Resource sensing and scheduling method of computing power network

Abstract

The application claims a resource sensing and scheduling method of a computing power network, and belongs to the field of computing power network resource scheduling, and specifically comprises the following steps: collecting historical resource state sequences of each node as endogenous variables, and obtaining associated exogenous variable sequences; using a multi-source information fusion model to asymmetrically encode the two, fusing information through a cross-attention and adaptive gating mechanism, and introducing a structure prior to perform conditional normalization; outputting expected values and prediction variances of each resource index in a future time window, and aggregating them into a node resource state vector. In the scheduling stage, a double-time-scale framework is used, an adaptive resource reservation and candidate path scoring are periodically performed in a slow decision layer, and the reservation amount is adaptively adjusted according to uncertainty; a fast decision layer is event-driven, and a reinforcement learning sequence decision is used to complete routing and deployment, and a reward function fuses service quality and uncertainty penalty. The application improves the computing power network scheduling acceptance rate and robustness while ensuring service quality.

Description

Technical Field

[0001] This invention relates to the field of computing network resource scheduling technology, and in particular to a computing network resource perception and scheduling method oriented towards future time windows. Background Technology

[0002] Computing power networks provide users with low-latency, high-reliability services by unifying the scheduling of distributed computing, storage, and network resources. In a computing power network, scheduling requests typically involve resource allocation and path selection across multiple computing nodes and network paths. Once a scheduling decision is executed, it often remains effective for a period of time. Therefore, the scheduling effect depends not only on the current resource state of the nodes but also heavily on how that resource state evolves within future time windows.

[0003] However, the resource status of nodes in a computing network exhibits significant time-varying and uncertainties. Scheduling based solely on instantaneous resource status can easily lead to resource congestion, performance degradation, or service quality breaches during the decision-making process. Therefore, constructing a resource awareness mechanism oriented towards future time windows, providing a stable and predictable representation of resource status for scheduling decisions, is crucial for improving the scheduling performance and robustness of computing networks.

[0004] Existing resource status modeling and prediction methods mainly fall into two categories: traditional time series models such as ARIMA and exponential smoothing, which rely heavily on the historical data of individual nodes, making it difficult to effectively characterize the relationships between nodes and limiting their perception accuracy and generalization ability; and spatiotemporal prediction methods based on graph neural networks such as STGCN, DCRNN, and GraphWaveNet, which model the relationships between nodes by explicitly constructing a graph structure, but typically suffer from problems such as complex graph structure construction, high message passing overhead, and numerous deployment constraints, while also having limited ability to characterize the uncertainty of resource status. Furthermore, existing methods primarily focus on the historical state of individual nodes, rarely considering the impact of exogenous correlation information such as domain-level environmental features and task-level load characteristics on node resource status, thus limiting resource perception capabilities. In terms of scheduling decisions, existing methods mainly fall into two categories: heuristic optimization and deep reinforcement learning. The former, such as integer linear programming, has high solution complexity and is difficult to adapt to dynamic environments; the latter mostly adopts reactive decision-making on a single time scale, which struggles to balance global optimization and real-time response, and fails to fully utilize resource perception results to guide scheduling decisions. Furthermore, existing methods generally ignore the impact of resource state uncertainty on scheduling robustness—when the perceived confidence is low, scheduling based on the perceived result may lead to insufficient resources or a decline in service quality.

[0005] In summary, how to integrate multi-source correlation information to achieve efficient resource perception, how to quantify resource state uncertainty to support robust scheduling, and how to achieve collaborative optimization of resource perception and scheduling decisions have become urgent problems to be solved. Therefore, this invention proposes a resource perception and scheduling method for computing power networks.

[0006] A search revealed that application publication number CN121008930A belongs to the field of computing power scheduling technology. The method described provides a holographic resource-aware intelligent agent computing power scheduling method, with the following steps: collecting current resource status information of each computing node, including CPU utilization, memory usage, network bandwidth, etc.; constructing a holographic representation model of resource status to uniformly model multi-dimensional resources; making scheduling decisions based on the intelligent agent learning framework according to the current resource status and task requirements; and achieving distributed scheduling through multi-agent collaboration to improve scheduling efficiency and system throughput.

[0007] Unlike the methods described above, this invention addresses the continuous evolution of computing network resource states over time by introducing a resource awareness mechanism oriented towards future time windows. By fusing endogenous and exogenous variables of nodes, it generates a resource state vector that simultaneously characterizes the future resource load level and its uncertainty. During the scheduling phase, this invention employs a dual-timescale framework. The slow decision layer performs adaptive resource reservation and candidate path scoring based on the resource state vector, while the fast decision layer uses reinforcement learning for routing and deployment. An uncertainty penalty mechanism guides the agent to adopt conservative strategies in high-uncertainty regions. Given the time-varying and uncertain nature of node resource states in computing network scenarios, scheduling based solely on the current resource state is insufficient to handle resource fluctuations during the decision-making process, easily leading to resource congestion or service quality degradation. The aforementioned patents primarily schedule based on the current resource state, lacking characterization of future resource state evolution characteristics and failing to incorporate resource state uncertainty into the scheduling decision process, thus limiting scheduling robustness in dynamic environments. This invention, by achieving synergistic optimization of resource awareness and scheduling, improves the scheduling performance and system robustness of the computing network while ensuring service quality. Summary of the Invention

[0008] This invention aims to solve the problems of the prior art mentioned above. It proposes a resource awareness and scheduling method for computing power networks. The technical solution of this invention is as follows:

[0009] A resource awareness and scheduling method for a computing power network, comprising the following steps:

[0010] S1. Collect historical resource data of each node in the computing power network, and construct the endogenous variable sequence and exogenous variable sequence of the target node;

[0011] S2. Perform multi-source information fusion processing on the endogenous variable sequence and the exogenous variable sequence, and output the expected value and prediction variance of the resource prediction of the target node in the future time window;

[0012] S3. Based on the expected value and the prediction variance, aggregate to generate the resource state vector of the target node;

[0013] S4. In the slow decision layer, adaptive resource reservation is performed based on the resource state vector of each node, and a set of candidate paths and their scores are generated.

[0014] S5. In the fast decision layer, based on the candidate path set, reinforcement learning sequence decision-making is used to complete the routing and deployment of scheduling requests.

[0015] Furthermore, step S1 specifically includes:

[0016] S11. Collect historical resource data of the target node from the monitoring system at fixed time intervals, and construct an endogenous variable sequence of length L. The resource data includes one or more of computing resource indicators, storage resource indicators, and network resource indicators.

[0017] S12. Collect the aggregated features of the availability zone to which the target node belongs, and construct a regional exogenous variable sequence. The aggregated features include the statistics of each resource indicator within the availability zone.

[0018] S13. Collect task load characteristics related to the target node and construct a sequence of task-level exogenous variables. The task load characteristics include one or more of instance running status indicators and resource requirement indicators.

[0019] S14. Obtain the category identifier of the availability zone to which the target node belongs, which will be used to introduce prior structural information later.

[0020] Furthermore, step S2 specifically includes:

[0021] S21. Perform piecewise embedding processing on the endogenous variable sequence, dividing the historical time series into several time segments and mapping them into segment-level embedding representations;

[0022] S22. Perform global embedding processing on the exogenous variable sequences to obtain variable-level embedding representations;

[0023] S23. Using the cross-attention mechanism, with the endogenous variable embedding as the query term, the information of the exogenous variable embedding is fused into the endogenous representation;

[0024] S24. Adaptively adjust the contribution weights of each exogenous variable to the resource perception results through a gating fusion mechanism;

[0025] S25. Based on the fused representation, output the expected value and variance of each resource indicator of the target node within the future time window to characterize the resource load level and its uncertainty.

[0026] Furthermore, step S3 specifically includes:

[0027] S31. Based on the expected value sequence, calculate multi-scale load statistics for each resource indicator, including short-term mean, long-term mean, peak value and trend of change;

[0028] S32. Based on the variance value sequence, extract uncertainty statistics for each resource indicator;

[0029] S33. Concatenate the multi-scale load statistics and uncertainty statistics of each resource indicator to generate the resource state vector of the target node.

[0030] Furthermore, step S4 specifically includes:

[0031] S41 uses the load peak in the resource state vector as the basic reservation amount;

[0032] S42 calculates the uncertainty margin based on the uncertainty statistics in the resource state vector; the higher the uncertainty, the larger the margin.

[0033] S43 calculates the trend margin based on the changing trend in the resource state vector and adds extra reservations for nodes with an upward trend;

[0034] S44 combines the basic reserve, uncertainty margin, and trend margin to obtain the resource reserve and remaining available resources for each node;

[0035] S45 generates a set of candidate paths for each source-destination node pair in the network;

[0036] S46 uses resource state vectors to comprehensively score each candidate path from dimensions such as resource load, uncertainty, transmission delay, and load balancing.

[0037] Furthermore, step S5 specifically includes:

[0038] S51 models the routing and deployment problem as a sequential decision process, which includes path selection and node placement for each functional unit.

[0039] S52 constructs a state space, which includes node resource characteristics, candidate path characteristics, scheduling request characteristics, and completed decision contexts.

[0040] S53 constructs the action space, selecting from the candidate path set during the path selection phase, and selecting from the nodes traversed by the selected path during the node placement phase.

[0041] The S54 design reward function integrates deployment result rewards, service quality penalties, load balancing rewards, and uncertainty penalties.

[0042] S55 employs a reinforcement learning algorithm training strategy, which makes it tend to choose a conservative scheduling strategy in regions of high uncertainty.

[0043] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements a resource awareness and scheduling method for a computing network as described in any one of the claims.

[0044] A non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements a resource awareness and scheduling method for a computing power network as described in any one of the claims.

[0045] The advantages and beneficial effects of this invention are as follows:

[0046] 1. Step S1 of this invention simultaneously collects the endogenous and exogenous variable sequences of the target node, incorporating the node's historical state and external correlation information into a unified resource perception framework. Most existing methods only focus on the node's own historical resource data, rarely considering the impact of external factors such as availability zone-level environmental characteristics and task-level load characteristics on the node's resource state. This invention, by introducing exogenous variables, can capture the correlation between the node's resource state and the external environment, thereby improving the accuracy and generalization ability of resource perception.

[0047] 2. Step S2 of this invention achieves asymmetric multi-source information encoding and fusion by segmenting and embedding the endogenous variable sequence and performing variable-level embedding on the exogenous variable sequence. Existing graph neural network-based methods typically use a uniform encoding method for all inputs, ignoring the differences in temporal granularity and information density between endogenous and exogenous variables. The asymmetric encoding strategy of this invention can differentiate the processing according to the characteristics of different types of variables, and adaptively adjust the contribution weight of each exogenous variable through cross-attention and gating fusion mechanisms, thereby improving perception accuracy while reducing computational complexity.

[0048] 3. Step S3 of this invention outputs expected and variance values ​​and aggregates them to generate a resource state vector containing multi-scale load statistics and uncertainty statistics, thus achieving a joint representation of resource load levels and their uncertainties. Most existing methods only output point estimates of resource states, lacking quantification of the confidence level of the perceived results, making it difficult for downstream scheduling decisions to distinguish between high-confidence and low-confidence perceived results. This invention provides a basis for risk perception for subsequent scheduling decisions by explicitly modeling uncertainty.

[0049] 4. Steps S4 and S5 of this invention achieve collaborative optimization of resource perception and scheduling decision-making through a dual-timescale framework. The slow decision-making layer performs adaptive resource reservation based on uncertainty, while the fast decision-making layer guides a conservative scheduling strategy through uncertainty penalty. Most existing scheduling methods adopt reactive decision-making with a single timescale, which is difficult to balance global optimization and real-time response, and fails to incorporate resource state uncertainty into the decision-making process. This invention, by transferring uncertainty from the perception stage to the scheduling stage, enables the scheduling strategy to automatically adopt conservative behavior in high-uncertainty regions, thereby improving the robustness of the system in dynamic environments. Attached Figure Description

[0050] Figure 1 This is a flowchart of the preferred embodiment provided by the present invention;

[0051] Figure 2 Diagram of the cross-attention fusion mechanism;

[0052] Figure 3 This is a diagram of the dual-head prediction output structure.

[0053] Figure 4 This is a diagram illustrating slow decision-making.

[0054] Figure 5 Diagram illustrating quick decision-making. Detailed Implementation

[0055] The technical solutions of the embodiments of the present invention will be clearly and thoroughly described below with reference to the accompanying drawings. The described embodiments are merely some embodiments of the present invention.

[0056] The technical solution of the present invention to solve the above-mentioned technical problems is:

[0057] like Figure 1 As shown, the present invention provides a resource awareness and scheduling strategy for a computing network, comprising the following steps:

[0058] S1. Collect historical resource data of each node in the computing power network, and construct the endogenous variable sequence and exogenous variable sequence of the target node;

[0059] S2. Perform multi-source information fusion processing on the endogenous variable sequence and the exogenous variable sequence, and output the expected value and prediction variance of the resource prediction of the target node in the future time window;

[0060] S3. Based on the expected value and the prediction variance, aggregate to generate the resource state vector of the target node;

[0061] S4. In the slow decision layer, adaptive resource reservation is performed based on the resource state vector of each node, and a set of candidate paths and their scores are generated.

[0062] S5. In the fast decision layer, based on the candidate path set, reinforcement learning sequence decision-making is used to complete the routing and deployment of scheduling requests.

[0063] The invention is illustrated below with specific examples:

[0064] Assume the computing network contains N = The historical observation window L = 12 time steps (5 minutes per step, totaling 60 minutes of history), and the future prediction T = 6 time steps (30 minutes). Nodes are interconnected via a wide area network, and each node is distributed across different availability zones.

[0065]

[0066] In this embodiment, the resource indicators of each node are expressed in the form of utilization rate, with a value range of 0-100%.

[0067] In this embodiment, step S1 involves collecting historical resource data from each node in the computing power network and constructing the endogenous variable sequence and exogenous variable sequence of the target node. Specifically, this includes the following steps:

[0068] S11. From the monitoring system at fixed time intervals Collect historical resource data of the target node every minute, and construct a data structure with a length of [length missing]. The sequence of endogenous variables. Based on nodes. For example, the sequence of endogenous variables is represented as:

[0069]

[0070] in This refers to the quantity of resource indicators. This embodiment uses... Several resource metrics, including CPU utilization, memory utilization, and network bandwidth utilization, are expressed as percentages. Node An example of an endogenous variable sequence is as follows:

[0071]

[0072] S12. Collect the aggregation features of the availability zone to which the target node belongs, and construct a regional-level exogenous variable sequence. For nodes For the corresponding availability zone AZ-1, calculate the mean of each resource indicator for all nodes within this availability zone to obtain the region-level exogenous variable sequence:

[0073] ;

[0074] The characteristic of each time step is the arithmetic mean of the corresponding indicators of each node within the availability zone:

[0075] ;

[0076] S13. Collect task load characteristics related to the target node and construct a task-level exogenous variable sequence, wherein the task load characteristics include:

[0077] ;

[0078] in This is the task feature dimension. This embodiment uses... ,include The number of instances reached The total planned CPU resources are plan_cpu_sum;

[0079] S14. Obtain the category identifier of the availability zone to which the target node belongs, which will be used to introduce prior structural information later. In this embodiment, there are 3 availability zones and nodes. The category identifier for the availability zone AZ-1 is az_id=1;

[0080] In this embodiment, step S2 involves multi-source information fusion processing of the endogenous variable sequence and the exogenous variable sequence to output the expected value and prediction variance of the resource prediction for the target node within a future time window. Specifically, this includes the following steps:

[0081] S21. Perform piecewise embedding on the endogenous variable sequence. Divide the historical sequence of length L=12 into P=4 time segments, each segment containing 3 time steps. Project each segment onto a linear mapping. The representation space of dimension:

[0082] ;

[0083] Where PatchEmbed is the fragment embedding layer, and PE is the positional encoding. For the i-th fragment:

[0084] ;

[0085] in The projection matrix for learning;

[0086] S22. Perform global embedding processing on the exogenous variable sequences to obtain variable-level embedding representations. Unlike the segmented processing of endogenous variables, exogenous variables are encoded at the variable level, mapping the entire sequence to a single embedding vector:

[0087] ;

[0088] ;

[0089] S23. Using a cross-attention mechanism, with the endogenous variable embedding as the query term, the information from the exogenous variable embedding is fused into the endogenous representation. For queries, each region is used as an exogenous embedding. and task-level exogenous embedding Extract relevant information for key / value pairs:

[0090] ;

[0091] ;

[0092] The formula for calculating cross attention is:

[0093] ;

[0094] S24. The contribution weights of each exogenous variable to the resource perception result are adaptively adjusted through a gating fusion mechanism. The contribution of each exogenous variable is adaptively adjusted through learnable gating coefficients.

[0095] ;

[0096] ;

[0097] in Let be the gating coefficient, when When the value is close to 1, regional-level exogenous information is more trusted. When the value is close to 0, task-level exogenous information is trusted more.

[0098] S25. Based on the fused representation, output the expected and variance values ​​of each resource indicator for the target node within the future time window to characterize the resource load level and its uncertainty. Take the final time step hidden state of the fused identifier, and output the expected and variance values ​​of each resource indicator within the next H=6 time steps through the prediction head:

[0099] ;

[0100] ;

[0101] ;

[0102] ;

[0103] in This is the numerically stable term. Model training uses the Gaussian negative log-likelihood loss function:

[0104] ;

[0105] node Here is an example of the predicted output:

[0106]

[0107] In this embodiment, step S3, which involves aggregating and generating the resource state vector of the target node based on the expected value and the prediction variance, specifically includes the following steps:

[0108] S31. Based on the expected value sequence, calculate multi-scale load statistics for each resource indicator, including short-term means. (In the near future) 2) Long-term mean (Full forecast window for the future), peak value and changing trends :

[0109] ;

[0110] ;

[0111] ;

[0112] ;

[0113] S32. Based on the variance value sequence Uncertainty statistics are extracted for each resource indicator. First, the variance is converted to standard deviation. Then calculate:

[0114] ;

[0115] ;

[0116] in Indicates average uncertainty. This indicates peak uncertainty;

[0117] S33. Concatenate the multi-scale load statistics and uncertainty statistics of each resource indicator to generate the resource state vector of the target node. For D=3 resource indicators, each indicator contains 6 statistics, and the resource state vector has a dimension of 18:

[0118] ;

[0119] node An example of a resource state vector is as follows:

[0120]

[0121] In this embodiment, step S4, at the slow decision layer, adaptive resource reservation is performed based on the resource state vector of each node, and a candidate path set and its score are generated. Specifically, this includes the following steps:

[0122] S41 uses the load peak in the resource state vector as the basic reservation amount:

[0123]

[0124] S42 calculates the uncertainty margin based on the uncertainty statistics in the resource state vector:

[0125]

[0126] in For the uncertainty sensitivity coefficient, this embodiment takes... The higher the uncertainty, the larger the margin should be.

[0127] S43 calculates the trend margin based on the changing trend in the resource state vector, and adds extra reservations for nodes showing an upward trend:

[0128]

[0129] in As the trend sensitivity coefficient, this embodiment takes... Additional reservations are added only for nodes showing an upward trend; reservations are not reduced for nodes showing a downward trend.

[0130] S44 integrates all reserved quantities and constrains them to not exceed the capacity limit:

[0131]

[0132]

[0133] in To maximize the reserved ratio, this embodiment takes... Taking CPU resources as an example, the CPU resource reservation and availability of each node are as follows:

[0134]

[0135] S45 generates a set of candidate paths for each source-destination node pair in the network. This embodiment uses the K-shortest path algorithm to generate a set of candidate paths for each source-destination node pair. Candidate paths. Starting from the source node. To the destination node For example:

[0136]

[0137]

[0138]

[0139] S46 uses resource state vectors to comprehensively score each candidate path from multiple dimensions. The lower the score, the better the path.

[0140] (a) Load score: Average long-term load of path nodes

[0141]

[0142] (b) Stability Score: Mean absolute value of the load change trend of path nodes

[0143]

[0144] (c) Uncertainty Score: Average Uncertainty of Path Nodes

[0145]

[0146] (d) Delay Score: Total Path Transmission Delay

[0147]

[0148] (e) Balance score: Coefficient of variation of available resources at path nodes

[0149]

[0150] The overall score is a weighted sum of the scores from each dimension:

[0151]

[0152] This embodiment uses weights. The following are examples of the scores for the three candidate paths:

[0153]

[0154] In this embodiment, in step S5, at the fast decision layer, based on the candidate path set, reinforcement learning sequence decision-making is used to complete the routing and deployment of the scheduling request. Assume the current scheduling request contains M=3 functional units and needs to be routed from... Routing to Specifically, it includes the following steps:

[0155] S51 models the routing and deployment problem as a sequential decision process. The decision sequence work... step:

[0156] Step 1: Select one path from K=3 candidate paths

[0157] Step 2: Select a placement node for functional unit f1

[0158] Step 3: Select a placement node for functional unit f2

[0159] Step 4: Select a placement node for functional unit f3

[0160] S52 constructs the state space and state vectors. Includes node resource characteristics Candidate path features Scheduling request characteristics and the completed decision context :

[0161] ;

[0162] ;

[0163] ;

[0164] ;

[0165] S53 constructs an action space. In the path selection phase, paths are chosen from the candidate path set, and in the node placement phase, nodes are chosen from those traversed by the selected path. A conditional action space is used, dynamically switching based on the current decision step.

[0166] ;

[0167] ;

[0168] To choose For example, the subsequent placement action space is {0,1,2}, corresponding to nodes respectively. .

[0169] The S54 design reward function integrates deployment result rewards, service quality penalties, load balancing rewards, and uncertainty penalties.

[0170] ;

[0171] in This indicates a reward for deployment results. This indicates the difference between the end-to-end delay and the delay requirement. This indicates the degree of load balance among the nodes. This represents the penalty for selecting nodes with high uncertainty.

[0172] The S55 uses the Proximal Policy Optimization (PPO) algorithm to train the policy network. The policy network employs a multilayer perceptron structure, with the input being a state vector. The output is the action probability distribution. .

[0173] The training objective is to maximize the expected value of the cumulative discount reward.

[0174] ;

[0175] in This is the discount factor. PPO ensures the stability of policy updates by pruning the objective function:

[0176] ;

[0177] in Importance sampling ratio, For the estimation of the advantage function, These are the trimming parameters.

[0178] Decision example: Suppose that in the current state, the agent outputs the following actions in sequence:

[0179] Step 1: Select (Lowest overall score, moderate uncertainty)

[0180] Step 2: Place f1 in

[0181] Step 3: Place f2 in

[0182] Step 4: Place f3 in

[0183] Final deployment plan: The nodes on are assigned as (f1, f2, f3 ), in Available resources: 22% 37% of resources are available. Available resources: 41%; end-to-end latency: 8ms; meets latency requirement of 10ms; deployment successful.

[0184] The systems, devices, modules, or units described in the above embodiments can be implemented by computer chips or entities, or by products with certain functions.

[0185] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0186] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0187] The above embodiments should be understood as illustrative only and not as limiting the scope of protection of the present invention. After reading the description of the present invention, those skilled in the art can make various alterations or modifications to the present invention, and these equivalent changes and modifications also fall within the scope defined by the claims of the present invention.

Claims

1. A resource awareness and scheduling method for a computing power network, characterized in that, Includes the following steps: S1. Collect historical resource data of each node in the computing power network, and construct the endogenous variable sequence and exogenous variable sequence of the target node; S2. Perform multi-source information fusion processing on the endogenous variable sequence and the exogenous variable sequence, and output the expected value and prediction variance of the resource prediction of the target node in the future time window; S3. Based on the expected value and the prediction variance, aggregate to generate the resource state vector of the target node; S4. In the slow decision layer, adaptive resource reservation is performed based on the resource state vector of each node, and a set of candidate paths and their scores are generated. S5. In the fast decision layer, based on the candidate path set, reinforcement learning sequence decision-making is used to complete the routing and deployment of scheduling requests.

2. The resource awareness and scheduling method for a computing network according to claim 1, characterized in that, Step S1 specifically includes: S11. Collect historical resource data of the target node from the monitoring system at fixed time intervals, and construct an endogenous variable sequence of length L. The resource data includes one or more of computing resource indicators, storage resource indicators, and network resource indicators. S12. Collect the aggregated features of the availability zone to which the target node belongs, and construct a regional exogenous variable sequence. The aggregated features include the statistics of each resource indicator within the availability zone. S13. Collect task load characteristics related to the target node and construct a sequence of task-level exogenous variables. The task load characteristics include one or more of instance running status indicators and resource requirement indicators. S14. Obtain the category identifier of the availability zone to which the target node belongs, which will be used to introduce prior structural information later.

3. The resource awareness and scheduling method for a computing power network according to claim 1, characterized in that, Step S2 specifically includes: S21. Perform piecewise embedding processing on the endogenous variable sequence, dividing the historical time series into several time segments and mapping them into segment-level embedding representations; S22. Perform global embedding processing on the exogenous variable sequences to obtain variable-level embedding representations; S23. Using the cross-attention mechanism, with the endogenous variable embedding as the query term, the information of the exogenous variable embedding is fused into the endogenous representation; S24. Adaptively adjust the contribution weights of each exogenous variable to the resource perception results through a gating fusion mechanism; S25. Based on the fused representation, output the expected value and variance of each resource indicator of the target node within the future time window to characterize the resource load level and its uncertainty.

4. The resource awareness and scheduling method for a computing power network according to claim 1, characterized in that, Step S3 specifically includes: S31. Based on the expected value sequence, calculate multi-scale load statistics for each resource indicator, including short-term mean, long-term mean, peak value and trend of change; S32. Based on the variance value sequence, extract uncertainty statistics for each resource indicator; S33. Concatenate the multi-scale load statistics and uncertainty statistics of each resource indicator to generate the resource state vector of the target node.

5. The resource awareness and scheduling method for a computing power network according to claim 1, characterized in that, Step S4 specifically includes: S41 uses the load peak in the resource state vector as the basic reservation amount; S42 calculates the uncertainty margin based on the uncertainty statistics in the resource state vector; the higher the uncertainty, the larger the margin. S43 calculates the trend margin based on the changing trend in the resource state vector and adds extra reservations for nodes with an upward trend; S44 combines the basic reserve, uncertainty margin, and trend margin to obtain the resource reserve and remaining available resources for each node; S45 generates a set of candidate paths for each source-destination node pair in the network; S46 uses resource state vectors to comprehensively score each candidate path from dimensions such as resource load, uncertainty, transmission delay, and load balancing.

6. The resource awareness and scheduling method for a computing power network according to claim 1, characterized in that, Step S5 specifically includes: S51 models the routing and deployment problem as a sequential decision process, which includes path selection and node placement for each functional unit. S52 constructs a state space, which includes node resource characteristics, candidate path characteristics, scheduling request characteristics, and completed decision contexts. S53 constructs the action space, selecting from the candidate path set during the path selection phase, and selecting from the nodes traversed by the selected path during the node placement phase. The S54 design reward function integrates deployment result rewards, service quality penalties, load balancing rewards, and uncertainty penalties. S55 employs a reinforcement learning algorithm training strategy, which makes it tend to choose a conservative scheduling strategy in regions of high uncertainty.

7. An electronic device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the resource awareness and scheduling method of the computing power network as described in any one of claims 1 to 6.

8. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the resource awareness and scheduling method of the computing power network as described in any one of claims 1 to 6.

Images