Multi-dimensional dynamic sensing and intelligent computing power scheduling method and device for computing power network

Abstract

The invention discloses a multi-dimensional dynamic sensing and intelligent computing power scheduling method and device oriented to a computing power network. The method comprises the following steps: deploying a plurality of target components forming multi-dimensional dynamic perception at a terminal, an edge node and a regional computing power center, and collecting computing power indexes and network indexes of each node; analyzing indexes in a preset period through a time sequence prediction model, and obtaining computing power bottleneck risk and time delay attenuation information; according to the type of a to-be-executed task, dynamically obtaining a computing power dimension weight and a time delay dimension weight, and combining a computing power bottleneck risk and time delay attenuation information to construct a cost function; determining a node corresponding to the minimum cost function and taking the node as a target node; and obtaining path costs corresponding to all feasible migration paths according to the target node, determining the migration path corresponding to the minimum path cost and taking the migration path as a target migration path, and triggering intelligent scheduling of computing power according to the migration path. According to the method, collaborative perception, intelligent decision and closed-loop control of computing power and network resources can be realized.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the technical field of computer, and particularly relates to a multi-dimensional dynamic perception and intelligent computing power scheduling method and device for computing power network. BACKGROUND

[0002] With the rapid development of 5G / 6G, cloud edge-end collaborative computing and artificial intelligence, computing resources show a large-scale, heterogeneous and distributed trend. In order to realize on-demand access and global scheduling, the industry proposes the concept of computing power network, aiming to unify the dispersed computing, storage and accelerator resources through network arrangement and routing, so that computing power can be discovered, selected and dynamically allocated like network bandwidth, to support low-latency, high-reliability and high-concurrency intelligent business needs. In typical scenarios such as intelligent manufacturing and Internet of Vehicles, tasks have strict constraints on end-to-end latency, jitter and stable computing power supply. The traditional single cloud center processing mode is difficult to meet the needs, and tasks need to be unloaded and migrated between terminals, edge nodes and clouds.

[0003] The existing task offloading and computing power scheduling technology mainly has three paths: local decision based on static rules, centralized global scheduling and programmable network driven service chain arrangement. However, these paths have many problems: in a high dynamic environment, computing power resources and network state are coupled across dimensions. The existing technology collects or evaluates two types of indicators separately, which is easy to misjudge the overall service availability. For diversified task characteristics, the existing scheduling and offloading algorithms use fixed or semi-fixed weight configuration, which is difficult to adapt to task type differences and environmental mutations. The execution control is mainly open-loop execution, and lacks the ability of intervention during runtime. In addition, the existing scheme combining network programmability and path redirection technology has limited support for computing power state prediction, network state synchronous perception and joint decision and linkage execution of the two.

[0004] There are obvious technical problems in task offloading and computing power scheduling in the current computing power network environment, including perception fragmentation and prediction deficiency, lack of joint modeling and forward-looking prediction of computing power load evolution and network quality fluctuation; decision-making adaptability is insufficient, it is difficult to dynamically adjust and optimize the weight and constraint according to different task types and environmental changes; the execution closed-loop capability is insufficient, and there is a lack of rapid back migration, cross-domain switching and path redirection mechanism based on real-time state during runtime. These defects seriously restrict the application effect of computing power network in low-latency and high-reliability businesses. SUMMARY

[0005] Therefore, the embodiment of the present disclosure provides a multi-dimensional dynamic perception and intelligent computing power scheduling method and device for computing power network, which can solve the problems in the existing computing power network task offloading and scheduling technology, such as the separation of computing power and network perception, insufficient computing power load prediction, rigid decision model, lack of closed-loop control in the execution process, and insufficient linkage between programmable network and computing power state, and realize the collaborative perception, intelligent decision and closed-loop control of computing power and network resources.

[0006] In a first aspect, the embodiment of the present disclosure provides a multi-dimensional dynamic perception and intelligent computing power scheduling method for computing power network, comprising: Deploying a plurality of target components constituting multi-dimensional dynamic perception at terminals, edge nodes and regional computing power centers and collecting computing power indicators and network indicators of each node; Analyzing the computing power indicators and the network indicators in a preset period corresponding to all nodes by a time series prediction model to obtain computing power bottleneck risk and time delay attenuation information; According to the type of the task to be executed, dynamically obtaining computing power dimension weight and time delay dimension weight; According to the computing power dimension weight, the time delay dimension weight, the computing power bottleneck risk and the time delay attenuation information, a cost function is constructed; Determining the node corresponding to the minimum cost function as a target node; According to the target node, the path cost corresponding to all feasible migration paths is obtained, and the migration path corresponding to the minimum path cost is determined as a target migration path; According to the target node and the target migration path, intelligent scheduling of computing power is triggered.

[0007] In a second aspect, the embodiment of the present disclosure further provides a computer device, which adopts the following technical solution: The computer device comprises: At least one processor; and A memory in communication connection with the at least one processor; wherein The memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the multi-dimensional dynamic perception and intelligent computing power scheduling method for computing power network as described above.

[0008] In a third aspect, the embodiment of the present disclosure further provides a computer readable storage medium storing computer instructions for causing a computer to execute the multi-dimensional dynamic perception and intelligent computing power scheduling method for computing power network as described above.

[0009] In a fourth aspect, the embodiments of the present disclosure further provide a computer program product comprising computer programs / instructions for implementing the steps of the method of any one of the preceding aspects when executed by a processor.

[0010] The multi-dimensional dynamic perception and intelligent computing power scheduling method for computing power network disclosed in the present application deploys target components at terminals, edge nodes and regional computing power centers to collect computing power indicators and network indicators of each node, changes the situation that the existing technology collects or evaluates the coupling indicators of computing power resources and network states across dimensions separately, can comprehensively and accurately obtain real-time state information of each node in the network through unified collection, and provides a richer and more reliable data basis for subsequent decision-making; uses a time series prediction model to analyze the computing power indicators and network indicators in a preset period to obtain computing power bottleneck risk and time delay decay information, which realizes joint modeling and forward-looking prediction of computing power load evolution and network quality fluctuation, makes up for the deficiencies of the existing technology in prediction, and avoids service interruption or performance degradation caused by sudden computing power bottlenecks or network quality degradation by predicting possible problems in advance; dynamically obtains computing power dimension weight and time delay dimension weight according to the type of the task to be executed, which is different from the existing scheduling and unloading algorithms that use fixed or semi-fixed weight configuration, so that the weight can be flexibly adjusted according to the characteristics and needs of the task, and the decision-making is more in line with the actual situation; constructs a cost function according to the computing power dimension weight, the time delay dimension weight, the computing power bottleneck risk and the time delay decay information, and determines the node corresponding to the minimum cost function as the target node, which is a decision-making method based on the cost function, comprehensively considers multiple factors, dynamically adjusts and optimizes the weight and constraint under different task types and environmental changes, makes the decision-making more scientific and reasonable, and improves the adaptability and accuracy of the decision-making; determines the migration path corresponding to the minimum path cost as the target migration path by calculating the path cost of all feasible migration paths, which provides an optimal path selection for the migration of the task, can select the best migration mode according to the real-time state during the running period, and realizes fast back migration, cross-domain switching and path redirection based on real-time state; triggers intelligent scheduling of computing power according to the target node and the target migration path, forming a closed-loop control link of “perception-prediction-decision-execution”, which makes the system able to monitor state changes in real time during the running process and adjust the scheduling strategy in time according to the changes, avoids the problem that the existing execution control is mostly open-loop execution and lacks intervention ability in the middle, and improves the reliability and stability of the system.

[0011] The above description is only a summary of the technical solutions of the present disclosure. In order to more clearly understand the technical means of the present disclosure, the above description can be implemented according to the content of the specification, and in order to make the above and other purposes, characteristics and advantages of the present disclosure more obvious and easy to understand, the following preferred embodiments are described in detail below, and the accompanying drawings are described as follows. Attached Figure Description

[0012] To more clearly illustrate the technical solutions of the embodiments of this disclosure, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this disclosure. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 This is a flowchart illustrating the multi-dimensional dynamic perception and intelligent computing power scheduling method for computing power networks provided in this embodiment of the disclosure.

[0014] Figure 2 A flowchart illustrating the method for obtaining computing power bottleneck risk and latency attenuation information provided in this embodiment of the disclosure.

[0015] Figure 3 This is a flowchart illustrating the method for dynamically obtaining computing power dimension weights and latency dimension weights provided in this embodiment of the disclosure.

[0016] Figure 4 A flowchart illustrating the method for constructing a cost function provided in this embodiment of the disclosure.

[0017] Figure 5 This is a flowchart illustrating the method for obtaining the target migration path provided in an embodiment of this disclosure.

[0018] Figure 6 This is a schematic diagram of the structure of a computer device provided in an embodiment of the present disclosure. Detailed Implementation

[0019] The embodiments of this disclosure will now be described in detail with reference to the accompanying drawings.

[0020] It should be understood that the following specific examples illustrate the implementation of this disclosure, and those skilled in the art can easily understand other advantages and effects of this disclosure from the content disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of this disclosure, and not all of them. This disclosure can also be implemented or applied through other different specific implementation methods, and the details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this disclosure. It should be noted that, in the absence of conflict, the following embodiments and features in the embodiments can be combined with each other. Based on the embodiments in this disclosure, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure.

[0021] It is important to note that the various aspects described herein are exemplary in nature and are not intended to limit the scope, applicability, or configuration of the disclosure in any way. Rather, the aspects have been presented for the sake of providing an example mode from which numerous alternatives can be derived. Furthermore, it should be understood that the aspects described herein can be implemented in a wide variety of forms and that any specific structure and / or function described herein is merely illustrative. Those skilled in the art will appreciate that one aspect described herein can be implemented independently of any other aspects, and that any listing of aspects does not imply that a combination of aspects must be implemented. Rather, aspects can be combined in any manner consistent with the subject technology.

[0022] It is also important to note that the present disclosure can be carried out in many ways, and that the application of each aspect disclosed herein is not limited to only those implementations that can be described in the detailed description. Rather, this detailed description is intended to describe many embodiments, applications, and alternatives based on the underlying principles and technologies described herein. Accordingly, the terms "exemplary," "by way of example," and / or "for example" as may be used herein in the detailed description and / or in the Examples should be understood as not necessarily limiting, and that the present disclosure can be implemented in a wide variety of ways.

[0023] In addition, in the following description, numerous specific details are provided for a thorough understanding of the examples. One skilled in the relevant art will recognize, however, that the aspects described herein can be practiced without one or more of these specific details.

[0024] Reference Figure 1 A multi-dimensional dynamic perception and intelligent computing power scheduling method for computing power network is disclosed, comprising: S100, deploying a plurality of lightweight target components constituting multi-dimensional dynamic perception at the terminal, edge node and regional computing power center. Synchronously collecting computing power indicators and network indicators of each node based on the lightweight target components.

[0025] The terminal includes a plurality of computing power nodes, the edge node refers to a node with weak resources, and the regional computing power center includes a plurality of high-performance nodes. Specifically, the regional computing power center is a cluster composed of a plurality of high-performance servers. The computing power indicators include CPU utilization, GPU utilization, video memory pressure, queue depth, etc. The network indicators include latency, packet loss rate, bandwidth, jitter, etc.

[0026] In this embodiment, the lightweight target component is preferably a lightweight probe for periodically collecting computing power and network key indicators. Specifically, the probe is implemented in dual mode of containerization and binary: the network probe is based on an improved version of iperf3, with unnecessary logic trimmed and link performance collection of bandwidth, latency, jitter, and packet loss rate completed through a multi-thread mechanism; the resource probe is implemented as a single static binary program in Go language, directly calling Linux cgroups interface and NVIDIA DCGM library to obtain underlying resource indicators such as CPU, GPU, memory, and I / O. The collected data is uploaded through a gRPC streaming interface, combined with Snappy and Delta incremental compression to reduce transmission load, and then input into a time series modeling engine to predict the bottleneck risk and quality decay trend of the future time window.

[0027] This step can realize unified collection and analysis of network indicators, breaking through the limitations of fragmented processing of two types of indicators in the prior art.

[0028] S200, analyzing the computing power indicators and network indicators in the preset period corresponding to all nodes through a time series prediction model to obtain computing power bottleneck risk and latency decay information.

[0029] Through this step, not only can the running state of computing power and network be grasped in real time, but also potential computing power overload and link decay can be identified in advance, so as to provide prospective input for the decision-making module before task offloading, avoiding scheduling failure problems caused by pure reliance on instantaneous state.

[0030] S300, dynamically obtaining computing power dimension weight and latency dimension weight according to the type of the task to be executed. According to the computing power state, the network state, and the computing power bottleneck risk, a cost function is constructed.

[0031] This step establishes a task-adaptive decision-making framework, which can dynamically adjust and optimize weights and constraints according to different task types and environmental changes, realize differentiated and refined resource optimization, and avoid computing power resource mismatch and efficiency reduction caused by fixed weight models.

[0032] S400, determining the node corresponding to the minimum cost function as the target node.

[0033] S500, obtaining the path cost corresponding to all feasible migration paths according to the target node, and determining the migration path corresponding to the minimum path cost as the target migration path.

[0034] S600, triggering intelligent scheduling of computing power according to the target node and the target migration path.

[0035] Further, the application adopts SRv6 technology as the underlying bearer, and through flexible operation on Segment Identifier (SID), closely combines the computing power state with the network routing process, so as to realize the dynamic service chain construction and rapid redirection of the computing power network. In combination with programmable network technologies such as SRv6 in the execution plane, path redirection, service chain orchestration and cross-domain computing power migration are realized, thereby forming a closed-loop control chain of “perception-prediction-decision-execution”.

[0036] The method can perform rapid migration, cross-domain switching and path redirection mechanism based on real-time state in the running period, realize the collaborative perception, decision and execution control of computing power and network resources under complex time-varying conditions, and effectively solve the problem of lack of runtime intervention under the existing open-loop mechanism.

[0037] The multi-dimensional dynamic perception and intelligent computing power scheduling method for the computing power network disclosed in the embodiment provides a combination of multi-dimensional perception, intelligent prediction, adaptive decision and closed-loop control, which can effectively solve the technical problems existing in task offloading and computing power scheduling in the current computing power network environment, and provide strong support for intelligent business needs with low latency, high reliability and high concurrency. In typical scenarios such as intelligent manufacturing and Internet of Vehicles, the strict constraints of end-to-end latency, jitter and stable computing power supply of tasks can be better met, the application effect of computing power network in these scenarios can be improved, and the development of related industries can be promoted.

[0038] In the embodiment, in the computing power dimension, the application focuses on monitoring the GPU and memory state. The GPU utilization is derived by counting the ratio of the number of active stream processors to the total number of processors, and the GPU utilization is : , wherein represents the number of processing units in the running state at time , and is the total number of GPU processing units. The index can accurately reflect the use degree of computing power resources.

[0039] The memory pressure is evaluated by the cache and swap area occupation, and the memory pressure is : , wherein represents the cache occupation, represents the swap area occupation, is the total memory capacity, is the weight coefficient for adjusting the influence of the swap area. The index can reveal the tension degree of the node memory resources.

[0040] In the network dimension, the system collects the core indexes such as bandwidth utilization, latency, packet loss rate, etc. The bandwidth utilization is : , wherein, is the actual throughput, is the link rated bandwidth, which can reflect the link congestion.

[0041] Further, S100 is preferably integrated in a multi-dimensional dynamic perception module, S200-S500 are preferably integrated in a task adaptive decision module, and S600 is preferably integrated in a programmable control execution module, and the present application can effectively improve the deficiencies in the task offloading process under the existing computing power network environment by constructing a complete system of multi-dimensional dynamic perception, task adaptive decision and programmable control execution. In terms of perception and prediction, the present application can simultaneously obtain multiple types of key indicators of computing power and network, and model their future trends, thereby avoiding scheduling failure caused by relying only on instantaneous state, and the system has the ability of forward-looking computing power bottleneck identification and network quality degradation judgment, providing reliable basis for subsequent decision. In terms of decision mechanism, the present application can dynamically adjust the weights of computing power, delay and other indicators in the comprehensive cost function according to the differentiated needs of different task types, realize the flexibility and adaptability of resource allocation, and avoid the mismatch of computing power resources caused by traditional fixed weight model. In terms of execution control, the present application provides closed-loop regulation and control capability, so that the system can trigger back migration, migration or cross-domain switching in time according to the changes of computing power or network state during task execution, ensuring the effectiveness and stability of offloading decision in dynamic environment.

[0042] Referring to Figure 2 For the method of S200 "analyzing the computing power indicators and network indicators of all node pairs in the preset period by the time series prediction model to obtain computing power bottleneck risk and delay attenuation information", the method for obtaining computing power bottleneck risk and delay attenuation information specifically includes: S210, normalizing the computing power indicators and network indicators of all node pairs in the preset period to obtain a normalized data set.

[0043] Specifically, the above multi-dimensional indicators are integrated into a node state vector after normalization and abnormal correction. Wherein, is the corresponding round-trip delay, is the corresponding packet loss rate.

[0044] At the feature level, the node state vector is constructed by multi-dimensional indicator fusion, which uniformly expresses the computing power and network features in the same vector space, ensuring the comparability of indicators of different dimensions.

[0045] S220, identify and correct outliers in the normalized data set to obtain a target data set.

[0046] Specifically, the improved boxplot method (IQR = 1.8) is used to identify and modify outliers in the normalized dataset to avoid bias in the prediction results caused by extreme samples. Specifically, Python language is used with the help of numpy and pandas libraries to complete this task, which includes: 1) Import the two important libraries, numpy and pandas, using Python language. The numpy library is mainly used for efficient numerical calculation, while the pandas library is good at data reading, processing and analysis, laying the foundation for subsequent operations; 2) Write a function named modified_boxplot, which receives two parameters: one is the input one-dimensional dataset, and the other is the interquartile range coefficient (IQR), with a default value of 1.8. The specific operation inside the function is as follows: 2.1) Calculate the quartiles: use the percentile function of the numpy library to calculate the first quartile (Q1) and the third quartile (Q3) of the dataset. The first quartile represents the 25% position in the dataset, and the third quartile represents the 75% position. 2.2) Calculate the interquartile range (IQR): subtract the first quartile from the third quartile to get the interquartile range (IQR), which reflects the distribution range of the middle 50% of the data. 2.3) Determine the upper and lower limits: according to the given interquartile range coefficient (here is 1.8), calculate the lower limit and upper limit respectively. The lower limit is the first quartile minus 1.8 times the interquartile range, and the upper limit is the third quartile plus 1.8 times the interquartile range. 2.4) Identify and modify outliers: use the where function of numpy to replace the values less than the lower limit with the lower limit and the values greater than the upper limit with the upper limit, which completes the identification and modification of outliers. Finally, return the modified dataset. 3) Load the normalized data: use the read_csv function of the pandas library to load the normalized dataset. It should be noted that the file path should be replaced with the path of your own normalized dataset file. 4) Apply the improved boxplot method to process the data: iterate through each column of the normalized dataset and call the modified_boxplot function defined earlier for each column of data. In this way, each column of the dataset is identified and modified for outliers, ensuring that the entire dataset is processed for outliers. 5) Obtain the target dataset: after modifying the outliers in each column of data, the resulting dataset is the final target dataset we need, which has removed the extreme samples that may affect the prediction results and is more suitable for subsequent analysis and prediction tasks.

[0047] S230, input the target dataset into the LSTM-based time series prediction model to obtain the computing power bottleneck risk and latency decay information.

[0048] For S230, specifically comprising the following steps: inputting the target data set into the LSTM-based time series prediction model, obtaining corresponding computing power risk probability, time delay attenuation information.

[0049] The computing power risk probability is : ; .

[0050] Wherein, , is a threshold value, is a parameter for controlling the steepness of the function, is the total number of processing units of the GPU.

[0051] The time delay attenuation information is : , wherein, is the predicted round-trip time delay, is the historical maximum time delay, is the predicted packet loss rate at , and is the weight of the packet loss item.

[0052] Further, it can also be judged whether the computing power risk probability and the time delay attenuation information are greater than the corresponding set threshold value, if yes, it is determined that the corresponding node has a potential computing power bottleneck; if not, it is determined that the corresponding node does not have a potential computing power bottleneck. In this application, it is preferred to select in the node without a potential computing power bottleneck.

[0053] In order to predict the future computing power and network evolution, this embodiment preferentially adopts an LSTM-based time series prediction model, takes the historical state sequence as input, and outputs the state estimation at the future time, i.e. the associated information of the computing power bottleneck risk and the network quality attenuation trend is : . Wherein, represents the prediction time window, represents the historical observation length, the prediction result can not only give the future GPU utilization and network time delay, but also can calculate the potential risk of computing power and network.

[0054] Specifically, in order to model the time series dependence, the present application introduces a double-layer LSTM structure, and assists with an Attention mechanism to enhance the weight recognition ability of key features; in the prediction stage, not only the numerical values such as GPU utilization and network time delay at the future time are output, but also the risk probability distribution is given. In the quantitative level, the present application respectively defines the computing power bottleneck risk probability and the network attenuation index , the former maps GPU utilization to risk probability through a sigmoid function, and the latter combines predicted latency and packet loss rate to evaluate link degradation trend. The above analysis methods not only guarantee the numerical accuracy of the prediction results, but also provide interpretable risk quantification indicators, providing an operable basis for the decision module to generate the optimal offloading strategy.

[0055] With reference to Figure 3 For the method of "obtaining the computing power dimension weight and the latency dimension weight dynamically according to the type of the task to be executed" in S300, the method of dynamically obtaining the computing power dimension weight and the latency dimension weight specifically includes: A100, determining a preset network delay weight and a preset computing power weight according to the type of the task to be executed.

[0056] Specifically, the type of the task to be executed includes real-time rendering, AI inference, data batch processing, etc. For different types of tasks, the preset weight is dynamically adjusted, which can provide a reasonable initial basis for subsequent weight allocation from the root of task demand, so that the scheduling scheme is more in line with the actual characteristics of the task; preferably, the higher the demand for the current task type to an indicator, the higher the corresponding preset weight is adjusted to the maximum weight.

[0057] A200, obtaining the environmental state sensitivity of each node based on the computing power utilization and the network utilization of each node.

[0058] The running environment of each node in the computing power network is dynamically changing, and the computing power and network utilization of the node will affect its processing capacity for tasks. Considering the environmental state sensitivity can make the determination of the weight more in line with the actual running situation of the node, and avoid making unreasonable scheduling decisions when the node resources are tight.

[0059] A300, determining a latency dimension score according to the preset network delay weight and the environmental state sensitivity.

[0060] A400, determining a computing power dimension score according to the preset computing power weight and the environmental state sensitivity.

[0061] Steps A300 and A400 respectively determine the latency dimension score and the computing power dimension score according to the preset weight and the environmental state sensitivity. This scoring method that comprehensively considers the preset weight and environmental factors can comprehensively measure the demand of the task in different dimensions and the actual situation of the node.

[0062] A500, determining a latency dimension weight and a computing power dimension weight according to the latency dimension score and the computing power dimension score.

[0063] The latency dimension weight is : ; The computing power dimension weight is : ; wherein, is a latency dimension score, is a computing power dimension score.

[0064] The dynamic weight obtained by the method comprehensively considers the task characteristic parameters and the environmental sensitivity, which not only avoids the uncertainty of manual parameter tuning, but also automatically adjusts when the task changes or the network fluctuates. Compared with the traditional fixed or semi-fixed weight configuration, it can better adapt to diversified task characteristics and complex and variable environments, improve the flexibility and accuracy of task offloading and computing power scheduling, and thus more effectively meet the requirements of different tasks for computing power and latency, and improve the overall performance and service quality of the computing power network.

[0065] For the method of A200 "obtaining the environmental state sensitivity of each node based on the computing power utilization rate and network utilization rate of each node", specifically comprising: A210, obtaining the change rate of the computing power utilization rate of each node ; , wherein, is the sum of the CPU utilization rate and the GPU utilization rate of the node at time t, is a preset time interval.

[0066] A220, obtaining the change rate of the network utilization rate of each node .

[0067] ; ; , wherein, is the bandwidth corresponding to time t, is the bandwidth corresponding to time t, is the total bandwidth of the corresponding link, is the bandwidth utilization rate from time t to time t.

[0068] A230, obtaining the environmental state sensitivity of each node according to the change rate of the computing power utilization rate and the change rate of the network utilization rate of each node .

[0069] .

[0070] In the computing power network, the resource usage of the node is dynamically changing, and the simple utilization value can only reflect the resource occupation degree at the current moment, while the change rate can reflect the dynamic trend of resource usage. For example, although the current computing power utilization rate of a node is not high, the change rate is very fast, which indicates that the resource usage of the node is changing rapidly and may face a resource shortage situation soon. By focusing on the change rate, the change of the node resource state can be perceived in advance, and more forward-looking information can be provided for subsequent decision-making.

[0071] The change rate of the computing power utilization rate and the change rate of the network utilization rate are multiplied to obtain the environment state sensitivity. This calculation method comprehensively considers the two key resource factors of computing power and network; in the actual computing power network, computing power and network resources are interrelated and interdependent, and the execution of a task not only needs computing power support, but also depends on a good network environment. Considering the change of computing power or network alone is not comprehensive, and by multiplying the change rates of the two, a comprehensive index can be obtained to measure the environment state sensitivity of the node, and the overall change of the node resource state can be more accurately reflected.

[0072] The environment state sensitivity can provide an important basis for dynamically determining the computing power dimension weight and the time delay dimension weight. When the environment state sensitivity of the node is high, it indicates that the resource state of the node changes dramatically, and more cautious consideration of the resource allocation of the node is needed when scheduling tasks, and the weights of the computing power and time delay dimensions need to be adjusted appropriately to avoid allocating tasks to nodes with unstable resources, thereby improving the accuracy and reliability of task offloading and computing power scheduling and ensuring the stable operation of the computing power network.

[0073] Reference Figure 4 For the method of "constructing a cost function according to the computing power state, the network state, and the computing power bottleneck risk" in S300, the construction method of the cost function includes: B100, obtaining the predicted network time delay corresponding to each node; B200, determining the cost function of the corresponding node according to the network time delay, the computing power bottleneck risk, the computing power dimension weight, and the time delay dimension weight.

[0074] The cost function of the i-th node is : .

[0075] Wherein, is the time delay dimension weight, is the computing power dimension weight, is the computing power bottleneck risk corresponding to the i-th node.

[0076] ​​​In step B100, the predicted network delay corresponding to each node is obtained. Network delay is one of the key indicators to measure network performance, especially for tasks with high real-time requirements (such as real-time rendering, online games, etc.), lower network delay is crucial. By obtaining the predicted network delay, the network transmission delay of each node in the future can be known in advance, not just relying on the current delay data, which helps to select nodes with smaller network delay in task scheduling, thereby ensuring the real-time and smoothness of the task.

[0077] In step B200, the cost function of the corresponding node is determined according to the network delay, the risk of computing power bottleneck, the weight of computing power dimension and the weight of delay dimension. This comprehensive consideration of multiple factors can comprehensively and objectively evaluate the comprehensive state of each node. The risk of computing power bottleneck reflects the possible problems of the node in computing power resources, and the network delay reflects the performance of network transmission. Considering both of them can avoid focusing only on network performance and ignoring computing power bottleneck, or only focusing on computing power and ignoring network delay, ensuring that the selected node can meet the task requirements in computing power and network.

[0078] The weight of computing power dimension and the weight of delay dimension are dynamically determined according to the type of task to be executed, so that the cost function can be adjusted according to the characteristics of different tasks, which can make the cost function more suitable for the actual needs of the task and improve the pertinence and effectiveness of task scheduling.

[0079] The cost function provides a quantitative index for task scheduling, which can easily determine the most suitable node for executing the task by comparing the cost function values of each node, and realize the optimal allocation of resources. In a complex computing power network environment, this scheduling method based on cost function can improve resource utilization, reduce task execution cost, and improve the operation efficiency and service quality of the entire network.

[0080] Referring to Figure 5 For the method of S500 "obtaining the path cost corresponding to all feasible migration paths according to the target node, determining the migration path corresponding to the minimum path cost as the target migration path", the method of obtaining the target migration path specifically includes: S510, obtaining the network delay corresponding to each feasible migration path and the processing overhead corresponding to the execution path analysis operation.

[0081] The network delay reflects the time required for data transmission during migration, and the processing overhead represents the computing resources consumed by operations such as path resolution. Different migration paths may differ significantly in these two aspects. For example, some paths may have low network delay but complex path resolution operations and high processing overhead, while other paths may have fast network transmission but time-consuming resolution processing. By considering both factors, the overall cost of each migration path can be comprehensively measured, avoiding the selection of a path that is not optimal due to focusing on a single factor.

[0082] S520, obtaining the path cost corresponding to the feasible migration path according to the network delay and the processing overhead.

[0083] The path cost corresponding to the feasible migration path is : ; The network delay of the migration path is denoted as T, the processing overhead caused by SID header resolution is denoted as P (default value: 80-120 microseconds), the network delay overhead weight is denoted as w1, and the processing overhead weight is denoted as w2. Through this cost model, the system can consider both network transmission delay and SID processing delay when selecting a path, avoiding the problem of traditional schemes that only focus on link delay and ignore programmable processing cost.

[0084] This step obtains the path cost of the corresponding feasible migration path according to the network delay and the processing overhead. This quantitative method integrates the two different dimensions of network delay and processing overhead into a unified path cost, making different paths comparable and providing a clear basis for subsequent path selection.

[0085] S530, sorting the path costs of all feasible migration paths in ascending order, determining the migration path corresponding to the smallest path cost as the target migration path.

[0086] This step sorts the path costs of all feasible migration paths in ascending order and selects the migration path corresponding to the smallest path cost as the target migration path. Among the many feasible paths, this sorting and selection mechanism can quickly and accurately find the migration path with the lowest cost. This helps to improve migration efficiency, reduce resource consumption and time cost during migration, and improve the overall performance of the system.

[0087] ​​​​Selecting the migration path with the minimum path cost means that the task can be completed with the least resource consumption during migration, reducing network latency can reduce the time cost of data transmission, reducing processing overhead can save computing resources, thereby improving the resource utilization efficiency of the entire system; the optimized migration path can reduce the probability of problems occurring during migration, such as data loss, transmission errors, etc., a stable migration process helps to maintain the normal operation of the system, avoids system failure or service interruption caused by migration problems, and improves the reliability and stability of the system.

[0088] Further, when it is monitored that the selected target node state deteriorates significantly, the system can immediately re-evaluate the cost function and trigger back migration or cross-domain switching, thereby realizing real closed-loop control. Specifically, when the computing power risk probability of the node performing the task at a future moment is greater than a preset threshold, immediate re-evaluation and dynamic regulation are triggered, so that the unloading decision is no longer dependent on a single static judgment, but can be continuously optimized during operation to ensure stable operation of the task in a dynamic environment.

[0089] Further, in order to ensure the stability of the configuration, the application also includes a consistency checking mechanism, which periodically compares the SID configuration issued by the controller with the actual configuration in the forwarding device, and detects the offset situation by using differential calculation, and immediately issues a repair instruction when the difference exceeds a threshold, ensuring that the configuration and execution remain strongly consistent. The application not only can complete the adjustment of the task path within milliseconds, but also can realize the dynamic migration of computing power resources in a cross-domain environment, thereby providing reliable execution guarantee for intelligent unloading in the computing power network environment.

[0090] The computer device according to the embodiments of the present disclosure includes a memory and a processor. The memory is configured to store non-transitory computer readable instructions. Specifically, the memory can include one or more computer program products, which can include various forms of computer readable storage media, such as volatile memory and / or non-volatile memory. The volatile memory may, for example, include random access memory (RAM) and / or cache memory, etc. The non-volatile memory may, for example, include read-only memory (ROM), hard disk, flash memory, etc.

[0091] The processor can be a central processing unit (CPU) or other forms of processing units with data processing capability and / or instruction execution capability, and can control other components in the computer device to perform desired functions. In an embodiment of the present disclosure, the processor is configured to run the computer readable instructions stored in the memory, so that the computer device performs all or part of the steps of the multi-dimensional dynamic perception and intelligent computing power scheduling method for computing power network according to the embodiments of the present disclosure. ​

[0092] Those skilled in the art should understand that, in order to solve the technical problem of how to obtain a good user experience effect, the embodiment can also include well-known structures such as a communication bus, an interface, and the like, which should also be included in the protection scope of the present disclosure.

[0093] As Figure 6 A structural schematic diagram of a computer device provided by the embodiment of the present disclosure is shown. The structural schematic diagram shows a structure suitable for implementing the computer device in the embodiment of the present disclosure. Figure 6 The computer device shown is only an example and should not impose any limitation on the functions and use range of the embodiment of the present disclosure.

[0094] As Figure 6 As shown, the computer device can include a processor (for example, a central processor, a graphics processor, and the like), which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) or a program loaded from a storage device into a random access memory (RAM). In the RAM, various programs and data required for the operation of the computer device are also stored. The processor, the ROM, and the RAM are connected to each other through a bus. An input / output (I / O) interface is also connected to the bus.

[0095] Generally, the following devices can be connected to the I / O interface: an input device including, for example, a sensor or a visual information collection device; an output device including, for example, a display screen; a storage device including, for example, a magnetic tape, a hard disk, and the like; and a communication device. The communication device can allow the computer device to communicate with other devices (such as an edge computing device) wirelessly or by wire to exchange data. Although Figure 6 The computer device with various devices is shown, but it should be understood that it is not required to implement or have all the devices shown. More or fewer devices can be alternatively implemented or provided.

[0096] In particular, according to the embodiment of the present disclosure, the processes described above with reference to the flowcharts can be implemented as a computer software program. For example, the embodiment of the present disclosure includes a computer program product including a computer program carried on a non-transitory computer readable medium, the computer program containing program code for executing the method shown in the flowchart. In such an embodiment, the computer program can be downloaded and installed from a network through the communication device, or installed from the storage device, or installed from the ROM. When the computer program is executed by the processor, all or part of the steps of the power-oriented network multi-dimensional dynamic perception and intelligent power scheduling method of the embodiment of the present disclosure are performed.

[0097] The detailed description of the embodiment can refer to the corresponding description in the foregoing embodiments, which will not be repeated here.

[0098] A computer readable storage medium according to embodiments of the present disclosure has non-transitory computer readable instructions stored thereon. When the non-transitory computer readable instructions are run by a processor, all or part of the steps of the method for multi-dimensional dynamic perception and intelligent computing power scheduling of computing power network according to the embodiments of the present disclosure are performed.

[0099] The computer readable storage medium described above includes, but is not limited to, an optical storage medium (for example, CD-ROM and DVD), a magneto-optical storage medium (for example, MO), a magnetic storage medium (for example, magnetic tape or a mobile hard disk), a medium with a built-in rewritable non-volatile memory (for example, a memory card), and a medium with a built-in ROM (for example, a ROM cartridge).

[0100] The detailed description of the present embodiment can refer to the corresponding description in the foregoing embodiments, which will not be described here.

[0101] The basic principles of the present disclosure are described above in combination with specific embodiments, but it should be noted that the advantages, advantages, effects and the like mentioned in the present disclosure are only examples and not limitations, and these advantages, advantages, effects and the like cannot be considered as the must-have of each embodiment of the present disclosure. In addition, the specific details of the above disclosure are only for the purpose of example and for the purpose of understanding, and the above details do not limit the present disclosure to the must-use specific details.

[0102] In the present disclosure, the relationship terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between the entities or operations. The block diagrams of devices, apparatuses, equipment, systems involved in the present disclosure are only illustrative examples and are not intended to require or imply the connection, arrangement, configuration shown in the block diagram. As those skilled in the art will recognize, these devices, apparatuses, equipment, systems can be connected, arranged, configured in any manner. Words such as "include", "contain", "have" and the like are open-ended words, which mean "including but not limited to", and can be used interchangeably. The words "or" and "and" used herein mean the word "and / or", and can be used interchangeably unless the context clearly indicates otherwise. The word "such as" used herein means the phrase "such as but not limited to", and can be used interchangeably.

[0103] In addition, as used herein, "or" used in the list of items starting with "at least one" indicates separate listing, so that, for example, the list of "at least one of A, B or C" means A or B or C, or AB or AC or BC, or ABC (i.e. A and B and C). In addition, the phrase "exemplary" does not mean that the described example is preferred or better than other examples.

[0104] It is also important to note that the systems and methods of the present disclosure can be embodied in a variety of contexts. For example, the systems and methods of the present disclosure can be implemented in the context of a computer system, a mobile device, a server, a network, a distributed computing environment, etc. It is also important to note that the various components and steps of the systems and methods of the present disclosure can be decomposed and / or recombined. Such decompositions and / or recombinations should be considered equivalents of the present disclosure.

[0105] Various changes, modifications, and alterations to the techniques described herein can be made without departing from the teachings of the technology defined by the appended claims. In addition, the scope of the claims of the present disclosure is not limited to the specific aspects described above. The presently existing or later developed processes, machines, manufactures, compositions of matter, means, methods, or steps that perform substantially the same function or achieve substantially the same results as those described herein can be utilized according to the teachings of the claims. Accordingly, the appended claims are intended to cover all such processes, machines, manufactures, compositions of matter, means, methods, or steps.

[0106] The above description of the disclosed aspects is given to enable any person skilled in the art to make or use the disclosure. Various modifications to these aspects will be readily apparent to those skilled in the art, and the generic principles defined herein can be applied to other aspects without departing from the scope of the disclosure.

[0107] The above description has been presented for the purpose of illustration and description. Furthermore, this description is not intended to limit the embodiments of the disclosure to forms disclosed herein. Although various example aspects and embodiments have been discussed above, those of skill in the art will recognize certain modifications, permutations, alterations, additions and sub-combinations thereof.

Claims

1. A multi-dimensional dynamic sensing and intelligent computing power scheduling method for computing power networks, characterized in that, include: Several target components constituting multi-dimensional dynamic perception are deployed at terminals, edge nodes, and regional computing centers, and computing power indicators and network indicators of each node are collected. By analyzing the computing power indicators and network indicators of all nodes within a preset period using a time-series prediction model, information on computing power bottleneck risks and latency decay can be obtained. Based on the type of task to be executed, dynamically obtain the weights of computing power and latency. A cost function is constructed based on the computing power dimension weight, the latency dimension weight, the computing power bottleneck risk, and the latency attenuation information; The node corresponding to the minimum cost function is determined and designated as the target node; Based on the target node, obtain the path costs corresponding to all feasible migration paths, determine the migration path corresponding to the minimum path cost, and use it as the target migration path. Intelligent scheduling of computing power is triggered based on the target node and the target migration path.

2. The multi-dimensional dynamic perception and intelligent computing power scheduling method for computing power networks according to claim 1, characterized in that, The step involves analyzing the computing power indicators and network indicators for all nodes within a preset period using a time-series prediction model to obtain information on computing power bottleneck risks and latency degradation, including: The computing power indicators and network indicators of all nodes within a preset period are normalized to obtain a normalized dataset. Identify and trim outliers in the normalized dataset to obtain the target dataset; The target dataset is input into an LSTM-based time series prediction model to obtain information on computing power bottleneck risks and latency degradation.

3. The multi-dimensional dynamic perception and intelligent computing power scheduling method for computing power networks according to claim 2, characterized in that, The step of inputting the target dataset into an LSTM-based time series prediction model to obtain information on computing power bottleneck risks and latency degradation includes: The target dataset is input into an LSTM-based time series prediction model to obtain... The corresponding computing power risk probability and latency decay information; The computing power risk probability is : ， ； in, , For the threshold, The parameter controls the kurtosis of the function. This represents the total number of processing units in the GPU. The time delay attenuation information is : ,in, For the predicted round-trip delay, This represents the longest delay in history.

4. The multi-dimensional dynamic perception and intelligent computing power scheduling method for computing power networks according to claim 3, characterized in that, The step of dynamically obtaining computing power dimension weights and latency dimension weights based on the type of task to be executed includes: Determine the preset network latency weight and preset computing power weight based on the type of task to be executed; Based on the computing power utilization and network utilization of each node, the environmental state sensitivity of each node is obtained; The latency dimension score is determined based on the preset network latency weight and the environmental state sensitivity. The computing power dimension score is determined based on the preset computing power weight and the environmental state sensitivity. The latency dimension weight and the computing power dimension weight are determined based on the latency dimension score and the computing power dimension score. The weight of the delay dimension is: : ; The weight of the computing power dimension is: : ;in, Scoring is done based on latency. Scoring is based on computing power.

5. The multi-dimensional dynamic perception and intelligent computing power scheduling method for computing power networks according to claim 4, characterized in that, The method of obtaining the environmental state sensitivity of each node based on the computing power utilization and network utilization of each node includes: Obtain the rate of change of computing power utilization for each node; The rate of change of computing power utilization is : ,in Let be the sum of the CPU utilization and GPU utilization of the node at time t. This is a preset time interval; Obtain the rate of change of network utilization for each node; The rate of change of network utilization is : ； in, From Maximum bandwidth utilization from time t to time t From Minimum bandwidth utilization from time t to time t; The environmental state sensitivity of each node is obtained based on the rate of change of computing power utilization and the rate of change of network utilization. Environmental state sensitivity is : .

6. The multi-dimensional dynamic perception and intelligent computing power scheduling method for computing power networks according to claim 4, characterized in that, The step of constructing a cost function based on the computing power dimension weight, the latency dimension weight, the computing power bottleneck risk, and the latency attenuation information includes: Obtain the predicted network latency for each node; Based on the network latency, the computing power bottleneck risk, the computing power dimension weight, and the latency dimension weight, determine the cost function for the corresponding node; No. The cost function for each node is: : ,in, As the weight of the latency dimension, Weights are based on computing power. For the first Nodes The corresponding computing power bottleneck risk.

7. The multi-dimensional dynamic perception and intelligent computing power scheduling method for computing power networks according to claim 6, characterized in that, The step of obtaining the path costs corresponding to all feasible migration paths based on the target node, and determining the migration path corresponding to the minimum path cost as the target migration path includes: Obtain the network latency and processing overhead for each feasible migration path and the path resolution operation. The path cost corresponding to the feasible migration path is obtained based on the network latency and the processing overhead. No. The path cost corresponding to each feasible migration path is : ; For network latency of the migration path, The processing overhead incurred by SID header parsing. Weighting network latency overhead To handle overhead weights; Sort the path costs of all feasible migration paths in ascending order, determine the migration path with the lowest path cost, and use it as the target migration path.

8. A computer device, characterized in that, The computer device includes: At least one processor; and, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, which, when executed by the at least one processor, enables the at least one processor to perform the multi-dimensional dynamic perception and intelligent computing power scheduling method for computing power networks as described in any one of claims 1-7.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing the computer to execute the multi-dimensional dynamic perception and intelligent computing power scheduling method for computing power networks as described in any one of claims 1-7.

10. A computer program product comprising computer instructions, characterized in that, When executed by a processor, the computer instructions implement the steps of the method according to any one of claims 1-7.

Images