Method and device for allocating computing power network resources

Abstract

The embodiment of the invention provides a computing power network resource allocation method and device, which can solve the technical problem of poor dynamic performance of computing power network resource allocation in related technologies. The method comprises the following steps: based on a general scheduling interface, inputting task characteristics and performance index constraints of a computing network task into a first model to obtain a combined association result; inputting the combined association result into a second model to obtain a time sequence feature between the task feature and the performance index constraint; and according to the task characteristics, the performance index constraint and the time sequence characteristics, determining a dynamic resource allocation result and a prediction performance index of the network calculation task. Adapting multiple types of computing network tasks and underlying resources based on a general scheduling interface; the first model is associated with task features, multi-dimensional resource requirements and performance constraints, so that the one-sidedness of single-dimensional matching is avoided, and the execution success rate is improved. The second model captures a task and a time sequence rule, and the scheduling precision is improved. Finally, resources are dynamically allocated in combination with multi-dimensional information, and the dynamic performance of resource allocation is improved.

Description

Technical Field

[0001] This application relates to the field of digital infrastructure, and more particularly to a method and device for allocating computing power network resources. Background Technology

[0002] Computing power (CP) refers to the ability of a device or system to process data and run tasks per unit of time, with the core metric being floating-point operations per second (FLOPS). Computing power has become a core production factor in digital infrastructure development. As a critical infrastructure for integrating distributed computing resources and enabling on-demand allocation of computing power, computing networks are essential for supporting emerging businesses such as cloud computing and artificial intelligence. While current computing networks have developed various resource scheduling and interface schemes, they generally suffer from insufficient dynamic adaptability, limited accuracy in capturing timing dependencies, and poor interface universality. These shortcomings make it difficult to cope with dynamic changes in business needs, fluctuations in resource load, and the rapid access of diverse services, resulting in low computing power utilization efficiency and high costs for adapting new services. Summary of the Invention

[0003] This application provides a method and device for allocating computing network resources, which can solve the technical problem of poor dynamic performance of computing network resource allocation in related technologies.

[0004] In a first aspect, embodiments of this application provide a method for allocating computing network resources. The method includes: inputting the task characteristics and performance constraints of computing network tasks into a first model based on a general scheduling interface to obtain a combined correlation result; inputting the combined correlation result into a second model to obtain time series characteristics between the task characteristics and performance constraints; and determining the dynamic resource allocation result and predicted performance indicators of the computing network tasks based on the task characteristics, performance constraints, and time series characteristics.

[0005] Based on the above description of the computing power network resource allocation method provided in the embodiments of this application, it can be seen that this method is based on a general scheduling interface, adapts to multiple types of computing network tasks and underlying resources, solves the problem of coupling between traditional scheduling logic and business, and improves versatility. The first model associates task characteristics, multi-dimensional resource requirements, and performance constraints, avoiding the one-sidedness of single-dimensional matching and providing support for task execution success rate. The second model captures task and time sequence patterns, makes up for the shortcomings of traditional scheduling that ignores time sequence, solves the problem of resource allocation lag or over-reservation, and improves scheduling accuracy. Finally, based on task characteristics, performance index constraints, and time sequence characteristics, resources are dynamically allocated in combination with multi-dimensional information, reducing the average task latency, improving resource utilization, and at the same time, interface decoupling makes the system easy to expand, comprehensively optimizing the computing power resource scheduling effect.

[0006] In one possible implementation of the first aspect, based on a general scheduling interface, the task characteristics and performance constraints of the computing network task are input into a first model to obtain a combined correlation result; this includes: determining the resource combination result corresponding to the computing network task based on the task characteristics; the task characteristics include at least one of task size, task request concurrency, and task algorithm; the resource combination result is used to characterize the computing resource requirements, including at least one of memory resource requirements, processor resource requirements, image processor resource requirements, and video memory requirements; obtaining the computing resource allocation result based on the resource combination result and the resource utilization threshold; obtaining the result of minimizing the prediction error based on the performance constraints and the performance threshold; and obtaining the combined correlation result based on the computing resource allocation result and the result of minimizing the prediction error.

[0007] In one possible implementation of the first aspect, the performance metric constraints include task response latency and task success rate, and the performance metric thresholds include a maximum allowable latency value and a success rate threshold. Based on the performance metric constraints and performance metric thresholds, a result minimizing the prediction error is obtained, including: obtaining a first association result based on the task response latency and the maximum allowable latency value; obtaining a second association result based on the task success rate and the success rate threshold; and obtaining a result minimizing the prediction error based on the first association result, a first weight, the second association result, and a second weight; wherein the first weight is the weight parameter corresponding to the first association result; and the second weight is the weight parameter corresponding to the second association result.

[0008] In one possible implementation of the first aspect, the dynamic resource allocation result and predicted performance index of the computing network task are determined based on task characteristics, performance index constraints, and time series characteristics. This includes: determining the input feature vector at the current moment based on the task size, the number of concurrent task requests, and the algorithm embedding vector of the task algorithm; obtaining the forget gate output result based on the input feature vector at the current moment and the hidden state at the previous moment using a first activation function; obtaining the input gate output result and candidate cell state result based on the input feature vector at the current moment and the hidden state at the previous moment using a second activation function; combining the forget gate output result, the cell state at the previous moment, the input gate output result, and the candidate cell state result to obtain the cell state at the current moment; obtaining the output gate output result based on the input feature vector at the current moment and the hidden state at the previous moment using a third activation function; obtaining the hidden state at the current moment based on the output gate output result and the cell state at the current moment; and obtaining the dynamic resource allocation result based on the hidden state at the current moment.

[0009] In one possible implementation of the first aspect, the input feature vector at the current moment is determined based on the task size, the number of concurrent task requests, and the algorithm embedding vector of the task algorithm. This includes: obtaining a historical dataset; the historical dataset includes business features, target output, embedding dimension, hidden layer size, and hyperparameters; normalizing the task size and the number of concurrent task requests in the business features to obtain normalized features; and concatenating the normalized features with the algorithm embedding vector of the task algorithm to generate the input feature vector at the current moment.

[0010] In one possible implementation of the first aspect, the method further includes: constructing a loss function based on latency loss, success rate loss, and resource demand loss; and updating model parameters through backpropagation based on the loss function.

[0011] In one possible implementation of the first aspect, the layered architecture of the general scheduling interface includes an application programming interface (API) layer, a logic layer, and a data layer. Based on the general scheduling interface, the task characteristics and performance constraints of the computing network task are input into a first model to obtain a combined association result. This includes: the API layer parsing and verifying the received task characteristics and performance constraints of the computing network task, and then passing them to the logic layer; the logic layer inputting the parsed and verified task characteristics and performance constraints into the first model, and obtaining the combined association result of the computing network task through the calculation of the first model; and the API layer receiving the combined association result fed back by the logic layer and returning the combined association result to the outside.

[0012] In one possible implementation of the first aspect, the method includes: a general scheduling interface limiting the number of calls per second using a token bucket algorithm; and rate limiting and concurrency control of received computing network task requests by setting user-level and task-level concurrency limits.

[0013] In one possible implementation of the first aspect, the method further includes: when the computing network task is a task in a high-concurrency scenario, the general scheduling interface realizes the asynchronous execution of the computing network task through a message queue; in response to receiving task characteristics and performance index constraints, it returns the task identity ID; in response to the first model operation obtaining the combined association result, it feeds back the combined association result to the outside through a query interface; the logic layer uses asynchronous input or output ports to input the task characteristics and performance index constraints into the first model.

[0014] Secondly, embodiments of this application provide a computing power network resource allocation device, which includes an execution unit for executing the computing power network resource allocation method described in the first aspect.

[0015] The computing power network resource allocation equipment mentioned in the second aspect above can refer to the beneficial effects of the first aspect above and any of its possible design methods, which will not be elaborated here. Attached Figure Description

[0016] The accompanying drawings are provided to better understand this solution and do not constitute a limitation on the embodiments of this application. Wherein: Figure 1 A flowchart illustrating a computing power network resource allocation method provided in an embodiment of this application; Figure 2 A flowchart illustrating a computing power network resource allocation method provided in an embodiment of this application; Figure 3 A flowchart illustrating a computing power network resource allocation method provided in an embodiment of this application; Figure 4 A flowchart illustrating a computing power network resource allocation method provided in an embodiment of this application; Figure 5 A flowchart illustrating a computing power network resource allocation method provided in an embodiment of this application; Figure 6 A flowchart illustrating a computing power network resource allocation method provided in an embodiment of this application; Figure 7 A flowchart illustrating a computing power network resource allocation method provided in an embodiment of this application; Figure 8 This is a schematic diagram of the structure of a general scheduling interface in a computing power network resource allocation method provided in an embodiment of this application. Detailed Implementation

[0017] The technical solutions of the embodiments of this application will be described below with reference to the accompanying drawings. In the description of this application, unless otherwise stated, " / " indicates that the objects before and after are in an "or" relationship. For example, A / B can represent A or B. "And / or" in this application is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, and B alone, where A and B can be singular or plural. Furthermore, in the description of this application, unless otherwise stated, "multiple" refers to two or more. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one of a, b, or c can represent: a, b, c, ab, ac, bc, or abc, where a, b, and c can be single or multiple. Furthermore, to facilitate a clear description of the technical solutions in the embodiments of this application, the terms "first" and "second" are used in the embodiments of this application to distinguish identical or similar items with substantially the same function and effect. Those skilled in the art will understand that the terms "first" and "second" do not limit the quantity or execution order, and that "first" and "second" are not necessarily different. Meanwhile, in the embodiments of this application, the terms "exemplary" or "for example" are used to indicate that something is being used as an example, illustration, or description. Any embodiment or design scheme described as "exemplary" or "for example" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design schemes. Specifically, the use of terms such as "exemplary" or "for example" is intended to present related concepts in a concrete manner for ease of understanding.

[0018] Furthermore, the network architecture and business scenarios described in the embodiments of this application are for the purpose of more clearly illustrating the technical solutions of the embodiments of this application, and do not constitute a limitation on the technical solutions provided in the embodiments of this application. As those skilled in the art will know, with the evolution of network architecture and the emergence of new business scenarios, the technical solutions provided in the embodiments of this application are also applicable to similar technical problems.

[0019] With the rapid development of the digital economy and the widespread application of emerging technologies such as cloud computing, big data, and artificial intelligence, the demand for computing power in various business scenarios is becoming increasingly diversified, dynamic, and large-scale. As a core technology for integrating and scheduling distributed computing resources and achieving efficient sharing and on-demand allocation of computing power, computing networks have become a critical infrastructure supporting the development of the digital economy.

[0020] Computing networks integrate and schedule computing resources distributed across different regions through network technology, building a distributed computing platform with dynamic resource awareness, intelligent task allocation, and efficient transmission mechanisms. This platform can meet the flexible computing power needs of different application scenarios, effectively improve overall computing efficiency, and reduce operating costs, making it of significant application value in multiple fields such as industrial internet, smart cities, and high-definition video processing.

[0021] To achieve the rational allocation and efficient utilization of computing resources, various computing network resource scheduling schemes have been proposed in existing technologies, mainly including the following three categories: The first type is a computing resource scheduling scheme based on fixed rules. This type of scheme typically pre-defines resource allocation rules, such as scheduling based on task submission time order, task priority level, or a fixed resource quota ratio. By pre-setting resource allocation thresholds and trigger conditions, resource allocation is completed directly by matching the corresponding rules after a task enters the scheduling queue. Its core characteristic is that the scheduling logic is fixed and cannot be adjusted according to dynamic changes in real-time computing load, network bandwidth fluctuations, or task resource requirements.

[0022] The second category is resource orchestration schemes based on traditional time-series models. For computing network tasks with time-series characteristics, this type of scheme uses traditional time-series prediction models such as ARIMA and sliding windows. It establishes a prediction model with fixed parameters by analyzing historical task resource demand time-series data, and then schedules resources based on the prediction results. However, during implementation, the model parameters need to be manually adjusted periodically to adapt to changes in data distribution, and it is difficult to capture the complex nonlinear time-series dependencies in task requirements.

[0023] The third category is resource scheduling interface solutions based on tight business coupling. Most resource scheduling systems' API interfaces are tightly coupled to specific business scenarios. The interface design is customized to meet the resource requirements and scheduling logic of specific businesses. For example, the scheduling interface for video processing only supports input of computing power and storage parameters related to video transcoding, and the interface is directly bound to the underlying resource scheduling logic. Adapting to new business types requires large-scale modifications to the interface structure and scheduling logic.

[0024] However, the aforementioned existing technical solutions all have significant drawbacks: For computing resource scheduling schemes based on fixed rules, they do not consider the dynamic and complex nature of computing network resource allocation, nor do they flexibly adapt to the diversity of computing resources required by tasks. This leads to a decrease in task execution success rate and an increase in average latency when task requirements change suddenly or resource load is unbalanced. For resource orchestration schemes based on traditional time-series models, they lack adaptive learning capabilities and cannot dynamically optimize model parameters to match the time-series characteristics of real-time tasks. This results in low accuracy in describing computing network orchestration and scheduling problems, especially when task requirement time-series patterns change abruptly. For resource scheduling interface schemes based on tightly coupled business operations, they lack universality and scalability. They cannot achieve intelligent allocation of computing, network, and storage resources for different business types through a unified interface. This leads to high development costs and long development cycles when adapting the system to new businesses, making it difficult to meet the rapid access needs of diverse businesses in the computing network.

[0025] Therefore, how to solve the problems of lack of dynamic adaptation capability in computing resource scheduling, insufficient accuracy of timing prediction, and poor interface universality and scalability in existing technologies has become an urgent technical challenge in the current computing network field.

[0026] To address the above issues, this application provides a method for allocating computing power network resources, which can provide an objective and effective method for allocating computing power network resources and is applicable to scenarios such as wireless networks, home broadband, and cloud computing.

[0027] Figure 1 This is a flowchart illustrating a computing power network resource allocation method provided in an embodiment of this application. Figure 1 As shown, in some embodiments, the computing power network resource allocation method includes the following steps: S101, based on the general scheduling interface, inputs the task characteristics and performance index constraints of the computing network task into the first model to obtain the combined association result.

[0028] A general scheduling interface is built based on a business decoupling model. In some embodiments, this general scheduling interface can receive key information such as task characteristics and performance constraints and transmit them in a standardized manner, while feeding back the subsequent processing results to the corresponding modules, realizing "separation of business and scheduling, and separation of scheduling and resources", thereby improving interface adaptability and the flexibility of the entire system.

[0029] The first model is a multi-element fusion resource scheduling model for computing power networks. This model integrates key elements related to computing network tasks, computing resources, and performance control across multiple dimensions, enabling systematic matching and computation from task input to associated result output.

[0030] In some embodiments, relying on a general scheduling interface, the core attributes of the computing network task are first used to determine the corresponding computing resource requirements, and then the performance constraints of the task are combined to determine the error control results. Finally, by integrating the above two types of results, a combined correlation result representing the relationship between computing network resource allocation and performance error control is obtained. In one implementation, relying on the general scheduling interface as a basic interaction carrier, the task characteristics of the computing network task are first extracted (including key information related to task execution such as task size, concurrency requirements, and execution algorithms). Based on these core attributes, the range of various computing resource requirements corresponding to the task is clarified (including the requirements for storage, computation, and dedicated processing resources). Then, combined with a preset resource utilization threshold, preliminary computing resource allocation results are formed. At the same time, for the performance constraints proposed by the task, the performance error control target result that can meet the expectation is determined by referring to the preset performance index threshold. Finally, by integrating the above "computing resource allocation results" and "performance error control target results", a combined correlation result that reflects the correspondence between the two and supports further processing is obtained.

[0031] S102, input the combined correlation results into the second model to obtain the time series characteristics between task characteristics and performance index constraints, and determine the dynamic resource allocation results and predicted performance indicators of the computing network task based on the task characteristics, performance index constraints and time series characteristics.

[0032] The second model can be an adaptive resource orchestration and scheduling model (LMFRS) based on Long Short-Term Memory (LSTM) networks, which can effectively capture the temporal dependencies of task requirements. In some embodiments, the second model includes an embedding layer, a sharing layer, and a branched fully connected layer. The design of sharing and branching aims to reduce interference between tasks while improving the accuracy of resource and performance predictions.

[0033] The shared layer captures key patterns in time series features using LSTM, including forget gate, input gate, update gate (cell state), and output gate.

[0034] The branch fully connected layers predict resource requirements and performance metrics, and include an input layer, a fully connected layer, and an output layer. In some embodiments, the branch fully connected layers of a computing network task are determined based on the task success rate and task response latency. Based on each branch fully connected layer, the fully connected network of the computing network task is determined. For example, each task (such as success rate prediction and latency prediction) still has its own specific requirements. By adding task-specific fully connected layers on top of shared features, accurate modeling of specific objectives can be achieved, i.e.: ; ; Each task corresponds to an independent fully connected network (i.e., a branched fully connected layer), with the following structure: Input shared features: ; Branched fully connected layer: ; in, , Let J represent the weights and biases for the j-th task. Let j be the output dimension of the j-th task; Let be the prediction function for the j-th task.

[0035] Predicted values ​​for each task They are calculated and output separately.

[0036] In some embodiments, the results of dynamic resource allocation can allocate appropriate computing resources (including CPU, GPU, memory, storage, etc.) to tasks to meet business needs.

[0037] In some embodiments, performance metrics are predicted, such as the success rate and latency of tasks, to provide feedback for resource optimization.

[0038] By executing steps S101 and S102, intelligent allocation of computing power, network, and storage resources is achieved through dynamic scheduling logic, which can adapt to high-demand scenarios with dynamic and complex allocation of computing power and network resources.

[0039] like Figure 2 As shown, in some embodiments, when performing step S101, the computing power network resource allocation method includes the following steps: S201, Based on the task characteristics, determine the resource combination results corresponding to the computing network task.

[0040] Task characteristics include at least one of the following: task size, number of concurrent task requests, and task algorithm.

[0041] The resource combination results are used to characterize computing resource requirements, including at least one of memory resource requirements, processor resource requirements, image processor resource requirements, and video memory requirements.

[0042] In some embodiments, a suitable combination of resources is allocated to computing tasks: ; in, Indicates the memory resource requirements; Indicates the CPU resource requirements; Indicates the GPU resource requirements; This indicates the video memory requirement.

[0043] S202. Based on the resource combination results and resource utilization threshold, the resource allocation results are calculated.

[0044] To avoid resource waste and ensure resource utilization during tasks, in some embodiments, the computational resource allocation result (i.e., the actual computational resources that need to be allocated) can be calculated based on a required resource utilization threshold. ; in, This indicates the threshold for resource utilization.

[0045] S203, based on performance index constraints and performance index thresholds, obtain the result of minimizing the prediction error.

[0046] In some embodiments, minimizing the prediction error result is used to characterize the correlation between latency and success rate targets. In one implementation, for the two performance metrics of task response latency and task success rate, the error of each individual metric is quantified based on its corresponding threshold (maximum allowable latency value, success rate threshold). Next, weight parameters are introduced to weight and sum the two types of errors, constructing an overall loss function. Finally, by minimizing this loss function, the minimized prediction error result that satisfies the performance metric constraints is obtained. By quantifying single-metric errors, weighting and fusing the importance of multiple metrics, and minimizing the total loss, precise satisfaction of performance constraints is achieved.

[0047] S204. Based on the results of resource allocation and minimizing prediction error, the combined correlation result is obtained.

[0048] like Figure 3 As shown, in some embodiments, performance constraints include task response latency and task success rate, and performance thresholds include a maximum allowed latency value and a success rate threshold. When executing step S203, the computing power network resource allocation method includes the following steps: S301, Based on the task response delay and the maximum allowable delay value, the first association result is obtained.

[0049] In some embodiments, computing tasks must meet the following performance metrics: , in, This is expressed as task response latency (in milliseconds). Expressed as the task success rate (in %).

[0050] In some embodiments, the task response latency is less than or equal to the maximum allowable latency value. ,Right now .

[0051] The first association result could be .in, This is expressed as task response latency (in milliseconds). This is expressed as the maximum allowable delay value; S302, based on the task success rate and success rate threshold, obtain the second correlation result.

[0052] In some embodiments, the task success rate is greater than or equal to a success rate threshold. ,Right now .

[0053] The second association result can be .in, Expressed as task success rate (in %); This represents the success rate threshold.

[0054] S303. Based on the first association result, the first weight, the second association result, and the second weight, the result of minimizing the prediction error is obtained.

[0055] Wherein, the first weight is the weight parameter corresponding to the first association result. The second weight is the weight parameter corresponding to the second association result.

[0056] In some embodiments, the formula for minimizing the prediction error includes: ; in, Indicates model parameters; express The weights; express Weights.

[0057] Understandable, , The value of can be determined based on the computing task. Different computing tasks , The value of can be different, and this application does not impose any restrictions.

[0058] By minimizing the prediction error, the model automatically adjusts its parameters. This approach minimizes the weighted sum of latency and success rate errors. Under performance constraints (maximum latency, minimum success rate), the prediction error is minimized, ensuring that the task's latency and success rate meet preset thresholds as much as possible, ultimately yielding a "minimized prediction error result" that meets business requirements. This is achieved by quantifying the constraint violation errors of individual performance metrics, balancing the importance of multiple metrics through weighting, and finally minimizing the weighted total error, thus ensuring precise satisfaction of performance constraints.

[0059] like Figure 4As shown, in some embodiments, when performing step S102, the computing power network resource allocation method includes the following steps: S401, determine the input feature vector at the current moment based on the task size, the number of concurrent task requests, and the algorithm embedding vector of the task algorithm.

[0060] In some embodiments, task features include discrete features, such as algorithm types. Different algorithms have different complexities and computational requirements. Directly using these discrete categories as numerical inputs (e.g., using one-hot encoding) may lead to dimensionality explosion and semantic loss. For example, if there are many algorithm types, one-hot encoding and other schemes can significantly increase the dimensionality of the input features, thereby increasing computational complexity and model training difficulty. Furthermore, one-hot encoding and other schemes cannot reflect the similarity or distance between categories. For algorithms with similar computational requirements, the model cannot perceive their potential connections. In this embodiment, the embedding layer introduces an embedded mapping method by executing step S401, mapping discrete features to a continuous low-dimensional space, enabling different algorithm types to have semantic similarity in a specific dimension.

[0061] In some embodiments, an embedding layer is a method for transforming high-dimensional sparse features into low-dimensional dense vectors. The embedding layer is a learnable parameter matrix whose values ​​are dynamically updated through model training. For an input class i, the embedding layer maps it to a d-dimensional vector. For category index i, the output vector is: ; Among them, matrix The size is , where is the embedding layer; N represents the total number of categories; d is the dimension of the embedding vector.

[0062] The dimension d of the embedding vector needs to be selected based on the total number of classes N and the task complexity, i.e. .

[0063] In some embodiments, the steps of embedding mapping include inputting discrete features, initializing the embedding matrix, embedding query, and training the embedding vector.

[0064] Input discrete features, index the discrete features (such as algorithm type), and index N algorithms. express.

[0065] Embedded matrix initialization, initial size is Embedding matrix It is initialized using a normal distribution.

[0066] Embedded query: Find the row corresponding to the embedded matrix based on the input index i. This yields a d-dimensional embedding vector.

[0067] Embedded vectors are used in training and are input as features into the subsequent LSTM network. Meanwhile, the embedding matrix is ​​used as part of the model and is optimized through backpropagation.

[0068] In some embodiments, given a computing task, the input feature vector of the computing task... It can be represented as: ; in, Indicates the size of the task to be processed each time; Indicates the number of concurrent task requests; This indicates the algorithm used in the task. It's understandable that different algorithms require different computing resources.

[0069] In some embodiments, the input feature vector The data structure is a tuple, which is an ordered, immutable set of elements. That is, the input feature vector... It can be a packaged set of related data, with elements arranged in a fixed order, and once created, it cannot be modified (e.g., elements cannot be added, deleted, or modified).

[0070] For example, input feature vector It can be (task type, computing power requirement, data volume, deadline, priority). For example, ("AI training", "GPU 4 cards + CPU 16 cores", 100GB, "2024-05-20 18:00", "high"). The order of elements in this tuple is fixed: the first is the task type, the second is the computing power requirement, and the third is the data volume. These cannot be arbitrarily changed. The content of the elements is immutable; for example, "AI training" cannot be changed to "video transcoding," nor can the element "storage requirement" be added. To modify it, a new tuple must be created. This ensures that the position of features is fixed, and the model can accurately identify "which position corresponds to which feature" (for example, the model knows that the second element must be the computing power requirement and will not be confused). It also prevents features from being mistakenly modified during transmission or processing (such as changing "high priority" to "low priority"), ensuring data consistency. In addition, tuples have high storage and access efficiency, making them suitable as input data formats for models, especially when batch processing a large number of tasks.

[0071] S402, based on the input feature vector at the current time step and the hidden state at the previous time step, the forget gate output result is obtained based on the first activation function.

[0072] In some embodiments, the impact of different features (such as task size, number of requests, and algorithm type) on latency and success rate may be highly nonlinear and time-dependent. The main purpose of the shared layer is to extract common time-series features for different tasks by sharing parameters, thereby mitigating the risk of overfitting and improving computational efficiency, which can be expressed as: ; The network structure of the shared layer is as follows: enter: ; in, , represents the input feature vector of the t-th time slot (i.e., the current time).

[0073] LSTM state updates control the transmission and forgetting of information through three "gates": the forget gate, the input gate, and the output gate. At the same time, it is also necessary to maintain two states: the cell state and the hidden state.

[0074] The forgetting gate controls the degree to which information from the previous time step is "forgotten" in a cell's state, that is: ; in: It is the input feature of the current time slot t; It is the hidden state of the previous moment (i.e., the output of the previous moment); and It is the weight matrix of the forget gate; It is the bias term of the forget gate; It is the Sigmoid activation function (i.e., the first activation function), with an output range of (0, 1), representing the degree of forgetting; the output of the forget gate The percentage of "forgotten" indicates that 1 represents complete retention and 0 represents complete forgetting.

[0075] By using the forget gate, this scheme controls the cell state from the previous moment. The part that needs to be retained is used to obtain the output of the forget gate.

[0076] S403, based on the input feature vector at the current time step and the hidden state at the previous time step, and using the second activation function, obtain the input gate output result and the candidate cell state result.

[0077] In some embodiments, the input gate determines how much of the cell state new information should be added to, i.e.: ; ; in: It is the output of the input gate, controlling the degree to which new information is received; This represents the current state of the candidate cells. and It is the weight matrix of the input gate; and It is the weight matrix of the candidate cell states; and These are the corresponding bias terms; It is the Sigmoid activation function. It is the hyperbolic tangent function; input gate The degree to which the information input at the current moment affects the cell state is controlled; candidate cell state This indicates the new candidate information at the current moment.

[0078] In some embodiments, the output of the input gate and candidate cell status Multiplication indicates the degree to which new information is added to the cell state, resulting in the input gate output and candidate cell state results.

[0079] The second activation function includes the Sigmoid activation function and the hyperbolic tangent function.

[0080] S404 combines the output of the forget gate, the cell state at the previous time step, the output of the input gate, and the candidate cell state to obtain the cell state at the current time step.

[0081] In some embodiments, the cell state is the core of the LSTM, storing historical information that has been relied upon for a long time. The update process combines the outputs of the forget gate and the input gate, i.e.: ; in, It is the output of the forget gate; It represents the cell state at the previous moment; It is the output of the input gate; It represents the current state of the cell; This indicates element-wise multiplication.

[0082] Current cell state It represents the cell state at the previous moment. The LSTM is a weighted combination of the forgotten information through the forget gate and the new information controlled by the input gate at the current moment. In this way, the LSTM can retain important information and discard unnecessary content according to the control of the forget gate and the input gate.

[0083] S405, based on the input feature vector at the current time step and the hidden state at the previous time step, the output gate output result is obtained based on the third activation function; and, based on the output gate output result and the cell state at the current time step, the hidden state at the current time step is obtained.

[0084] In some embodiments, the output gate determines the hidden state at the current moment. (i.e., the output of the model), that is: ; ; in, It is the output of the output gate; It is the hidden state (i.e., the output) at the current moment; and It is the weight matrix of the output gate; It is the bias term of the output gate. It is the result of the cell state after passing through the third activation function.

[0085] The output gate controls the cell state. What portion of the data will be used as the hidden state at the current moment? Output the current hidden state. Through output gate control, the cell state is transformed into an output suitable for the next layer of processing.

[0086] S406, Based on the current hidden state, obtain the dynamic resource allocation result.

[0087] In some embodiments, the output Where T represents the sequence length, This is represented as the hidden layer dimension.

[0088] like Figure 5 As shown, in some embodiments, when performing step S401, the computing power network resource allocation method includes the following steps: S501, Obtain historical dataset.

[0089] Historical datasets include business features, target output, embedding dimensions, hidden layer size, and hyperparameters. For example, a historical dataset can be represented as: , in, Including task characteristics, Includes target output; embedding dimension d; hidden layer size h; hyperparameter: learning rate Weight parameters .

[0090] In some embodiments, hyperparameters include at least one of the learning rate and weight parameters.

[0091] S502, normalize the task size and the number of concurrent task requests in the business features to obtain normalized features, and concatenate the normalized features with the algorithm embedding vector of the task algorithm to generate the input feature vector at the current time.

[0092] The normalized features and embedded features are concatenated to generate the input feature vector.

[0093] In some embodiments, the embedding matrix is ​​initialized. LSTM layer parameters, fully connected layer weights and biases , .

[0094] like Figure 6 As shown, in some embodiments, the computing power network resource allocation method further includes the following steps: S601, construct a loss function based on latency loss, success rate loss, and resource demand loss.

[0095] In some embodiments, the total loss is defined as the error in combining resource requirements and performance indicators: ; The delay loss is used to constrain the response time, i.e.: ; Success rate loss is to improve the task success rate, that is: ; Resource demand loss is to reduce resource forecasting errors, i.e.: .

[0096] S602 updates model parameters through backpropagation based on a loss function.

[0097] In some embodiments, backpropagation is used to update parameters. .

[0098] For example, compute the hidden state at each time step and input the hidden state into the multi-task branch layer; repeat the following steps until convergence: sample a batch of data from the dataset: ; Calculate the LSTM hidden state and the prediction value for each task. ; Calculate composite loss Update parameters using backpropagation: For each input sample Input the preprocessed samples into the trained model Obtain the predicted value for each target task. .

[0099] To achieve efficient, stable, and secure scheduling of computing network resources, such as Figure 7As shown, in some embodiments, the layered architecture of the general scheduling interface includes an application programming interface (API) layer, a logic layer, and a data layer. When executing step S101, the computing power network resource allocation method includes the following steps: S701, after the API layer parses and verifies the task characteristics and performance constraints of the received computing network task, it passes it to the logic layer.

[0100] The API layer is used to receive external requests and is responsible for input parsing, parameter validation, and result return.

[0101] The logic layer is used to handle resource prediction, allocation, and dynamic scheduling logic.

[0102] S702, the logic layer inputs the parsed and verified task characteristics and performance index constraints into the first model, and obtains the combined association result of the computing network tasks through the calculation of the first model.

[0103] The data layer is used to store and manage resource status, task information, and scheduling logs.

[0104] S703, the API layer receives the combined association results fed back from the logic layer and returns the combined association results to the outside.

[0105] like Figure 8 As shown, in some embodiments, after a client or user initiates a request to the Application Programming Interface (API) layer, the API layer processes the request sequentially through input validation, authentication, and access control, and then passes the request to the logic layer while returning a response to the client. The task parsing module in the logic layer receives the request and passes it to the resource prediction module. After processing by the resource prediction module, the request flows to the dynamic scheduling module. The dynamic scheduling module obtains the resource status from the real-time resource status management in the data layer, processes it, and passes it to the scheduling optimization module, while also updating the resource status with the real-time resource status management. The scheduling optimization module obtains task data from the task information database in the data layer, processes it, writes the task results to the task information database, writes scheduling logs to the scheduling log record, and simultaneously obtains historical optimization data from the scheduling log record. In the data layer, the scheduling log record stores the scheduling log and provides historical optimization data; the real-time resource status management maintains the resource status and interacts with the dynamic scheduling module; and the task information database stores task data and interacts with the scheduling optimization module. These modules collaborate to complete the entire process from request processing to data management.

[0106] For example, a client or user initiates an operation, such as an enterprise administrator submitting a task request for "bulk sales data statistics and analysis" on the cloud platform interface. The Application Programming Interface (API) layer then initiates the process: the input validation module checks the "data range parameters and format specifications" of the task request to ensure there is no illegal input; the identity authentication module verifies that the user is a legitimate platform user, "Administrator Zhang San," using the "enterprise account token" submitted by the user; after the access control module confirms that the user has the operation permission for "bulk data processing tasks," it passes the request to the logic layer and simultaneously returns a "request accepted" response to the client. The logic layer begins task processing: the task parsing module breaks down the "batch sales data statistics task" into sub-tasks such as "regional data reading, multi-dimensional calculation, and result visualization"; the resource prediction module predicts that "4 computing servers and 3TB of temporary storage resources" are needed to complete this task based on the resource consumption of historical sales data tasks; the dynamic scheduling module obtains the current resource status of "only 2 idle computing servers and 2TB of idle storage" from the real-time resource status management of the data layer, and then adjusts the scheduling strategy and requests temporary expansion of resources; the scheduling optimization module retrieves "resource allocation efficiency records of similar historical sales data tasks" from the task information database, optimizes the resource scheduling of this task, for example, prioritizes the allocation of "high-frequency regional sales calculation sub-tasks" to GPU computing nodes, and writes task results such as "task completion time and result storage path" to the task information database, and writes scheduling logs such as "resource application time and node allocation details" to the scheduling log record. At the same time, it obtains "scheduling bottleneck cases of similar tasks in the past three months" from the scheduling log record to further improve the scheduling strategy. The three modules of the data layer work together to support: real-time resource status management updates resource usage in real time, such as marking a computing server as "occupied" after it is allocated; the task information database stores "the initiator, execution status, and result link of each batch of sales data tasks"; and the scheduling log records "the resource allocation process and anomaly investigation records for each scheduling", providing historical data reference for the scheduling optimization module.

[0107] In this way, through modular design, combined with technologies such as concurrent processing and security protection, the interface is ensured to have versatility, flexibility and reliability in complex environments with large scale and multiple tasks.

[0108] In some embodiments, the computing power network resource allocation method includes the following steps: The S801 general scheduling interface uses a token bucket algorithm to limit the number of calls per second. Furthermore, it performs rate limiting and concurrency control on received network task requests by setting user-level and task-level concurrency limits.

[0109] In some embodiments, the computing power network resource allocation method further includes the following steps: In the case of high-concurrency computing network tasks, the S801 general scheduling interface enables asynchronous execution of computing network tasks through message queues.

[0110] S802, in response to receiving task characteristics and performance constraints, returns the task identity ID.

[0111] S803, in response to the first model operation to obtain the combined association result, feeds back the combined association result to the outside through the query interface.

[0112] S804, the logic layer uses asynchronous input or output ports to input task characteristics and performance index constraints into the first model.

[0113] This application's embodiments consider the dynamic and complex nature of computing network resource allocation and propose a multi-element resource orchestration and scheduling scheme. Combining the diverse characteristics of computing network tasks, it ensures the success rate and average latency of computing network task execution.

[0114] This application's embodiments address the time-dependent nature of computing network task requirements by implementing an LSTM-based adaptive resource orchestration and scheduling model, thereby accurately describing the adaptive orchestration and scheduling problem of computing networks.

[0115] This application embodiment combines core computing network capabilities to construct a general business API capability encapsulation and scheduling interface based on a business decoupling model. Through dynamic scheduling logic, it realizes intelligent allocation of computing power, network and storage resources, and has strong versatility and decoupling.

[0116] Compared to related technologies, computing resource scheduling often employs static or single-dimensional resource matching strategies, which struggle to cope with the diversity and temporal variations of task requirements. The multi-element resource orchestration and scheduling scheme proposed in this application comprehensively considers the multi-dimensional characteristics of computing power, network, and storage resources, and performs precise matching based on task requirement characteristics, significantly improving task execution success rate and reducing average latency.

[0117] Compared to related technologies, scheduling methods often ignore the temporal evolution of task requests, leading to delayed resource allocation or over-reservation. This application introduces an LSTM-based adaptive resource orchestration and scheduling model, which can effectively predict future resource demand trends and dynamically adjust resource allocation strategies, thereby improving scheduling accuracy and response speed. This solves the problems of low resource utilization and poor task execution efficiency caused by the lack of temporal modeling in existing technologies.

[0118] In contrast to related technologies, systems typically couple resource scheduling logic with upper-layer business logic, making it difficult for the system to adapt to different business scenarios. This application's embodiments construct a universal API interface based on a business decoupling model, abstracting resource scheduling logic into a unified interface capability. This supports resource invocation and combination under various business scenarios, achieving decoupling between resource scheduling and business logic, and greatly improving the system's versatility, maintainability, and scalability.

[0119] In some solutions, multiple embodiments of this application can be combined, and the combined solution can be implemented. Optionally, some operations in the processes of each method embodiment may be combined, and / or the order of some operations may be changed. Furthermore, the execution order between the steps of each process is merely exemplary and does not constitute a limitation on the execution order between steps; other execution orders are also possible. It is not intended to indicate that the execution order is the only possible order in which these operations can be performed. Those skilled in the art will conceive of various ways to reorder the operations described herein. In addition, it should be noted that the process details involved in one embodiment of this document are similarly applicable to other embodiments, or different embodiments may be combined.

[0120] Furthermore, some steps in the method embodiments can be equivalently replaced with other possible steps. Alternatively, some steps in the method embodiments may be optional and can be deleted in certain use cases. Or, other possible steps may be added to the method embodiments. Moreover, the various method embodiments can be implemented individually or in combination.

[0121] It should be noted that the foregoing explanation of the method embodiments also applies to the apparatus (such as device nodes and first target network nodes) in the embodiments of this application, and the principle is the same. Therefore, the embodiments of this application are not limited.

[0122] It's important to note that artificial intelligence (AI) is the study of enabling computers to simulate certain human thought processes and intelligent behaviors (such as learning, reasoning, thinking, and planning). It encompasses both hardware and software technologies. AI hardware technologies generally include sensors, dedicated AI chips, cloud computing, distributed storage, and big data processing. AI software technologies primarily include computer vision, speech recognition, natural language processing, machine learning / deep learning, big data processing, and knowledge graph technologies.

[0123] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for allocating computing power network resources, characterized in that, The method includes: Based on the general scheduling interface, the task characteristics and performance index constraints of the computing network task are input into the first model to obtain the combined association result; The combined correlation results are input into the second model to obtain the time series features between the task features and the performance index constraints; and, based on the task features, the performance index constraints and the time series features, the dynamic resource allocation results and predicted performance indicators of the computing network task are determined.

2. The computing power network resource allocation method according to claim 1, characterized in that, The method, based on a general scheduling interface, inputs the task characteristics and performance constraints of the computing network task into the first model to obtain a combined correlation result; including: Based on the task characteristics, the resource combination result corresponding to the computing network task is determined; the task characteristics include at least one of task size, number of concurrent task requests, and task algorithm; the resource combination result is used to characterize computing power resource requirements, including at least one of memory resource requirements, processor resource requirements, image processor resource requirements, and video memory requirements. Based on the resource combination results and resource utilization threshold, the resource allocation results are calculated. Based on the performance index constraints and performance index thresholds, the result of minimizing the prediction error is obtained; The combined association result is obtained based on the computational resource allocation result and the minimized prediction error result.

3. The computing power network resource allocation method according to claim 2, characterized in that, The performance constraints include task response latency and task success rate, and the performance thresholds include a maximum allowable latency value and a success rate threshold. Obtaining the result that minimizes the prediction error based on the performance constraints and performance thresholds includes: Based on the task response latency and the maximum allowed latency value, a first association result is obtained; Based on the task success rate and the success rate threshold, a second correlation result is obtained; The minimized prediction error result is obtained based on the first association result, the first weight, the second association result, and the second weight; wherein, the first weight is the weight parameter corresponding to the first association result; and the second weight is the weight parameter corresponding to the second association result.

4. The computing power network resource allocation method according to claim 2 or 3, characterized in that, The step of determining the dynamic resource allocation result and predicted performance index of the computing network task based on the task characteristics, the performance index constraints, and the time series characteristics includes: The input feature vector at the current moment is determined based on the task size, the number of concurrent task requests, and the algorithm embedding vector of the task algorithm. Based on the input feature vector at the current moment and the hidden state at the previous moment, the forget gate output is obtained based on the first activation function; Based on the input feature vector at the current time and the hidden state at the previous time, the input gate output result and the candidate cell state result are obtained based on the second activation function; The cell state at the current time is obtained by combining the output of the forget gate, the cell state at the previous time step, the output of the input gate, and the candidate cell state. Based on the input feature vector at the current time and the hidden state at the previous time, the output gate output result is obtained based on the third activation function; and based on the output gate output result and the cell state at the current time, the hidden state at the current time is obtained. The dynamic resource allocation result is obtained based on the current hidden state.

5. The computing power network resource allocation method according to claim 4, characterized in that, Determining the input feature vector at the current moment based on the task size, the number of concurrent task requests, and the algorithm embedding vector of the task algorithm includes: Obtain historical datasets; the historical datasets include business features, target output, embedding dimension, hidden layer size, and hyperparameters; The task size and the number of concurrent task requests in the business features are normalized to obtain normalized features. The normalized features are then concatenated with the algorithm embedding vector of the task algorithm to generate the input feature vector at the current time.

6. The computing power network resource allocation method according to claim 4, characterized in that, The method further includes: Construct a loss function based on latency loss, success rate loss, and resource requirement loss; Based on the loss function, the model parameters are updated through backpropagation.

7. The computing power network resource allocation method according to any one of claims 1-3, characterized in that, The layered architecture of the general scheduling interface includes an application programming interface (API) layer, a logic layer, and a data layer. Based on the general scheduling interface, the task characteristics and performance constraints of the computing network task are input into the first model to obtain a combined correlation result, including: The API layer parses and verifies the task characteristics and performance constraints of the received computing network task, and then passes them to the logic layer. The logic layer inputs the parsed and verified task features and performance index constraints into the first model, and obtains the combined association result of the computing network task through the calculation of the first model; The API layer receives the combined association result fed back by the logic layer and returns the combined association result to the outside.

8. The computing power network resource allocation method according to claim 7, characterized in that, The method includes: The general scheduling interface uses a token bucket algorithm to limit the number of calls per second; and it uses user-level and task-level concurrency limits to perform rate limiting and concurrency control on the received computing network task requests.

9. The computing power network resource allocation method according to claim 8, characterized in that, The method further includes: In the case of computing network tasks in high-concurrency scenarios, the general scheduling interface realizes the asynchronous execution of the computing network tasks through message queues; In response to receiving task characteristics and performance constraints, return the task identity ID; In response to obtaining the combined association result from the first model operation, the combined association result is fed back to the outside through a query interface; The logic layer uses asynchronous input or output ports to input the task characteristics and performance index constraints into the first model.

10. A computing power network resource allocation device, characterized in that, It includes an execution unit for performing the method according to any one of claims 1 to 9.

Images