inference results, a long short-term memory network is constructed as a short-term prediction model by inputting the primary features and the anchored inference results into the temporal decomposition network, a gated recurrent unit network is constructed as a medium-term prediction model by inputting the intermediate features and the anchored inference results into the gated recurrent unit network, and an attention mechanism network is constructed as a long-term prediction model by inputting the high-level features and the anchored inference results into the attention mechanism network, thus obtaining a multi-scale prediction model set;

[0059] Based on the multi-scale prediction model set and the resource state tensor, the five-minute prediction results output by the short-term prediction model, the one-hour prediction results output by the medium-term prediction model, and the twenty-four-hour prediction results output by the long-term prediction model are calculated by the model fusion algorithm, and then fused by the weighted fusion method to obtain the prediction resource state tensor set.

[0060] This invention also provides a unified modeling and tensor representation system for optical-computational multidimensional resources, comprising:

[0061] The resource feature acquisition module is used to acquire feature parameters of heterogeneous computing resources and optical network resources. Through resource detection, normalization processing and quantization encoding, a standardized multi-dimensional resource feature vector is obtained.

[0062] The tensor construction module is used to construct a seven-dimensional resource tensor based on the standardized multidimensional resource feature vector through dimension mapping and coordinate space construction, which includes computing power dimension, memory dimension, storage dimension, wavelength dimension, spectrum dimension, path dimension and time dimension. It also defines the tensor subtraction operation rules for resource occupation, the tensor addition operation rules for resource release and the tensor multiplication operation rules for resource reachability, thus obtaining the tensor operation system.

[0063] The state mapping module is used to map real-time resource states to resource state tensors and task resource requirements to resource requirement tensors according to the tensor operation system, and to realize resource state transformation through tensor operation to obtain resource state tensors and corresponding transformation functions.

[0064] The resource scheduling module is used to obtain the resource allocation tensor and the corresponding joint scheduling decision scheme based on the resource state tensor, the resource demand tensor, the resource state tensor and the corresponding transformation function, through tensor space search and optimization, and to obtain resource bottleneck report and resource conflict early warning report through historical resource state tensor sequence analysis.

[0065] The optical-computational multidimensional resource unified modeling and tensor representation method and system of the present invention have the following beneficial effects:

[0066] 1. A seven-dimensional tensor model R(c, m, s, w, f, p, t) is proposed to unify the representation of heterogeneous computing resources and optical network resources. This model maps the originally separate computing resources (computing power, memory, storage) and network resources (wavelength, spectrum, path, time) to a unified mathematical space, solving the problem of inconsistent resource representation in traditional technologies.

[0067] 2. A complete set of tensor operation rules was designed, including tensor subtraction for resource occupancy, tensor addition for resource release, and tensor multiplication for resource reachability. This realizes the mathematical expression and operation of resource status, and improves the standardization and computability of resource management.

[0068] 3. A mapping method from resource state to tensor space was established. Through the state mapping algorithm, the real-time resource state and task requirements were quantified into tensor representation, realizing unified management of heterogeneous resources and enhancing the system's ability to express complex resource states.

[0069] 4. An innovative joint resource scheduling mechanism based on tensor space search is proposed, which can simultaneously optimize the selection of computing nodes and the configuration of optical paths in a unified tensor space, avoiding the resource fragmentation problem of traditional "computation first, then network" or "network first, then computation", and improving the global optimality of resource allocation.

[0070] 5. A resource bottleneck identification and conflict prediction technology based on tensor analysis was developed. By analyzing the tensor sequence of historical resource status and predicting future resource status, potential resource problems can be identified in advance and optimization suggestions can be provided, thereby enhancing the system's foresight and reliability. Attached Figure Description

[0071] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0072] Figure 1 This is a flowchart illustrating the optical-computational multidimensional resource unified modeling and tensor representation method of the present invention;

[0073] Figure 2 This is a schematic diagram illustrating the process of constructing the seven-dimensional resource tensor and coordinate space in this invention.

[0074] Figure 3 This is a schematic diagram of the structure of the optical-computational multidimensional resource unified modeling and tensor representation system of the present invention. Detailed Implementation

[0075] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0076] Example 1

[0077] like Figure 1 As shown, this invention provides a unified modeling and tensor representation method for optical-computational multidimensional resources, including:

[0078] Step S1: Obtain the characteristic parameters of heterogeneous computing resources and optical network resources, and obtain a standardized multidimensional resource feature vector through resource detection, normalization processing and quantization encoding.

[0079] Step S2: Based on the standardized multidimensional resource feature vector, a seven-dimensional resource tensor containing computing power dimension, memory dimension, storage dimension, wavelength dimension, spectrum dimension, path dimension and time dimension is obtained through dimension mapping and coordinate space construction. Tensor subtraction operation rules for resource occupation, tensor addition operation rules for resource release and tensor multiplication operation rules for resource reachability are defined to obtain the tensor operation system.

[0080] Step S3: According to the tensor operation system, the real-time resource state is mapped to a resource state tensor and the task resource requirement is mapped to a resource requirement tensor through the state mapping algorithm. The resource state transformation is realized through tensor operation to obtain the resource state tensor and the corresponding transformation function.

[0081] Step S4: Based on the resource state tensor, the resource demand tensor, and the resource state tensor and its corresponding transformation function, the resource allocation tensor and the corresponding joint scheduling decision scheme are obtained through tensor space search and optimization. Resource bottleneck report and resource conflict early warning report are obtained through historical resource state tensor sequence analysis.

[0082] Specifically, firstly, this step involves collecting characteristic parameters of various heterogeneous computing resources in the system through a resource detection agent. These parameters include the number of CPU cores, clock speed, architecture type, internal cache size, single-precision and double-precision floating-point operation capabilities, GPU / TPU accelerator computing power indicators, total memory capacity, memory bandwidth, storage capacity, and storage I / O rate. Simultaneously, it acquires optical network resource characteristic parameters through the monitoring interface of the optical network management system. These parameters include the number and topology of fiber optic links, the total number of available wavelength channels on each link, the center frequency and bandwidth of each wavelength, the spectrum allocation for each wavelength, the available routing paths between nodes in the optical network and their physical characteristics (such as transmission distance, hop count, and signal quality parameters), and the time slot configuration of the time-division multiplexing system. The collected raw parameters typically have different units of measurement and numerical ranges. Normalization is used to convert these heterogeneous parameters to a uniform numerical range (usually between 0 and 1) to eliminate dimensional differences. Different normalization strategies are employed for different types of parameters: linear parameters (such as utilization) use Min-Max normalization; parameters with exponential growth (such as bandwidth) use logarithmic normalization; and discrete categorical parameters (such as processor type) use one-hot encoding. After normalization, quantization encoding is further applied to map continuous normalized values ​​to discrete quantization levels, improving system processing efficiency and robustness. The quantization process sets different quantization precisions based on parameter importance and sensitivity to change; core resource parameters use high-precision quantization (such as 16-bit or 32-bit), while auxiliary parameters use low-precision quantization (such as 8-bit or 4-bit). The resulting standardized multidimensional resource feature vector is a structured dataset containing complete feature descriptions of computational and network resources, providing a unified data foundation for subsequent tensor construction.

[0083] Second, the standardized multidimensional resource feature vectors are mapped to seven core dimensions through dimensional mapping: computing power represents the processor's computing capabilities, memory represents system memory resources, storage represents persistent storage resources, wavelength represents wavelength channel resources in the optical network, spectrum represents spectrum resources within each wavelength, path represents network connection paths, and time represents the temporal allocation of resources. Each dimension has its specific physical meaning and metric. A dedicated coordinate space is constructed for each dimension, defining the coordinate scale and resolution. For example, the computing power dimension uses a logarithmic scale to cover a wide range of computing power from low-end to high-end; the wavelength dimension directly maps to physical wavelength numbers; and the time dimension supports multi-scale representation, from milliseconds to days in time granularity. The coordinate spaces of each dimension are integrated to form a unified seven-dimensional coordinate system, in which each resource state point has a unique coordinate representation. The physical meaning and value rules of tensor elements are also defined: element values ​​may represent resource occupancy status (e.g., used / available), resource allocation ratio, or resource quality indicators. Based on the constructed seven-dimensional resource tensor, three core tensor operation rules are defined: tensor subtraction represents the state change after a resource is allocated or occupied, applicable to the resource allocation process; tensor addition represents the state restoration after a resource is released, applicable to the resource reclamation process; and tensor multiplication represents the connectivity and reachability between resources, used for evaluating resource combinations and path verification. These operation rules also include constraint verification mechanisms to ensure that the operation results meet physical constraints, such as resource non-negativity, capacity limits, and path continuity. Through software implementation, these concepts and rules constitute a complete tensor operation system, providing a mathematical foundation for subsequent resource state representation and operations.

[0084] Third, the latest resource status information is collected from various monitoring platforms, including CPU / GPU utilization, memory usage, storage I / O performance, wavelength usage in the network, spectrum allocation status, active routing paths, and time slice usage. This raw status data, after preprocessing (such as noise removal and missing value imputation), is mapped to corresponding positions in a seven-dimensional coordinate space using a state mapping algorithm, forming a tensor representation of the current resource status. Similarly, the resource requirements of a task or application are converted into a resource requirement tensor, representing the specific amount of resources required by the task in each dimension. The requirement description may come from directly specified resource configuration requirements or may be derived from a model based on performance goals. The conversion process considers both the determinism and flexibility of the requirements; some dimensions may have explicit hard requirements, while others may have room for flexible adjustment. With the resource status tensor and resource requirement tensor, the impact of resource operations on the system state is simulated using defined tensor operation rules. For example, the resource allocation process is simulated using tensor subtraction: subtracting the resource requirement tensor from the current resource status tensor yields the new state after resource allocation. Similarly, the resource release process is simulated using tensor addition. These simulations undergo constraint checks to ensure the results meet physical constraints. The system abstracts these state transition processes into state transition functions, each defining the state change pattern under a specific operation (such as allocation, release, reservation, and migration). These functions not only contain basic state calculation logic but also include precondition checks and postcondition verifications to ensure that each transition is valid and safe.

[0085] Fourth, the optimization objective function is defined first, typically including multiple objectives: maximizing resource utilization, minimizing task execution time, minimizing energy consumption, etc. To comprehensively consider these objectives, a weighted optimization function is constructed, including components such as resource matching score, latency penalty, and cost penalty. The optimization process employs a multi-stage search strategy: first, a causal discovery algorithm is used to analyze the causal relationship between resource dimension variables and the optimization objective, identifying the dimension with the greatest impact; then, based on the causal analysis results, a search order is designed, prioritizing the optimization of key dimensions; finally, an optimization search is performed in the seven-dimensional tensor space to find an allocation scheme that satisfies resource constraints and maximizes the optimization objective value. The search process combines algorithms such as branch and bound, simulated annealing, and references historical best solutions for decision boundary alignment, balancing theoretical calculations and practical experience. The resulting resource allocation tensor precisely defines the allocation decisions for resources in each dimension, converting them into specific scheduling instructions, including executable operation commands such as CPU core allocation, memory allocation, storage channel configuration, wavelength allocation, spectrum configuration, path selection, and time slice scheduling. Simultaneously, the historical resource state tensor sequence is analyzed to identify bottlenecks and potential conflicts in system operation. This analysis process includes multi-granularity historical information sampling, progressive historical coding, and anchor point inference. By calculating resource utilization and historical trends across various dimensions, bottleneck dimensions with utilization exceeding preset thresholds are identified, generating detailed resource bottleneck reports. Multi-timescale predictive network analysis of future resource status trends can predict potential resource allocation conflicts and generate early warning reports, including conflict timing, involved dimensions, and severity. These analysis reports provide system administrators with valuable decision support information, helping to optimize resource allocation and plan resource expansion in advance.

[0086] Through the above steps, this invention achieves unified modeling and representation of heterogeneous computing resources and optical network resources, overcoming the limitations of separate management of computing and network resources in traditional resource management, and providing an efficient and flexible unified resource scheduling method. The seven-dimensional resource tensor model can comprehensively capture the multidimensional characteristics of system resources, the tensor operation system provides a mathematical foundation for resource management, the causal analysis-based optimization search significantly improves resource allocation efficiency, and historical sequence analysis enhances the system's predictive and proactive management capabilities. This invention is particularly suitable for large-scale heterogeneous computing environments and optical network convergence scenarios, such as edge computing, cloud-network converged systems, and next-generation data centers.

[0087] Example 2

[0088] In this embodiment, the acquisition of feature parameters of heterogeneous computing resources and optical network resources, through resource detection, normalization processing, and quantization encoding, yields a standardized multidimensional resource feature vector, including:

[0089] Based on the hardware specifications of heterogeneous computing nodes, raw parameter data of computing power, memory capacity, cache size, and storage bandwidth of each computing node are collected to obtain a raw feature set of computing resources. In this embodiment, this step is implemented through a resource probing agent deployed on each computing node. For computing power characteristics, the number of CPU cores, clock speed, architecture type, floating-point operation capability (FLOPS), and computing power indicators of GPU / TPU accelerators are collected; for memory capacity, the total memory size, currently available memory, memory speed, and number of memory channels are collected; for cache characteristics, the size and read / write speed of L1 / L2 / L3 cache levels are collected; for storage characteristics, indicators such as disk I / O bandwidth, read / write latency, and throughput are collected. These raw parameters are stored in structured data form, constituting the raw feature set of computing resources. This feature set comprehensively reflects the differences in computing power and resource characteristics of each node in the heterogeneous computing environment, providing a data foundation for subsequent unified modeling.

[0090] Based on the optical network topology and device status information, the number of available wavelengths, spectrum range, time slot allocation, and hop count and delay parameters of optional routing paths are collected to obtain the original feature set of optical network resources. This step directly obtains the optical network device status through the southbound interface of the optical network management system. For wavelength resources, the total number of wavelengths on each link, the number of occupied wavelengths, the list of available wavelengths and their center frequencies are collected; for spectrum resources, the total spectrum bandwidth, occupied spectrum intervals, and spectrum utilization are collected; for time slot resources, the total number of time slots in the time-division multiplexing system, the allocated time slot mapping table, and the time slot granularity are collected; for routing resources, the network topology map, the set of optional paths between each node pair, the optical signal quality parameters of each path, the hop count, and the end-to-end delay are collected. These network resource parameters are initially normalized through the optical network resource abstraction layer to constitute the original feature set of optical network resources. This feature set describes in detail the availability, connectivity, and performance characteristics of optical network resources and is an important component of the unified modeling of optical-computing resources.

[0091] Based on the original feature sets of computing resources and optical network resources, resource parameters of different dimensions and orders of magnitude are converted into standardized values ​​within a unified numerical range using a normalization function, resulting in a set of normalized feature values. Normalization is a crucial step in resolving inconsistencies in the dimensions of different resource parameters. This embodiment employs an adaptive normalization method, selecting different normalization strategies for different types of resource parameters. For continuous numerical parameters (such as CPU utilization and memory capacity), a Min-Max normalization method is used, mapping the original values ​​to the [0, 1] interval. For unevenly distributed parameters (such as computing power and bandwidth), logarithmic normalization is applied followed by Min-Max normalization. For discrete parameters (such as wavelength number and path ID), an ordinal normalization method is used. The normalization function also considers the physical upper limit of resource parameters and actual scenario requirements, dynamically adjusting the boundary values ​​of the normalization interval to improve numerical accuracy. Through this process, the original features of computing and network resources are converted into standardized values ​​with unified dimensions, facilitating subsequent tensor construction and computation.

[0092] Based on the set of normalized eigenvalues, continuous normalized eigenvalues ​​are mapped to discrete quantization levels using a quantization function and then digitally encoded to obtain the standardized multidimensional resource feature vector. Quantization encoding is the process of discretizing normalized continuous values, which helps reduce computational complexity and improve system robustness. This embodiment employs a non-uniform quantization method, setting different quantization levels and encoding precision according to the importance and sensitivity of resource parameters. Specifically, for key resource dimensions such as computing power and wavelength, fine-grained quantization (e.g., 16 or 32 levels) is used to maintain high precision; for parameters that do not change frequently (e.g., path hop count), coarse-grained quantization (e.g., 4 or 8 levels) is used. The quantization function first determines the number of quantization levels for each parameter, then maps the normalized eigenvalues ​​to the corresponding levels, and finally generates digital codes. After this processing, the resulting standardized multidimensional resource feature vector has a unified representation and encoding rules, providing standardized data input for constructing a seven-dimensional resource tensor. This quantization encoding strategy significantly improves the efficiency of tensor operations and system scalability while maintaining resource representation precision.

[0093] Example 3

[0094] like Figure 2 As shown, in this embodiment, based on the standardized multidimensional resource feature vector, a seven-dimensional resource tensor is obtained through dimension mapping and coordinate space construction, including computing power dimension, memory dimension, storage dimension, wavelength dimension, spectrum dimension, path dimension, and time dimension. Tensor subtraction rules for resource occupancy, tensor addition rules for resource release, and tensor multiplication rules for resource reachability are defined, resulting in a tensor operation system, including:

[0095] Step S21: Based on the standardized multi-dimensional resource feature vector, the tensor dimensions, including computing power, memory, storage, wavelength, spectrum, path, and time dimensions, are determined through dimension mapping rules to obtain a tensor dimension definition set. Dimension mapping rules are a set of rules that map each element in the resource feature vector to the corresponding dimension in the tensor model. In this embodiment, dimension mapping uses a combination of feature clustering and dimension projection. First, correlation analysis is performed on the standardized multi-dimensional resource feature vector, clustering highly correlated features into the same dimension; then, a clear physical meaning and coordinate definition are defined for each dimension. The computing power dimension reflects the processing power of a node, including the comprehensive computing power level of heterogeneous computing resources such as CPUs and GPUs; the memory dimension characterizes the system's storage capacity, including memory capacity and access speed; the storage dimension describes persistent storage resources, including storage capacity and IO bandwidth; the wavelength dimension corresponds to the available wavelength resources in the optical network; the spectrum dimension represents the allocatable spectrum resources within each wavelength; the path dimension describes the connection paths between nodes in the network; and the time dimension reflects the dynamic changes of resources over time. Through this mapping rule, a transformation relationship from standardized feature vectors to a seven-dimensional tensor model was established, resulting in a tensor dimension definition set that includes specific dimension definitions, ranges, granularities, and physical meanings.

[0096] Step S22: Based on the tensor dimension definition set, construct corresponding coordinate spaces for each of the computing power dimension, memory dimension, storage dimension, wavelength dimension, spectrum dimension, path dimension, and time dimension through spatial coordinate mapping, resulting in a seven-dimensional coordinate space. Spatial coordinate mapping is the process of establishing a specific coordinate system for each tensor dimension. In this embodiment, for the computing power dimension, the coordinate space adopts a logarithmic scale distribution to better represent a wide range from low to high computing power; for the memory and storage dimensions, a piecewise linear coordinate mapping is used, with fine-grained partitioning within the commonly used capacity range and coarse-grained partitioning within the maximum or minimum value range; for the wavelength dimension, the coordinate space directly corresponds to the physical wavelength channel number; for the spectrum dimension, a uniformly divided spectrum slot representation is used; for the path dimension, a path index space is constructed based on the network topology graph, with each coordinate point corresponding to a possible optical path; for the time dimension, a multi-scale time scale representation is used, supporting time granularity from seconds to days. Through these specially designed coordinate mapping rules, the system establishes a structured coordinate space for each resource dimension, which together constitute a unified seven-dimensional coordinate system, enabling any resource state to find a unique corresponding position in this coordinate space.

[0097] Step S23: Based on the seven-dimensional coordinate space, the physical meaning and value rules of each element in the tensor are determined through the element assignment function, thus obtaining the tensor element value definition rules. The element assignment function defines the specific meaning and numerical rules of each element in the tensor. In this embodiment, the value types of tensor elements are divided into three categories according to actual needs: binary, discrete, and continuous. Binary elements represent the availability status of resource points (0 represents unavailable, 1 represents available); discrete elements represent a finite level of resource quantity; and continuous elements represent a precise resource utilization rate or allocable ratio. Special element values ​​are also defined to represent specific resource states, such as -1 representing resource failure and 999 representing reserved resources. The element assignment function not only specifies the range and type of element values ​​but also defines the mapping relationship from physical meaning to numerical value. For example, in the computing power dimension, the element value may represent the CPU core allocation percentage; in the wavelength dimension, the element value may represent the occupancy status of wavelength channels; and in the time dimension, the element value may represent the pre-occupancy ratio of time slots. These element value definition rules constitute the semantic basis of tensor data content, transforming tensors from purely mathematical structures into resource representation models with physical meaning.

[0098] Step S24: Based on the seven-dimensional coordinate space and the tensor element value definition rules, the standardized multi-dimensional resource feature vector is filled into the corresponding coordinate positions using a tensor initialization algorithm to obtain the seven-dimensional resource tensor. The tensor initialization algorithm is the process of mapping standardized feature vector data to a seven-dimensional coordinate space. This algorithm adopts a hierarchical mapping strategy, first determining the tensor dimension and coordinate position corresponding to each element in the feature vector, then converting it into tensor element values ​​according to the element value definition rules, and finally filling it into the tensor structure. Due to the sparsity of the seven-dimensional space, a sparse tensor storage structure is adopted, storing only non-zero or non-default elements to save storage space. During initialization, a consistency check is also performed to ensure that the dependencies between tensor elements conform to physical constraints. For example, if a certain wavelength resource is unavailable, then all spectrum resources corresponding to that wavelength should also be marked as unavailable. Through this initialization algorithm, a resource tensor containing seven dimensions is successfully constructed. This tensor comprehensively and uniformly represents the state information of heterogeneous computing resources and optical network resources, providing a basic data structure for subsequent tensor operations.

[0099] Step S25: Based on the seven-dimensional resource tensor, the tensor subtraction operation is defined to represent the remaining resource state after the resource is allocated or occupied, thus obtaining the tensor subtraction operation rules. The tensor subtraction operation is a mathematical representation of the resource allocation process, defining the state transition rules after the resource is occupied. In this embodiment, the tensor subtraction operation is not a simple element-to-element subtraction, but a semantic subtraction based on the physical characteristics of the resource. For divisible resources (such as CPU computing power and memory capacity), the subtraction operation adopts conventional numerical subtraction; for indivisible resources (such as wavelength channels), the subtraction operation adopts a state transition model, that is, from the available state (1) to the occupied state (0); for resources with dependencies (such as wavelengths and their spectrum resources), the subtraction operation needs to consider the cascading effect to ensure resource consistency. The tensor subtraction operation also introduces a "resource reservation" mechanism, which allows marking specific resources as reserved states during the subtraction process to support the resource pre-allocation function. The subtraction operation rules also define a constraint checking mechanism in the operation process to ensure that the subtraction result does not produce physically impossible resource states (such as negative value resources). These carefully designed subtraction rules enable the system to accurately simulate state changes during resource allocation, providing a mathematical basis for resource scheduling decisions.

[0100] Step S26: Based on the seven-dimensional resource tensor, tensor addition is defined to represent the increased resource state after resource release or reclamation, thus obtaining tensor addition operation rules. Tensor addition operation corresponds to the resource release process and defines the state recovery rules after resource reclamation. Similar to subtraction, addition is also a specific operation based on resource semantics, rather than simple algebraic addition. For divisible resources, addition operation adds the released resource amount back to the available resource pool; for indivisible resources, addition operation restores the resource state from occupied (0) to available (1); for composite resources (such as an optical path containing multiple wavelengths), addition operation needs to handle the overall release of the resource combination. Tensor addition operation rules also include two modes: "partial release" and "full release," corresponding to partial and complete resource reclamation scenarios, respectively. In addition, addition operation rules define the priority of resource release and conflict handling mechanism to ensure the consistency of system state when multiple tasks release resources concurrently. Addition operation also includes resource state validity verification to ensure that the addition result does not exceed the physical limit of the resource (such as CPU utilization not exceeding 100%). This complete set of addition rules enables the system to accurately track state changes during resource release, providing the necessary mathematical tools for dynamic resource management.

[0101] Step S27: Based on the seven-dimensional resource tensor, tensor multiplication is defined to represent the connectivity and reachability between different resource nodes, thus obtaining the tensor multiplication operation rules. Tensor multiplication is the core mechanism for expressing the relationships between resources. Unlike traditional matrix multiplication, tensor multiplication here is an operation under a specific semantic meaning. In this embodiment, multiplication is mainly used in three scenarios: resource reachability calculation, resource combination utility evaluation, and resource path verification. For resource reachability, multiplication checks whether a valid connection path exists between two resource points; for resource combination, multiplication calculates the comprehensive utility of using multiple resource combinations; for path verification, multiplication checks whether all nodes on a resource allocation path satisfy the constraints. Tensor multiplication uses a "conditional propagation" mechanism, meaning the multiplication result is valid only when all nodes on the path satisfy the conditions. The multiplication operation rules also define the concept of "resource compatibility" to test the matching degree of different types of resources when used in combination. This specially defined multiplication operation enables the system to express and compute complex resource relationship networks in a unified tensor model, providing a mathematical foundation for resource path planning and multidimensional resource joint optimization.

[0102] Step S28: Based on the tensor subtraction, tensor addition, and tensor multiplication rules, a set of constraint conditions for tensor operations is obtained by defining non-negativity constraints, capacity limit constraints, and path continuity constraints that must be satisfied by tensor operations through constraint condition functions. The constraint condition function is a set of rules that ensures the results of tensor operations conform to physical laws and system limitations. The non-negativity constraint ensures that the remaining resources after resource allocation do not have negative values, reflecting the basic principle that physical resources cannot be over-allocated; the capacity limit constraint ensures that the total amount of resources after resource release does not exceed the physical limit, reflecting the objective fact that the total amount of resources is limited; the path continuity constraint requires that the continuity and integrity of the path be guaranteed in network path allocation, and no breakpoints are allowed. In addition to these three basic constraints, the constraint condition function also defines resource mutual exclusion constraints (some resources cannot be allocated to different tasks simultaneously), timing constraints (resource allocation must satisfy temporal order), and quality constraints (resource allocation must meet minimum service quality requirements). A hierarchical constraint verification mechanism is adopted, dividing the constraints into hard constraints (must be satisfied) and soft constraints (satisfy as much as possible), and designing a dedicated verification algorithm for each type of constraint. These constraints together constitute the safety boundary of tensor operations, ensuring that all resource operations are performed within physically feasible limits.

[0103] Step S29: Based on the tensor subtraction rules, tensor addition rules, tensor multiplication rules, and the set of tensor operation constraints, a complete tensor operation library containing subtraction, addition, multiplication, and constraint verification functions is implemented through software programming, resulting in the tensor operation system. The tensor operation system is a complete software implementation that transforms the aforementioned mathematical definitions into executable computer programs. In this embodiment, the operation library adopts a layered architecture. The bottom layer provides basic tensor data structures and atomic operations; the middle layer implements semantic subtraction, addition, and multiplication operations; and the top layer provides constraint verification and compound operation functions. To improve performance, the system designs a dedicated storage structure and parallel computing algorithm for sparse seven-dimensional tensors, supporting accelerated tensor operations on heterogeneous computing platforms such as GPUs. The operation library also implements "lazy computation" and "operation merging" optimizations to reduce intermediate result generation and storage overhead. To ensure correctness, the operation library has a built-in unit testing framework and correctness verification mechanism, which can automatically detect whether the operation results meet various constraints. Furthermore, the computation library provides a rich set of API interfaces, supports integration with external systems, and possesses runtime adaptive optimization capabilities, dynamically adjusting computation strategies based on problems of varying scales. This complete tensor computation system provides a powerful computing engine for subsequent resource state representation and scheduling optimization, and is a core technical component for achieving unified management of optical and computing resources.

[0104] Example 4

[0105] In this embodiment, the step of mapping real-time resource states to resource state tensors and task resource requirements to resource requirement tensors using a state mapping algorithm based on the tensor operation system, and then performing resource state transformations through tensor operations to obtain resource state tensors and corresponding transformation functions, includes:

[0106] Based on actual resource status data and the seven-dimensional coordinate space of the seven-dimensional resource tensor, a state mapping algorithm is used to map the CPU utilization, memory usage, storage input / output rate, wavelength occupancy, spectrum allocation status, path selection information, and time slice usage of computing nodes in the current system to specific element values ​​in the seven-dimensional coordinate space, thus obtaining the resource status tensor. The state mapping algorithm is the core mechanism for converting raw resource status data acquired through real-time system monitoring into a seven-dimensional tensor representation. In this embodiment, the state mapping adopts a "multi-source data fusion" strategy. First, raw status data is collected from different monitoring systems. Then, noise and outliers are eliminated through data preprocessing. Finally, coordinate mapping rules are applied to map the cleaned data to the seven-dimensional tensor space. For computing resource status, the system collects CPU utilization (core-level precision), memory usage (including physical memory and swap space), cache usage at all levels, and storage I / O performance metrics for each computing node. For optical network resource status, it collects wavelength occupancy tables (recording the allocation status of each wavelength channel), spectrum allocation maps (displaying fine-grained spectrum resource usage), optical path configuration information (all currently active optical paths and their attributes), and time slice allocation tables (displaying the resource usage of the time-division multiplexing system). The state mapping algorithm also supports unified processing of heterogeneous data sources, capable of handling state data from different monitoring systems, different data formats, and different acquisition periods. Through time alignment and data normalization techniques, it ensures that the final generated resource state tensor has a consistent time label and a unified numerical representation. This complex state mapping process enables the system to integrate resource state information originally scattered across different management platforms into a unified seven-dimensional tensor model, providing a data foundation for global resource optimization.

[0107] Based on the resource requirement description of the task or application, a requirement mapping algorithm quantifies and maps the task's requirements for computing power, memory capacity, storage bandwidth, number of wavelengths, spectrum resources, path length, and time window to the seven-dimensional coordinate space, resulting in the resource requirement tensor. The requirement mapping algorithm converts the task's resource requirement specifications into a resource requirement tensor compatible with the resource state tensor. In this embodiment, requirement mapping first needs to handle diverse requirement description formats, including explicit requirements (clearly specifying the quantity and type of resources) and implicit requirements (deriving the required resources based on task performance goals). Three main requirement description methods are supported: template-based descriptions (using predefined resource configuration templates), performance-based descriptions (specifying expected performance metrics), and history-based descriptions (referencing historical resource usage of similar tasks). The requirement mapping algorithm parses and standardizes these different forms of description, converting them into a unified set of resource requirement parameters. Then, the algorithm applies a specially designed requirement quantification model to quantify the task's requirements for different resource dimensions into specific numerical values. For example, "high CPU requirement" is quantified into a specific number of cores or computing power value, and "low latency communication requirement" is quantified into a specific number of wavelengths and path constraints. The demand mapping algorithm also addresses the uncertainty of resource requirements by introducing fuzzy demand representations and probabilistic demand models, providing possible value ranges and confidence levels for each demand parameter. Finally, the algorithm maps the quantified demand parameters to a seven-dimensional coordinate space, generating a tensor structure representing the task's resource requirements. This precise demand representation enables the system to compare resource requirements and availability within a unified mathematical framework, providing a basis for resource allocation decisions.

[0108] Based on the resource state tensor, the resource demand tensor, and the tensor subtraction rules, the new state after resource allocation is calculated through tensor subtraction. The set of tensor operation constraints is then used to verify whether each element in the new state satisfies the non-negativity constraint, resulting in the state tensor after resource allocation. This step implements a mathematical simulation of the resource allocation process. First, the resource demand tensor and the resource state tensor are subtracted to calculate the new state after resource allocation. Unlike ordinary numerical subtraction, this tensor subtraction follows the previously defined semantic subtraction rules, employing different subtraction strategies based on different resource types. During the calculation, the dependencies between resources are considered to ensure that the state changes of related resources remain consistent. For example, when a wavelength channel is allocated, the states of all spectral resources on that wavelength are updated accordingly. After calculating the new state, constraint checks are performed, primarily checking whether the non-negativity constraint is satisfied, ensuring that no resource dimension experiences over-allocation. The constraint verification employs a multi-level checking strategy: first, a rapid global check is performed to confirm the legality of all elements; then, key resource points are verified in detail to ensure the physical feasibility of state transitions; finally, a resource consistency check is performed to verify whether the states of related resources are coordinated. If any constraint violation is found, a detailed violation report is generated, along with possible solutions. Through this process, a new tensor representing the state after resource allocation is obtained. This result satisfies both resource requirements and physical constraints, providing a reliable basis for subsequent resource management decisions.

[0109] Based on the resource state tensor, the resource to be released tensor, and the tensor addition rules, the new state after resource release is calculated using tensor addition. The set of tensor operation constraints is then used to check whether each element in the new state exceeds the capacity limit constraint, resulting in the state tensor after resource release. This step simulates the state changes during resource release. The resource to be released tensor represents the resource to be returned to the resource pool. It is added to the current resource state tensor to calculate the new state after resource release. The addition operation follows the previously defined semantic addition rules, with different state recovery logics for different types of resources. For divisible resources (such as CPU computing power), the released resource amount is directly added back to the available pool; for indivisible resources (such as wavelength channels), the resource state is restored from occupied to available. The addition process also handles the side effects of resource release, such as the release of a primary resource potentially causing related subordinate resources to be released simultaneously. After the addition calculation is completed, a capacity limit constraint check is performed to ensure that no resource dimension exceeds its physical limit. This check is particularly important because in the case of concurrent resource release by multiple tasks, errors such as duplicate resource release may occur. The constraint verification employs a hierarchical validation strategy based on resource type: for discrete resources, it checks whether the maximum available quantity is exceeded; for continuous resources, it checks whether the utilization rate exceeds 100%; and for composite resources, it checks whether all components are within reasonable limits. If any violation of the upper limit constraint is found, the release process is automatically adjusted to ensure the legality of the final state. Through this process, a new tensor representing the state after resource release is obtained, providing an accurate state representation for dynamic resource management.

[0110] Based on the state tensors after resource allocation and resource release, the output state tensor after the input state tensor undergoes operations is defined through state transition functions, thus obtaining the resource state tensor and its corresponding transition function. State transition functions are a set of functions that mathematically describe the state changes caused by resource operations. In this embodiment, state transition functions include not only the basic subtraction and addition operations described earlier, but also more complex combination and conditional operations. A series of atomic transition functions are defined, such as the resource allocation function (Allocate), resource release function (Release), resource reservation function (Reserve), and resource migration function (Migrate). Each atomic function has clearly defined input parameters, preconditions, transition logic, and postconditions. Based on these atomic functions, composite transition functions are constructed to handle more complex resource operation scenarios, such as batch resource allocation, conditional release, and resource reconstruction. State transition functions also support transactional operations, meaning that multiple resource operations either all succeed or all rollback, ensuring system state consistency. To improve flexibility, parameterized transition functions are introduced, allowing dynamic adjustment of the transition logic based on external conditions. For example, resource allocation strategies can be adjusted based on system load status, or resource release priorities can be changed based on time factors. All these transformation functions together constitute a complete resource state transformation system, accurately describing the impact of various resource operations on the system state and providing a mathematical basis for subsequent resource scheduling decisions. In this way, resource state changes can be accurately tracked and predicted, enabling unified management and optimization of heterogeneous resources.

[0111] Example 5

[0112] In this embodiment, before obtaining the resource allocation tensor and the corresponding joint scheduling decision scheme through tensor space search and optimization based on the resource state tensor, the resource demand tensor, and the resource state tensor and its corresponding transformation function, the method further includes:

[0113] Based on the resource state tensor and its corresponding transformation function, and the state transition sequences recorded in the system evolution history, a sequence coding algorithm is used to combine the resource state tensor and operation sequences into a temporal decision sequence, resulting in a resource state transition sequence model. The sequence coding algorithm is a technique for organizing discrete resource states and operational behaviors into a coherent temporal decision sequence. In this embodiment, resource state records and corresponding operation records from a historical database over a past period (typically the last 24 hours to 7 days) are first extracted to form the original state-operation pair sequence. This original data contains the time series of the resource state tensor and the operation commands corresponding to each state change. The sequence coding algorithm employs a three-stage processing strategy: first, data cleaning is performed to remove abnormal states and invalid operations; then, time alignment is performed to ensure that all state-operation pairs have consistent timestamps; finally, semantic encoding is performed to convert the state tensor and operation commands into a unified encoding format. Semantic encoding is the core step of this algorithm. It uses feature compression and semantic embedding techniques to compress the high-dimensional resource state tensor into a low-dimensional feature vector and map complex operation commands into standardized operation codes. After this processing, each state-operation pair is represented as a combined code, and multiple codes are concatenated in chronological order to form a temporal decision sequence. This sequence not only records the evolution trajectory of the system state but also contains the decision information that triggers state transitions, providing training data for subsequent sequence transition models. The resource state transition sequence model is a mathematical model built upon these coded sequences. It can capture state change patterns and decision-making rules, providing historical experience references for resource scheduling.

[0114] Based on the tensor subtraction, addition, and multiplication operations in the tensor operation system, a set of allowed resource operation tags is defined using a tag set construction algorithm, resulting in a tag vocabulary. The tag set construction algorithm is a method for defining standardized representations for all legal resource operations in the system. In this embodiment, the tag vocabulary is a complete set of executable operations in the system, with each tag corresponding to a basic resource operation or a composite operation. The tag set construction algorithm first identifies all atomic operations supported by the system, including basic operation types such as resource allocation (Allocate), resource release (Release), resource migration (Migrate), and resource reservation (Reserve). Then, the algorithm defines a parameterized template for each operation, such as "Allocate(ResourceType, Amount, Target)" which represents allocating a specified type and quantity of resources to a target object. Based on these atomic operations, the algorithm further constructs composite operation tags to represent multi-step resource operation combinations, such as "AllocateAndRoute" which represents allocating computing resources and establishing a network route. During the label construction process, the algorithm also verifies the validity of each operation, ensuring its compatibility with the tensor operation system and enabling state transitions through tensor subtraction, addition, or multiplication. The final label vocabulary contains three types of labels: basic operation labels (operations corresponding to a single resource dimension), composite operation labels (cooperative operations involving multiple resource dimensions), and conditional operation labels (intelligent operations containing execution conditions). Each label has a clear syntactic definition, semantic interpretation, parameter specification, and state transition mapping, forming the basic vocabulary of the system's decision language. This standardized operation representation enables the system to express complex resource management decisions as formalized operation sequences, facilitating subsequent sequence generation and optimization.

[0115] Based on the resource state transition sequence model and the labeled vocabulary, a probability distribution for selecting the next operation in the current state is constructed using a conditional probability model, and the tensor operation constraint set is used as hard constraints to obtain a constrained sequence transition model. The constrained sequence transition model is a decision model combining conditional probability and constraints, used to predict the probability distribution of various possible next operations given the current system state. In this embodiment, the conditional probability model adopts an improved Markov decision process framework, using the resource state as the state space and the operation labels as the action space. First, state transition data is extracted from the resource state transition sequence model, including a set of triples of state-operation-next state. Then, this data is used to train the conditional probability model, learning the mapping relationship between states and operations. The conditional probability model adopts a multi-layered architecture: the bottom layer is a statistical model based on historical frequency, calculating the frequency of each operation in historical data; the middle layer is a context-based association model, considering the contextual information of the current state; and the top layer is an evaluation model based on resource utilization efficiency, evaluating the impact of each operation on the overall system efficiency. These three layers of models are combined to generate a preliminary operation probability distribution. Then, a set of tensor operation constraints is applied to this initial distribution to eliminate operation options that violate hard constraints, and the probability values ​​of the remaining operations are adjusted accordingly. Hard constraints include resource non-negativity constraints, capacity cap constraints, and path continuity constraints, ensuring that the operation sequences predicted by the model are physically feasible. Constraint processing employs a constraint propagation algorithm, iteratively adjusting the operation probabilities until all constraints are satisfied. The resulting constrained sequence transition model can generate an operation probability distribution that conforms to physical constraints for any given system state, providing guidance for subsequent sequence generation.

[0116] Based on the constrained sequence transition model, the resource state tensor, and the resource demand tensor, a sequence generation algorithm generates an operation sequence within the operation space of the labeled vocabulary that satisfies the resource demand tensor and maximizes the state transition probability, thus obtaining a resource allocation operation sequence plan. The sequence generation algorithm is an algorithm that generates the optimal operation sequence based on the current state and the target demand. In this embodiment, sequence generation employs a deep reinforcement learning method based on Monte Carlo Tree Search (MCTS). This algorithm models the resource allocation problem as a multi-step decision-making process: starting from the initial state (current resource state tensor), through a series of operations (selected from the labeled vocabulary), the target state (satisfying the resource demand tensor) is finally reached. The core of the algorithm is the construction and exploration of a decision tree: each node of the tree represents a resource state, and each edge represents an operation, expanding from the root node (current state) towards the target state. The expansion process consists of four phases: Selection phase, starting from the root node, selects the most promising path based on the UCB (Upper Confidence Bound) value; Expansion phase, adds a new node at the end of the selected path, representing a new state transition; Simulation phase, randomly executes operations starting from the new node until a termination condition is met; and Backtracking phase, updates the value assessments of all nodes on the path. In the expansion and simulation phases, the algorithm uses a constrained sequence transition model as prior knowledge to guide operation selection. Simultaneously, the algorithm checks whether each operation violates resource constraints and calculates the matching degree between the state and the target requirement as the basis for node evaluation. Through multiple iterations, the algorithm gradually builds and optimizes the decision tree, ultimately selecting the path with the highest value as the operation sequence plan. This plan not only meets resource requirements but also has a high transition probability, indicating that it aligns with the system's historical best practices and possesses reliability and predictability.

[0117] Based on the resource allocation operation sequence plan, the resource state tensor, and the corresponding transformation function, the intermediate state tensors corresponding to each step of the operation are calculated sequentially using the sequence-tensor mapping function, resulting in an intermediate state tensor sequence and a final resource allocation state tensor. The sequence-tensor mapping function is a mathematical mapping that converts an operation sequence into a specific state transition process. In this embodiment, the function receives the resource allocation operation sequence plan and the initial resource state tensor as input and outputs a state tensor sequence containing all intermediate states. The mapping process adopts a sequential execution mode: starting from the initial state, each operation in the sequence is applied one by one, calculating the new state after the operation. For each operation flag, according to its type and parameters, the corresponding transformation function is called to perform specific tensor operations. For example, for the allocation operation, a resource allocation function based on tensor subtraction is called; for the release operation, a resource release function based on tensor addition is called; for the path establishment operation, a connectivity verification function based on tensor multiplication is called. After each operation, a complete constraint check is performed to ensure that the intermediate state satisfies all physical constraints. If any constraint violation is found, a rollback mechanism is triggered, restoring the previous valid state and attempting an alternative operation. The entire mapping process employs a transaction management mechanism to ensure that either all operations are executed completely and the final state is reached, or a complete rollback to the initial state is achieved in the event of an unrecoverable error. This rigorous mapping process yields a complete state transition trajectory, including the initial state, each intermediate state, and the final resource allocation state. This detailed state sequence not only illustrates the execution path of resource allocation but also provides crucial information for subsequent decision-making and monitoring.

[0118] Example 6

[0119] In this embodiment, the step of obtaining the resource allocation tensor and the corresponding joint scheduling decision scheme based on the resource state tensor, the resource demand tensor, and the resource state tensor and its corresponding transformation function through tensor space search and optimization includes:

[0120] Based on the resource demand tensor and the resource state tensor, the matching degree between task requirements and available system resources in the dimensions of computing power, memory, storage, wavelength, spectrum, path, and time is calculated using a tensor matching degree calculation function, resulting in a set of resource matching degree indicators. The tensor matching degree calculation function is a mathematical tool for evaluating the fit between task requirements and available system resources. In this embodiment, the function designs a specific matching degree calculation method for each dimension of the seven-dimensional resource tensor. For the computing power dimension, the matching degree is based on the ratio of required computing power to available computing power, while also considering the compatibility of computing power types; for the memory dimension, the matching degree comprehensively considers capacity matching and access speed matching; for the storage dimension, the matching degree is based on the dual indicators of storage capacity and I / O bandwidth; for the wavelength dimension, the matching degree reflects whether the required number of wavelengths can be met and whether the wavelength quality meets the requirements; for the spectrum dimension, the matching degree evaluates the continuity and bandwidth sufficiency of spectrum resources; for the path dimension, the matching degree considers whether available paths meet connection requirements and service quality requirements; for the time dimension, the matching degree reflects the degree of overlap of time windows and scheduling flexibility. The calculation process also considers the correlation between resources, such as the collaborative matching of computing nodes and network connections. For each dimension, the matching degree calculation function generates a standardized matching score between 0 and 1, where 1 represents a perfect match and 0 represents a complete mismatch. These scores together form a resource matching degree index set, comprehensively reflecting the fit between tasks and resources across various dimensions. This index set not only describes the static matching degree but also includes dynamic adaptive assessments, predicting the potential impact of resource state changes on the matching degree.

[0121] Based on the resource matching index set and system optimization objectives, weights are assigned to each dimension of the matching index through weight allocation. A multi-objective optimization function is constructed, including a weighted summation term for matching degree, a latency penalty term, and a cost penalty term, resulting in the optimization objective function. In this embodiment, the system optimization objectives typically include maximizing resource utilization, minimizing task completion time, and minimizing energy consumption. The weight allocation process employs an adaptive weighting mechanism, dynamically adjusting the importance weights of each dimension based on the current system state and optimization preferences. For example, when system computing resources are strained, the weight of the computing power dimension will increase accordingly; when communication latency requirements are high, the weights of the wavelength and path dimensions will increase. The weight allocation uses a combination of the analytic hierarchy process (AHP) and historical performance feedback to ensure that the weight settings both conform to management strategies and adapt to real-time needs. The construction of the multi-objective optimization function consists of three main components: a weighted summation term that integrates the matching degree of each dimension, representing the basic suitability of the resource allocation scheme; a delay penalty term that measures the task execution delay and communication delay caused by resource allocation, where task execution delay is related to computational resource allocation and communication delay is related to network resource allocation; and a cost penalty term that considers the economic cost and energy consumption of computational resources, including computational cost, storage cost, bandwidth cost, and energy cost. These three components are combined through weighted summation to form the final optimization objective function. This function transforms the complex multi-dimensional optimization objective into a single numerical evaluation standard, providing a clear optimization direction for subsequent search algorithms. The design of the optimization objective function fully considers the trade-offs and mutual influences between different objectives, enabling it to find a balance point under various constraints.

[0122] Based on the optimization objective function and the resource state tensor, a causal discovery algorithm is used to analyze the causal dependencies between the resource state variables of each dimension in the resource state tensor and the weighted summation term, delay penalty term, and cost penalty term in the optimization objective function. A causal relationship graph is constructed, and an intervention operator is defined to represent the operation of setting the numerical values ​​of the resource dimension variables in the causal relationship graph. By calculating the difference between the first function value of the optimization objective function before the intervention operator is applied and the second function value of the optimization objective function after the intervention operator is applied, a causal effect matrix is ​​obtained. The causal discovery algorithm is an algorithm for identifying causal relationships between variables. Unlike traditional correlation analysis, it can reveal the causal mechanism of "changes in variable A lead to changes in variable B". In this embodiment, the causal discovery algorithm adopts a structure learning method based on the PC algorithm, combined with domain knowledge constraints to infer causal structure. The algorithm first extracts the time series of resource state variables and each component of the objective function from historical data, and then constructs a preliminary causal skeleton through conditional independence testing. Next, the algorithm applies direction determination rules to determine the direction of the edges, and at the same time, combines prior knowledge of resource physical characteristics to correct the direction. The resulting causal graph is a directed acyclic graph (DAG), where nodes represent resource state variables and components of the objective function, edges represent causal relationships, and edge weights represent causal strength. Intervention operators are a core concept in causal inference, representing the artificial setting of a variable's value and observing its impact on other variables. In this system, intervention operators correspond to operations that make allocation decisions for specific resource dimensions. By calculating the changes in the objective function value before and after intervention, a causal effect matrix is ​​constructed. Each element (i, j) in this matrix represents the degree of influence of intervening in resource dimension i on component j of the objective function. This causal analysis method surpasses traditional correlation analysis, identifying "which resource decisions truly affect system performance," providing more effective guidance for subsequent optimization.

[0123] Based on the causal effect matrix and the optimization objective function, the resource dimension variable with the largest causal effect value is selected as the priority intervention target through the causal intervention optimization algorithm. A search strategy based on causal effect ranking in the seven-dimensional coordinate space is designed to obtain a causal relationship-based search strategy. The causal intervention optimization algorithm is a decision optimization method based on causal reasoning, which uses causal effect evaluation to guide the exploration of the search space. In this embodiment, the algorithm first analyzes the causal effect matrix and calculates the cumulative causal effect value of each resource dimension, that is, the sum of the influence of this dimension on all components of the objective function. Then, the resource dimensions are ranked according to the cumulative effect value to determine the intervention priority. Two factors are considered in the ranking: effect strength (the magnitude of the influence of the intervention on the objective function) and effect certainty (the predictability of the intervention result). Based on this ranking, the algorithm designs a multi-stage search strategy: first optimize the dimension with the largest causal effect, fix this dimension, then optimize the secondary dimensions, and so on. This strategy significantly reduces the dimensions of the search space and improves search efficiency. For different types of resource dimensions, the algorithm employs different search methods: enumeration search for discrete dimensions (such as wavelength selection); gradient descent for continuous dimensions (such as computing power allocation); and heuristic search for combined dimensions (such as path selection). During the search process, the algorithm dynamically updates the causal model, adjusting the causal effect evaluation of subsequent dimensions based on the decisions already made. This causal-based search strategy not only improves optimization efficiency but also enhances the interpretability of decisions, clearly explaining why specific resource allocation decisions were made.

[0124] Based on the causal search strategy, the resource state tensor, and the resource demand tensor, a tensor space search algorithm is used to find the resource allocation tensor that maximizes the value of the objective function while satisfying the resource non-negativity constraint. The tensor space search algorithm is a search method for finding the optimal resource allocation scheme in a multi-dimensional tensor space. In this embodiment, the algorithm uses an improved branch-and-bound method combined with simulated annealing to efficiently search in a seven-dimensional resource tensor space. The algorithm first determines the dimension processing order based on the causal search strategy, and then processes each dimension sequentially according to priority. For each dimension, the algorithm sets the search range and step size, constructing a candidate value set for that dimension. The search process uses a tree structure, where each node represents a partial decision (the value of a partial dimension has been determined), and branches represent candidate decision options. To improve efficiency, the algorithm applies various pruning strategies: feasibility pruning removes branches that violate resource constraints; upper bound pruning removes branches that cannot be better than the current optimal solution; and similarity pruning merges branches with similar decision effects. During the exploration process, the algorithm dynamically balances exploration and utilization, adjusting the search temperature through simulated annealing. Initially, it favors extensive exploration, while later it focuses on local optimization. To handle large-scale problems, the algorithm employs a hierarchical search strategy, first determining the general direction in a coarse-grained space and then precisely locating the target in a fine-grained space. At each search step, the algorithm verifies the non-negativity constraints of resources to ensure the allocation scheme is physically feasible. Finally, the algorithm outputs a resource allocation tensor that satisfies all constraints and maximizes the objective function value. This tensor precisely defines the allocation scheme for resources in each dimension. This tensor-space-based search method can simultaneously optimize the allocation of computational and network resources, overcoming the limitations of traditional hierarchical allocation methods.

[0125] Based on the resource allocation tensor, the element values ​​of each dimension of the allocation tensor in the tensor space are converted into CPU core allocation instructions, memory block allocation instructions, storage channel allocation instructions, wavelength allocation instructions, spectrum slot allocation instructions, routing path selection instructions, and time slice scheduling instructions through a decision mapping function, thus obtaining the joint scheduling decision scheme. The decision mapping function is a conversion mechanism that transforms an abstract tensor space representation into specific system-executable instructions. In this embodiment, the function has designed specific mapping rules for each of the seven resource dimensions. For the computing power dimension, the mapping function generates CPU / GPU core allocation instructions, specifying the specific processor ID, number of cores, priority, and affinity settings. For the memory dimension, it generates memory block allocation instructions, specifying the memory region, size, access permissions, and NUMA node binding. For the storage dimension, it generates storage channel allocation instructions, including storage device selection, IO bandwidth limits, and caching strategies. For the wavelength dimension, it generates optical network wavelength allocation instructions, specifying the wavelength channels used and the configuration of optical transmitting / receiving devices. For the spectrum dimension, it generates spectrum slot allocation instructions, detailing the frequency range and modulation format. For the path dimension, it generates routing path selection instructions, including complete end-to-end paths and intermediate node configurations. For the time dimension, it generates time-slice scheduling instructions, defining the time window for resource usage and scheduling priority. During the mapping process, the characteristics of different devices and platforms are considered, generating instruction sets optimized for specific hardware environments. Simultaneously, inter-instruction dependency analysis is performed to ensure that the generated instruction sequences are coordinated in execution order. The final generated joint scheduling decision scheme is a complete, directly executable set of instructions, containing collaborative configuration information for computing and network resources. This end-to-end decision mapping mechanism ensures that the optimization results in the tensor space can be accurately transformed into actual system operations, achieving a seamless connection between theory and practice.

[0126] Example 7

[0127] In this embodiment, the step of finding the resource allocation tensor that maximizes the value of the optimization objective function based on the causal search strategy, the resource state tensor, and the resource demand tensor using a tensor space search algorithm under the premise of satisfying the resource non-negativity constraint, and obtaining the resource allocation tensor, includes:

[0128] Based on the causal effect ranking results in the causal relationship-based search strategy and the resource state tensor, intervention operations are sequentially performed on multiple resource dimension variables with the largest causal effect values ​​through causal intervention tensor space search, generating multiple sets of candidate intervention operations and corresponding resource allocation tensor sets, and obtaining a set of candidate resource allocation schemes.

[0129] Causal intervention tensor space search is an advanced search method that combines causal reasoning with tensor space search. In this embodiment, the method determines the optimization order of resource dimensions based on the "causal effect ranking result". The system first selects the resource dimension with the largest causal effect value (usually the dimension with the most significant impact on the optimization objective) as the primary intervention target. For the selected dimension, a series of intervention operations are designed, each representing a possible allocation method for the resource in that dimension. These intervention operations are designed based on the physical characteristics of the resource and historical best practices to ensure that each operation is physically feasible. When executing intervention operations, the "do-calculus" framework is used to accurately calculate the impact of the intervention on the system state. Unlike the traditional "hypothesis-verification" method, do-calculus can accurately simulate the causal effect of "how we would react if we set resource X to value Y". The system performs multiple intervention operations on each primary dimension, then fixes the optimal result and continues to process secondary dimensions. This progressive intervention process forms a decision tree structure, where each path in the tree represents a possible resource allocation scheme. During the generation of candidate schemes, parallel intervention evaluation is performed, exploring different combinations of multiple dimensions simultaneously to avoid getting trapped in local optima. The effect of each intervention is evaluated by optimizing the objective function, while simultaneously validating the resource constraints. Ultimately, a set of high-quality candidate resource allocation schemes is generated, each corresponding to a specific resource allocation tensor representing a specific allocation decision across seven dimensions of resources. This causal intervention-based search method can efficiently explore complex decision spaces and find candidate schemes that meet physical constraints and offer high optimization objective values.

[0130] Based on the set of candidate resource allocation schemes and the set of historically optimal resource allocation tensors stored in historical scheduling experience, the Euclidean distance between the candidate scheme and the historically optimal scheme is calculated using a decision boundary alignment algorithm. The historically optimal scheme with the smallest distance is selected as the reference scheme, and the allocation values ​​of each dimension of the candidate scheme are adjusted so that the allocation values ​​of each dimension of the candidate scheme are close to the corresponding allocation values ​​of the reference scheme, thus obtaining the resource allocation tensor.

[0131] The decision boundary alignment algorithm is a scheme optimization method that combines historical experience and current computation. In this embodiment, the algorithm first extracts historically optimal resource allocation cases from the system database. These cases represent resource allocation decisions that performed exceptionally well in past system operations and have been verified to have good results in real-world environments. The algorithm uses multidimensional Euclidean distance to measure the similarity between the current candidate scheme and the historical optimal scheme. The distance calculation considers different weights and scales for each dimension. Similarity calculation is based on two aspects: resource demand characteristics and system state characteristics. Resource demand characteristics compare the resource demand patterns of the current task with those of historical tasks; system state characteristics compare the similarity between the current system environment and historical environments. Through comprehensive evaluation, the algorithm selects the historical optimal scheme that is most similar to the current situation as a reference benchmark. Then, the algorithm applies the "decision boundary alignment" technique to bring the decision boundary of the candidate scheme closer to the reference scheme. This process is not a simple replication but a selective adjustment: for dimensions with high certainty (such as when resource capacity is clear), the current computation results are maintained; for dimensions with high uncertainty (such as when performance prediction is difficult to be accurate), more historical experience is referenced. The adjustment process uses a weighted average method, with weights based on the predictive certainty and historical similarity of each dimension. The resulting resource allocation tensor combines the advantages of theoretical calculation and practical experience: it retains the main decision-making direction based on causal analysis while drawing on successful experiences validated in the past. This hybrid decision-making approach of "computation + experience" significantly improves the reliability and performance stability of resource allocation, avoiding practical application deviations that may occur with purely theoretical models.

[0132] Example 8

[0133] In this embodiment, obtaining resource bottleneck reports and resource conflict early warning reports through historical resource state tensor sequence analysis includes:

[0134] Based on resource status data recorded during system operation, a multi-scale sampling strategy is used to extract historical resource status tensors at three time granularities: second, minute, and hour, resulting in a multi-granularity historical information set. The multi-scale sampling strategy is a data acquisition method for multi-granularity extraction of time-series data. In this embodiment, the strategy designs differentiated sampling schemes at the three time scales for different resource types and analysis objectives. For second-level sampling, a high-frequency acquisition method is used, collecting instantaneous status of computing resources and rapid changes in optical network resources at 1-5 second intervals, focusing on capturing short-term fluctuation patterns and sudden events in resource usage. Second-level sampling primarily targets high-rate-of-change indicators such as CPU utilization, memory access frequency, and network packet transmission rate. For minute-level sampling, the system uses a medium-frequency acquisition method, aggregating the average utilization of computing resources and changes in optical network link status at 1-5 minute intervals, focusing on medium-term trends and periodic patterns in resource usage. Minute-level sampling primarily targets medium-rate-of-change indicators such as memory usage, storage I / O throughput, and wavelength allocation status. For hourly sampling, a low-frequency acquisition method is adopted, statistically calculating the long-term usage of resources and changes in the optical network topology at 1-6 hour intervals, focusing on analyzing the seasonal changes and growth trends of resource usage over the long term. Hourly sampling mainly targets indicators with low rate of change, such as storage capacity growth, spectrum resource allocation patterns, and path selection preferences. During the sampling process, the raw data undergoes preliminary processing, including outlier filtering, missing value imputation, and time alignment, to ensure the quality and consistency of multi-scale data. Through this multi-scale sampling strategy, an information set containing historical resource state tensors at different time granularities is constructed, providing a rich and comprehensive data foundation for subsequent multi-level analysis.

[0135] Based on the aforementioned multi-granularity historical information set, a three-layer encoding network encodes second-level historical information into primary features, fuses minute-level historical information with primary features into intermediate features, and fuses hour-level historical information with intermediate features into high-level features, resulting in a progressive historical encoding representation. The three-layer encoding network is a neural network architecture for processing multi-timescale data, effectively integrating multi-granularity information through progressive encoding and feature fusion. In this embodiment, the network adopts a stacked encoding structure, with each layer responsible for processing information at a specific time granularity and fusing the encoding results of the previous layer. The first encoder layer processes second-level historical information, using a convolutional neural network (CNN) structure to extract local feature patterns in the time-series data, such as sudden peaks, periodic fluctuations, and short-term correlations in resource usage. The CNN captures different features across seven resource dimensions through multi-channel convolutional layers, reduces feature dimensions through pooling layers, and extracts salient features, ultimately outputting a primary feature vector that mainly contains the short-term dynamic characteristics of resource usage. The second encoder layer processes minute-level historical information and simultaneously receives the primary features output from the first layer. This layer employs a Bidirectional Long Short-Term Memory (BiLSTM) network structure, capable of simultaneously considering historical and future information to capture mid-term trends in resource use and inter-dimensional dependencies. BiLSTM selectively fuses minute-level information with primary features through an attention mechanism, reinforcing important time points and resource dimensions, and outputting a mid-level feature vector. This vector primarily contains mid-term trends in resource use and dimensional correlations. The third encoder layer processes hourly historical information while simultaneously receiving mid-level features from the second layer. This layer uses a Transformer structure, leveraging its powerful long-distance dependency modeling capabilities to capture long-term patterns and global dependencies in resource use. The Transformer processes hourly time-series data through a multi-head self-attention mechanism and fuses mid-level features through a cross-modal attention mechanism, outputting a high-level feature vector. This vector primarily contains long-term trends, cyclical patterns, and global dependency structures in resource use. This progressive encoding architecture effectively integrates information from different time scales, forming a multi-granular, multi-layered representation of resource use, providing rich feature representations for subsequent bottleneck analysis and early warning prediction.

[0136] Based on the seven-dimensional coordinate space of the seven-dimensional resource tensor and historical resource usage patterns, a key information extraction algorithm is used to define a representative query anchor point in each of the dimensions of computing power, memory, storage, wavelength, spectrum, path, and time, resulting in a set of query anchor points. The key information extraction algorithm is a method for locating key monitoring points in a high-dimensional resource space, achieving efficient monitoring of complex resource spaces by setting representative query anchor points. In this embodiment, the query anchor point is a specific coordinate point in the seven-dimensional coordinate space, representing the resource status that the system needs to focus on. The algorithm first analyzes historical resource usage data to identify key feature points in each dimension, including: high utilization areas (areas where resources are frequently under high load), high volatility areas (areas where resource status changes frequently), bottleneck history areas (areas where resource bottlenecks have occurred in the past), and key dependency areas (areas that have a strong impact on other resource dimensions). Then, the algorithm applies clustering and anomaly detection methods to select the most representative point in each dimension as the query anchor point for that dimension. Specifically, for the computing power dimension, anchors are typically set in high-computing-power areas required by CPU-intensive applications; for the memory dimension, anchors are set in the working areas of memory-sensitive applications; for the storage dimension, anchors are set in critical areas of I / O-intensive operations; for the wavelength dimension, anchors are set in the wavelength areas used by core communication links; for the spectrum dimension, anchors are set in the spectrum areas required for high-quality transmission; for the path dimension, anchors are set on critical connection paths; and for the time dimension, anchors are set during peak resource demand periods. Each anchor not only contains specific coordinate values ​​but also includes the query range and the attributes of interest, forming a structured set of query anchors. These anchors, acting as "sentinels" of the resource space, can efficiently monitor the most critical and problematic resource points in the system, providing a precise query focus for subsequent attention mechanisms.

[0137] Based on the set of query anchors and the progressive historical encoding representation, an attention weight between each query anchor and historical encoding features is calculated using a parallel attention mechanism. The historical encoding features are then weighted and summed to obtain the context vector for each query anchor, resulting in the anchoring inference result. The parallel attention mechanism is a multi-head attention computation framework capable of simultaneously processing the correlation analysis between multiple query anchors and historical features. In this embodiment, the mechanism constructs a dedicated attention head for each query anchor, enabling parallel processing of seven-dimensional anchors. For each query anchor, it is first converted into a query vector, and the features in the progressive historical encoding representation are converted into key and value vectors. Attention computation employs a scaled dot product attention method, calculating the similarity between the query vector and all key vectors to obtain an attention score. These scores are normalized to weights using a softmax function, representing the degree of correlation between the query anchor and historical features at different time points. Then, these weights are used to weight and sum the value vectors to generate the context vector for that query anchor. This process achieves the goal of extracting the most relevant information for a specific anchor from a large number of historical features. To enhance expressive power, multiple attention heads are configured for each anchor point, each focusing on different types of patterns, such as short-term fluctuations, long-term trends, and periodic changes. These multi-head results are merged into a final context vector through concatenation and linear transformation. The seven anchor points are processed simultaneously by seven sets of parallel attention heads, forming seven context vectors. These vectors are further integrated through a cross-dimensional attention layer to capture the inter-dimensional interactions, ultimately forming the anchored inference result. This anchor-based parallel attention mechanism enables efficient and targeted information extraction from historical data, providing accurate feature representations for subsequent bottleneck identification and conflict prediction.

[0138] Based on the anchored inference results and the resource state tensor, an enhanced utilization analysis function calculates the ratio of used resources to total available resources in each dimension and combines it with historical trend data to obtain a resource utilization tensor containing utilization values ​​for each of the seven dimensions. The utilization values ​​for each dimension are extracted from this resource utilization tensor, and a bottleneck identification algorithm selects resource dimensions with utilization values ​​exceeding a preset threshold as bottleneck points, resulting in a resource bottleneck report containing bottleneck dimension identification, bottleneck severity, and historical trend analysis. The enhanced utilization analysis function is a resource utilization assessment method that comprehensively considers both the current state and historical trends. In this embodiment, the function not only calculates static utilization but also analyzes dynamic change characteristics, providing a more comprehensive resource state assessment. For each resource dimension, the function first extracts the allocated resources and total available resources from the current resource state tensor to calculate the basic utilization rate. Then, based on the different characteristics of the resource type, specialized calculation methods are applied: for divisible resources (such as CPU and memory), linear proportional calculation is used; for discrete resources (such as wavelength channels), occupancy proportional calculation is used; and for combined resources (such as paths), weighted combined calculation is used. After calculating the base utilization rate, the function incorporates historical trend correction, using historical information from anchored inference results to adjust the current utilization rate assessment. The correction considers three key factors: growth rate (the rate of change in resource utilization), volatility (the degree of instability in utilization), and persistence (the duration of high utilization). These corrections result in a more predictive "effective utilization rate" that reflects the actual pressure level on resources. The seven dimensions of effective utilization together form a resource utilization tensor, comprehensively describing the system's resource status. The bottleneck identification algorithm uses this tensor to detect bottleneck points, employing an adaptive threshold method with different thresholds set for different resource types. For example, CPU utilization exceeding 85%, memory utilization exceeding 90%, and wavelength utilization exceeding 95% may be identified as bottlenecks. The algorithm also considers the severity, duration, and development trend of bottlenecks, rating and classifying each identified bottleneck point. The final resource bottleneck report contains three parts: bottleneck dimension identification (which resources are bottlenecked), bottleneck severity assessment (the severity and urgency of the bottleneck), and historical trend analysis (the bottleneck formation process and future development predictions). This enhanced utilization analysis method can identify potential bottlenecks in a timely manner, providing precise guidance for system optimization.

[0139] Based on the progressive historical encoding representation and the anchored inference results, a multi-timescale prediction network is used to construct a short-term prediction model to predict the resource status for the next five minutes, a medium-term prediction model to predict the resource status for the next one hour, and a long-term prediction model to predict the resource status for the next twenty-four hours. Weighted fusion is then used to generate future resource status predictions, resulting in a predicted resource status tensor set. The multi-timescale prediction network is a hierarchical prediction architecture designed for different prediction durations, capable of simultaneously meeting the needs of accurate short-term prediction and long-term trend prediction. In this embodiment, the network consists of three specialized prediction models, each optimized for different timescales. The short-term prediction model focuses on the resource status for the next five minutes, employing a Long Short-Term Memory (LSTM) network structure, which is particularly suitable for capturing subtle patterns in short-term time series. This model takes second-level sampled data and primary features as input, learning the temporal dependence and abrupt change patterns of resource use in the short term through multi-layer LSTM units. The model emphasizes prediction accuracy, employing a high sampling density and a low level of abstraction, enabling accurate prediction of short-term resource fluctuations. The medium-term prediction model focuses on the resource status for the next one hour, employing a gated recurrent unit (GRU) network structure, balancing computational efficiency and memory capacity. This model takes minute-level sampled data and intermediate features as input, capturing medium-term resource usage trends and cyclical patterns through a bidirectional GRU layer. It emphasizes prediction stability and trend accuracy, effectively filtering short-term noise and predicting the overall trend of resource usage. The long-term prediction model focuses on resource status over the next 24 hours, employing a Transformer-based attention network structure with strong long-sequence modeling capabilities. This model takes hour-level sampled data and high-level features as input, learning long-term seasonal patterns and global dependencies of resource usage through a multi-layer self-attention mechanism. It emphasizes grasping macro trends and warning of abnormal patterns; although its accuracy in specific numerical values ​​is lower, it accurately predicts long-term trends in resource usage. The outputs of these three prediction models are integrated using a weighted fusion algorithm to generate resource status predictions at different time points. During fusion, short-term predictions mainly rely on the short-term model, medium-term predictions consider both short-term and medium-term models, and long-term predictions mainly rely on the long-term model, achieving a smooth transition between predictions at different time scales. The final predicted resource status tensor set contains seven-dimensional resource status predictions for different time points within the next 24 hours, providing comprehensive prediction data for subsequent conflict detection.

[0140] Based on the predicted resource state tensor set and the known future task demand tensor set, a conflict detection function is used to check whether any dimension of the difference between the predicted resource state and the future task demand is negative. If so, it is marked as a resource allocation conflict, resulting in a resource conflict warning report containing the time of conflict occurrence, the dimension of the conflicting resource, and the severity of the conflict. The conflict detection function is a predictive resource conflict analysis tool that identifies potential resource shortages by comparing future resource states with task demands. In this embodiment, the function receives two key inputs: a predicted resource state tensor set (the system's predicted future resource availability state) and a future task demand tensor set (known or planned future task resource demands). The conflict detection process uses a sliding time window approach, analyzing each time point within the next 24 hours. For each time point, the function first extracts the corresponding predicted resource state tensor and task demand tensor, and then applies tensor subtraction to simulate the resource allocation process. The subtraction result represents the remaining resource state after allocation. If any negative element exists in the result, it indicates a conflict where demand exceeds availability in that resource dimension. Conflict detection not only focuses on the existence of conflicts but also analyzes the specific characteristics of the conflicts. The function calculates the severity of the conflict (the ratio of the absolute value of the negative value to the total amount of resources), duration (the length of time the conflict lasts), and scope of impact (the number of resource dimensions affected). To improve detection reliability, the function also considers prediction uncertainty, calculates confidence intervals for each prediction time point, and makes conservative estimates under boundary conditions to avoid false alarms and false negatives. For detected conflicts, the system generates a structured resource conflict early warning report containing three core pieces of information: the time of conflict occurrence (when resource shortages are expected to occur), the conflicting resource dimensions (which types of resources will be insufficient), and the severity of the conflict (the degree and impact of resource shortages). The report also includes conflict trend analysis (how the conflict evolves over time) and mitigation suggestions (possible resource adjustment schemes). This predictive conflict detection mechanism enables the system to identify potential resource problems in advance, providing managers with sufficient response time to avoid actual conflicts by pre-adjusting resource allocation or scheduling plans.

[0141] Example 9

[0142] In this embodiment, based on the progressive historical encoding representation and the anchored inference results, a short-term prediction model is constructed using a multi-timescale prediction network to predict the resource status for the next five minutes, a medium-term prediction model to predict the resource status for the next one hour, and a long-term prediction model to predict the resource status for the next twenty-four hours. A weighted fusion is then used to generate a future resource status prediction, resulting in a predicted resource status tensor set, including:

[0143] Based on the progressive historical encoding representation and the anchored inference results, a long short-term memory network is constructed as a short-term prediction model by inputting the primary features and the anchored inference results into a temporal decomposition network, a gated recurrent unit network is constructed as a medium-term prediction model by inputting the intermediate features and the anchored inference results into a gated recurrent unit network, and an attention mechanism network is constructed as a long-term prediction model by inputting the high-level features and the anchored inference results into a high-level feature network, thus obtaining a multi-scale prediction model set.

[0144] Temporal decomposition networks are neural network architectures that decompose time series data into components of different frequencies, enabling more efficient processing of multi-scale time series data. In this embodiment, the network first performs temporal decomposition on the progressively encoded historical representation, separating trend components, periodic components, and residual components. The trend component represents the long-term trend of resource use, the periodic component represents the cyclical pattern of resource use, and the residual component represents random fluctuations and anomalous events. The decomposition process uses wavelet transform implemented by a neural network, which can adaptively capture time patterns of different frequencies. The decomposed components, along with the anchored inference results, are input into three specialized prediction models. The short-term prediction model is built on a Long Short-Term Memory (LSTM) network and is specifically designed to handle high-frequency time series data. This model receives primary features (containing second-level resource status information) and anchored inference results as input. The core advantage of LSTM lies in its gating mechanism, which can selectively remember and forget historical information, making it particularly suitable for capturing subtle change patterns in the short term. The model adopts a stacked architecture, with lower-level LSTM units capturing basic time series patterns and higher-level LSTM units integrating lower-level features to form a more abstract representation. The output layer maps the LSTM state to resource state predictions for the next five minutes through a fully connected network. The intermediate-term prediction model is built on a Gated Recurrent Unit (GRU) network, balancing model complexity and expressive power. This model receives intermediate-level features (containing minute-level resource state information) and anchored inference results as input. GRU achieves long-term dependency learning through a simpler gating mechanism, with higher computational efficiency than LSTM, making it particularly suitable for prediction tasks of medium-length sequences. The model employs a bidirectional architecture, considering both past and future contextual information, improving prediction stability. The attention layer helps the model focus on the most relevant parts of historical data, enhancing the model's interpretability. The output layer, through a residual connection mechanism, combines the original input and GRU output to generate resource state predictions for the next hour. The long-term prediction model is built on a Transformer attention mechanism network, possessing powerful long-distance dependency modeling capabilities. This model receives high-level features (containing hour-level resource state information) and anchored inference results as input. Transformer directly establishes dependencies at any position in the sequence through a self-attention mechanism, overcoming the sequence processing limitations of RNN-type models, making it particularly suitable for long-term prediction tasks. The model employs an encoder-decoder architecture, where the encoder processes historical information and the decoder generates future predictions. A multi-head attention mechanism enables the model to simultaneously focus on different types of time-series patterns, such as daily, weekly, and monthly variations. The output layer uses a probabilistic prediction approach, providing not only point predictions but also confidence intervals to reflect the uncertainty of the predictions. These three models together constitute a multi-scale prediction model set, providing specially optimized solutions for prediction tasks at different time scales.

[0145] Based on the multi-scale prediction model set and the resource state tensor, the five-minute prediction results output by the short-term prediction model, the one-hour prediction results output by the medium-term prediction model, and the twenty-four-hour prediction results output by the long-term prediction model are calculated by the model fusion algorithm, and then fused by the weighted fusion method to obtain the prediction resource state tensor set.

[0146] Model fusion algorithms are a technique that integrates the results of multiple prediction models, fully leveraging the strengths of different models to improve overall prediction performance. In this embodiment, the algorithm first runs prediction models at three time scales independently, generating their respective prediction results. When the short-term prediction model runs, it uses the most recent high-precision data as input to generate a prediction sequence for the next five minutes at 30-second intervals, totaling 10 time points for seven-dimensional resource status predictions. When the medium-term prediction model runs, it uses historical data over a longer period as input to generate a prediction sequence for the next hour at 5-minute intervals, totaling 12 time points for predictions. When the long-term prediction model runs, it uses historical data spanning multiple periods as input to generate a prediction sequence for the next twenty-four hours at 1-hour intervals, totaling 24 time points for predictions. These three sets of prediction results overlap in time and need to be integrated into a consistent prediction result using a weighted fusion method. The weighted fusion method employs a performance-based dynamic weight allocation strategy, automatically adjusting the weights based on the prediction accuracy of each model on the historical validation set. For short-term overlapping regions (the next five minutes), the short-term model receives a higher weight (approximately 0.7), while the medium-term model receives a lower weight (approximately 0.3). For medium-term overlapping regions (5 minutes to 1 hour), the weight gradually shifts from the short-term model to the medium-term model. For long-term overlapping regions (1 hour to 24 hours), the weight gradually shifts from the medium-term model to the long-term model. The weight allocation also considers the uncertainty of the prediction, with models having lower uncertainty receiving higher weights. During the fusion process, the system not only merges point predictions but also prediction intervals to comprehensively evaluate the reliability of the predictions. For non-overlapping regions, the prediction results of the corresponding model are directly used. The final predicted resource state tensor set is a complete time-series prediction result, covering seven-dimensional resource state predictions at different time points within the next 24 hours, combining the high accuracy of short-term predictions with the trend accuracy of long-term predictions. This multi-model fusion method effectively overcomes the limitations of a single model, providing a more comprehensive and reliable prediction of future resource states, and providing a solid data foundation for resource conflict detection and early warning decisions.

[0147] Example 10

[0148] like Figure 3 As shown, the present invention also provides a unified modeling and tensor representation system for optical-computational multidimensional resources, comprising:

[0149] The resource feature acquisition module 10 is used to acquire the feature parameters of heterogeneous computing resources and optical network resources, and obtain a standardized multi-dimensional resource feature vector through resource detection, normalization processing and quantization encoding.

[0150] Tensor construction module 20 is used to obtain a seven-dimensional resource tensor containing computing power dimension, memory dimension, storage dimension, wavelength dimension, spectrum dimension, path dimension and time dimension based on the standardized multidimensional resource feature vector through dimension mapping and coordinate space construction, and to define the tensor subtraction operation rules for resource occupation, the tensor addition operation rules for resource release and the tensor multiplication operation rules for resource reachability, thus obtaining the tensor operation system;

[0151] The state mapping module 30 is used to map real-time resource states to resource state tensors and task resource requirements to resource requirement tensors according to the tensor operation system and through the state mapping algorithm, and to realize resource state transformation through tensor operation to obtain resource state tensors and corresponding transformation functions.

[0152] The resource scheduling module 40 is used to obtain the resource allocation tensor and the corresponding joint scheduling decision scheme based on the resource state tensor, the resource demand tensor, the resource state tensor and the corresponding transformation function, through tensor space search and optimization, and to obtain resource bottleneck report and resource conflict early warning report through historical resource state tensor sequence analysis.

[0153] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A unified modeling and tensor representation method for optical-computational multidimensional resources, characterized in that, include: The characteristic parameters of heterogeneous computing resources and optical network resources are obtained, and standardized multidimensional resource feature vectors are obtained through resource detection, normalization processing and quantization encoding. Based on the standardized multidimensional resource feature vector, a seven-dimensional resource tensor containing computing power dimension, memory dimension, storage dimension, wavelength dimension, spectrum dimension, path dimension and time dimension is obtained through dimension mapping and coordinate space construction. Tensor subtraction operation rules for resource occupation, tensor addition operation rules for resource release and tensor multiplication operation rules for resource reachability are defined to obtain a tensor operation system. According to the tensor operation system, the real-time resource state is mapped to a resource state tensor and the task resource requirement is mapped to a resource requirement tensor through the state mapping algorithm. The resource state transformation is realized through tensor operation to obtain the resource state tensor and the corresponding transformation function. Based on the resource state tensor, the resource demand tensor, and the resource state tensor and its corresponding transformation function, the resource allocation tensor and the corresponding joint scheduling decision scheme are obtained through tensor space search and optimization. Furthermore, resource bottleneck reports and resource conflict early warning reports are obtained through historical resource state tensor sequence analysis.

2. The method according to claim 1, characterized in that, The process of acquiring the characteristic parameters of heterogeneous computing resources and optical network resources, through resource detection, normalization processing, and quantization encoding, yields a standardized multi-dimensional resource feature vector, including: Based on the hardware specifications of heterogeneous computing nodes, raw parameter data of computing power, memory capacity, cache size and storage bandwidth of each computing node are collected to obtain the raw feature set of computing resources. Based on the optical network topology and device status information, the number of available wavelengths, spectrum range, time slot allocation, and hop count and delay parameters of the optional routing paths are collected to obtain the original feature set of optical network resources; Based on the original feature set of computing resources and the original feature set of optical network resources, resource parameters of different dimensions and orders of magnitude are converted into standardized values ​​within a unified numerical range through a normalization function, thereby obtaining a set of normalized feature values. Based on the set of normalized feature values, the continuous normalized feature values ​​are mapped to discrete quantization levels through a quantization function and digitally encoded to obtain the standardized multidimensional resource feature vector.

3. The method according to claim 1, characterized in that, Based on the standardized multidimensional resource feature vector, a seven-dimensional resource tensor is obtained through dimensional mapping and coordinate space construction, including dimensions of computing power, memory, storage, wavelength, spectrum, path, and time. Tensor subtraction rules for resource occupancy, tensor addition rules for resource release, and tensor multiplication rules for resource reachability are defined, resulting in a tensor operation system, including: Based on the standardized multidimensional resource feature vector, the tensor dimensions, including computing power dimension, memory dimension, storage dimension, wavelength dimension, spectrum dimension, path dimension and time dimension, are determined by the dimension mapping rules, and a tensor dimension definition set is obtained. Based on the tensor dimension definition set, a seven-dimensional coordinate space is obtained by mapping spatial coordinates to construct corresponding coordinate spaces for computing power dimension, memory dimension, storage dimension, wavelength dimension, spectrum dimension, path dimension and time dimension. Based on the seven-dimensional coordinate space, the physical meaning and value rules of each element in the tensor are determined by the element assignment function, thus obtaining the tensor element value definition rules; Based on the seven-dimensional coordinate space and the tensor element value definition rules, the standardized multi-dimensional resource feature vector is filled into the corresponding coordinate positions through the tensor initialization algorithm to obtain the seven-dimensional resource tensor. Based on the seven-dimensional resource tensor, the tensor subtraction operation is defined to represent the remaining resource state after resources are allocated or occupied, and the tensor subtraction operation rule is obtained. Based on the seven-dimensional resource tensor, the tensor addition operation is defined to represent the increased resource state after the resource is released or recycled, and the tensor addition operation rule is obtained. Based on the seven-dimensional resource tensor, the tensor multiplication operation rules are obtained by defining tensor multiplication operations to represent the connectivity and reachability between different resource nodes. Based on the tensor subtraction operation rules, the tensor addition operation rules, and the tensor multiplication operation rules, the non-negativity constraints, capacity upper limit constraints, and path continuity constraints that tensor operations must satisfy are defined through constraint condition functions, thus obtaining the tensor operation constraint set. Based on the tensor subtraction operation rules, the tensor addition operation rules, the tensor multiplication operation rules, and the tensor operation constraint set, a complete tensor operation library containing subtraction operation functions, addition operation functions, multiplication operation functions, and constraint verification functions is implemented through software programming, thus obtaining the tensor operation system.

4. The method according to claim 3, characterized in that, According to the tensor operation system, the real-time resource state is mapped to a resource state tensor and the task resource requirement is mapped to a resource requirement tensor through a state mapping algorithm. Resource state transformation is achieved through tensor operations to obtain the resource state tensor and the corresponding transformation function, including: Based on the actual resource status data and the seven-dimensional coordinate space of the seven-dimensional resource tensor, the CPU utilization, memory usage, storage input / output rate, wavelength occupancy, spectrum allocation status, path selection information, and time slice usage of the computing nodes in the current system are mapped to specific element values ​​in the seven-dimensional coordinate space through a state mapping algorithm, thereby obtaining the resource status tensor. Based on the resource requirement description of the task or application, the resource requirement tensor is obtained by quantifying and mapping the task's requirements for computing power, memory capacity, storage bandwidth, number of wavelengths, spectrum resources, path length and time window to the seven-dimensional coordinate space through a requirement mapping algorithm. Based on the resource state tensor, the resource demand tensor, and the tensor subtraction operation rules, the new state after resource allocation is calculated by tensor subtraction operation, and the state tensor after resource allocation is obtained by checking whether each element in the new state satisfies the non-negativity constraint through the tensor operation constraint set. Based on the resource state tensor, the resource to be released tensor, and the tensor addition operation rules, the new state after resource release is calculated by tensor addition operation, and the tensor operation constraint set is used to check whether each element in the new state exceeds the capacity limit constraint, thus obtaining the state tensor after resource release. Based on the state tensor after resource allocation and the state tensor after resource release, the output state tensor after the input state tensor is defined by the state transition function, thus obtaining the resource state tensor and the corresponding transition function.

5. The method according to claim 4, characterized in that, Before obtaining the resource allocation tensor and the corresponding joint scheduling decision scheme through tensor space search and optimization based on the resource state tensor, the resource demand tensor, the resource state tensor, and the corresponding transformation function, the process further includes: Based on the resource state tensor and its corresponding transition function, and the state transition sequence recorded in the system evolution history, the resource state tensor and operation sequence are combined and represented as a time-series decision sequence through a sequence coding algorithm to obtain the resource state transition sequence model. Based on the tensor subtraction, tensor addition, and tensor multiplication operations in the tensor operation system, a set of allowed resource operation tags is defined through a tag set construction algorithm to obtain a tag vocabulary; Based on the resource state transition sequence model and the labeled vocabulary, a probability distribution for selecting the next operation in the current state is constructed through a conditional probability model, and the tensor operation constraint set is used as a hard constraint to obtain a constraint sequence transition model. Based on the constrained sequence transition model, the resource state tensor, and the resource demand tensor, an operation sequence that satisfies the resource demand tensor and maximizes the state transition probability is generated in the operation space of the labeled vocabulary by a sequence generation algorithm, thereby obtaining a resource allocation operation sequence plan. Based on the resource allocation operation sequence plan and the resource state tensor and its corresponding transformation function, the intermediate state tensor corresponding to each step of the operation is calculated sequentially through the sequence-tensor mapping function to obtain the intermediate state tensor sequence and the final resource allocation state tensor.

6. The method according to claim 5, characterized in that, The process of obtaining a resource allocation tensor and a corresponding joint scheduling decision scheme based on the resource state tensor, the resource demand tensor, the resource state tensor, and the corresponding transformation function through tensor space search and optimization includes: Based on the resource demand tensor and the resource state tensor, the matching degree between task requirements and available system resources in the dimensions of computing power, memory, storage, wavelength, spectrum, path, and time is calculated using the tensor matching degree calculation function to obtain a set of resource matching degree indicators. Based on the resource matching degree index set and the system optimization objective, weights are assigned to each dimension of the matching degree index through weight allocation, and a multi-objective optimization function containing a matching degree weighted summation term, a delay penalty term, and a cost penalty term is constructed to obtain the optimization objective function; Based on the optimization objective function and the resource state tensor, the causal dependencies between the resource state variables of each dimension in the resource state tensor and the weighted summation term, delay penalty term, and cost penalty term in the optimization objective function are analyzed using a causal discovery algorithm. A causal relationship graph is constructed, and an intervention operator is defined to represent the operation of setting the numerical values ​​of the resource dimension variables in the causal relationship graph. By calculating the difference between the first function value of the optimization objective function before the intervention operator is applied and the second function value of the optimization objective function after the intervention operator is applied, a causal effect matrix is ​​obtained. Based on the causal effect matrix and the optimization objective function, the resource dimension variable with the largest causal effect value is selected as the priority intervention object through the causal intervention optimization algorithm. A search strategy based on causal effect ranking in the seven-dimensional coordinate space is designed to obtain a search strategy based on causal relationship. Based on the causal search strategy, the resource state tensor, and the resource demand tensor, the resource allocation tensor is obtained by using a tensor space search algorithm to find the resource allocation tensor that maximizes the value of the optimization objective function under the premise of satisfying the resource non-negativity constraint. Based on the resource allocation tensor, the element values ​​of each dimension of the allocation tensor in the tensor space are converted into CPU core allocation instructions, memory block allocation instructions, storage channel allocation instructions, wavelength allocation instructions, spectrum slot allocation instructions, routing path selection instructions, and time slice scheduling instructions through the decision mapping function, thereby obtaining the joint scheduling decision scheme.

7. The method according to claim 6, characterized in that, The step of finding the resource allocation tensor that maximizes the value of the optimization objective function based on the causal search strategy, the resource state tensor, and the resource demand tensor using a tensor space search algorithm under the premise of satisfying the resource non-negativity constraint, and obtaining the resource allocation tensor, includes: Based on the causal effect ranking results in the causal relationship-based search strategy and the resource state tensor, intervention operations are sequentially performed on multiple resource dimension variables with the largest causal effect values ​​through causal intervention tensor space search, generating multiple sets of candidate intervention operations and corresponding resource allocation tensor sets, and obtaining a set of candidate resource allocation schemes; Based on the set of candidate resource allocation schemes and the set of historically optimal resource allocation tensors stored in historical scheduling experience, the Euclidean distance between the candidate scheme and the historically optimal scheme is calculated using a decision boundary alignment algorithm. The historically optimal scheme with the smallest distance is selected as the reference scheme, and the allocation values ​​of each dimension of the candidate scheme are adjusted so that the allocation values ​​of each dimension of the candidate scheme are close to the corresponding allocation values ​​of the reference scheme, thus obtaining the resource allocation tensor.

8. The method according to claim 7, characterized in that, The resource bottleneck report and resource conflict early warning report are obtained through historical resource state tensor sequence analysis, including: Based on the resource status data recorded during system operation, a multi-scale sampling strategy is used to extract historical resource status tensors at three time granularities: second, minute, and hour, to obtain a multi-granularity historical information set. Based on the multi-granularity historical information set, a three-level coding network is used to encode second-level historical information into primary features, minute-level historical information is fused with primary features and then encoded into intermediate features, and hour-level historical information is fused with intermediate features and then encoded into advanced features, resulting in a progressive historical coding representation. Based on the seven-dimensional coordinate space of the seven-dimensional resource tensor and the historical resource usage pattern, a representative query anchor point is defined in each of the computing power dimension, memory dimension, storage dimension, wavelength dimension, spectrum dimension, path dimension and time dimension through the key information extraction algorithm to obtain a set of query anchor points; Based on the set of query anchors and the progressive historical encoding representation, the attention weight between each query anchor and the historical encoding feature is calculated through a parallel attention mechanism. The historical encoding features are weighted and summed to obtain the context vector of each query anchor, thus obtaining the anchoring inference result. Based on the anchored inference results and the resource state tensor, the ratio of used resources to total available resources is calculated in each dimension using an enhanced utilization analysis function and combined with historical trend data to obtain a resource utilization tensor containing utilization values ​​for each of the seven dimensions. The utilization values ​​of each dimension are extracted from the resource utilization tensor, and the resource dimensions with utilization values ​​exceeding a preset threshold are selected as bottleneck points using a bottleneck identification algorithm to obtain a resource bottleneck report containing bottleneck dimension identification, bottleneck degree, and historical trend analysis. Based on the progressive historical encoding representation and the anchored inference results, a short-term prediction model is constructed through a multi-timescale prediction network to predict the resource status in the next five minutes, a medium-term prediction model to predict the resource status in the next hour, and a long-term prediction model to predict the resource status in the next twenty-four hours. The future resource status prediction is generated by weighted fusion, and the predicted resource status tensor set is obtained. Based on the predicted resource state tensor set and the known future task requirement tensor set, a conflict detection function is used to check whether any dimension element of the predicted resource state minus the future task requirement is negative. If it exists, it is marked as a resource allocation conflict, and a resource conflict early warning report is obtained, which includes the time of conflict occurrence, the dimension of conflicting resources, and the severity of conflict.

9. The method according to claim 8, characterized in that, Based on the progressive historical encoding representation and the anchored inference results, a multi-timescale prediction network is used to construct a short-term prediction model to predict the resource status for the next five minutes, a medium-term prediction model to predict the resource status for the next one hour, and a long-term prediction model to predict the resource status for the next twenty-four hours. Weighted fusion is then used to generate future resource status predictions, resulting in a predicted resource status tensor set, including: Based on the progressive historical encoding representation and the anchored inference results, a long short-term memory network is constructed as a short-term prediction model by inputting the primary features and the anchored inference results into the temporal decomposition network, a gated recurrent unit network is constructed as a medium-term prediction model by inputting the intermediate features and the anchored inference results into the gated recurrent unit network, and an attention mechanism network is constructed as a long-term prediction model by inputting the high-level features and the anchored inference results into the attention mechanism network, thus obtaining a multi-scale prediction model set; Based on the multi-scale prediction model set and the resource state tensor, the five-minute prediction results output by the short-term prediction model, the one-hour prediction results output by the medium-term prediction model, and the twenty-four-hour prediction results output by the long-term prediction model are calculated by the model fusion algorithm, and then fused by the weighted fusion method to obtain the prediction resource state tensor set.

10. A unified modeling and tensor representation system for optical-computational multidimensional resources, characterized in that, include: The resource feature acquisition module is used to acquire feature parameters of heterogeneous computing resources and optical network resources. Through resource detection, normalization processing and quantization encoding, a standardized multi-dimensional resource feature vector is obtained. The tensor construction module is used to construct a seven-dimensional resource tensor based on the standardized multidimensional resource feature vector through dimension mapping and coordinate space construction, which includes computing power dimension, memory dimension, storage dimension, wavelength dimension, spectrum dimension, path dimension and time dimension. It also defines the tensor subtraction operation rules for resource occupation, the tensor addition operation rules for resource release and the tensor multiplication operation rules for resource reachability, thus obtaining the tensor operation system. The state mapping module is used to map real-time resource states to resource state tensors and task resource requirements to resource requirement tensors according to the tensor operation system, and to realize resource state transformation through tensor operation to obtain resource state tensors and corresponding transformation functions. The resource scheduling module is used to obtain the resource allocation tensor and the corresponding joint scheduling decision scheme based on the resource state tensor, the resource demand tensor, the resource state tensor and the corresponding transformation function, through tensor space search and optimization, and to obtain resource bottleneck report and resource conflict early warning report through historical resource state tensor sequence analysis.

Images