Edge computing task dynamic migration method supporting service continuity

Abstract

The invention discloses an edge computing task dynamic migration method supporting service continuity, which comprises the following steps of: anchoring an original edge node operated by a computing task, and marking M first candidate nodes to be migrated of the computing task; obtaining a static attribute vector corresponding to the node index number; screening Q second candidate nodes from the M first candidate nodes; predicting a plurality of dynamic operation parameters of the calculation task in the Q second candidate nodes; calculating the dynamic operation quality and Euclidean distance of the Q second candidate nodes; generating a distance node sequence and a quality node sequence according to the Euclidean distance and the dynamic operation quality; aligning the distance node sequence and the quality node sequence based on the descending order number, generating a node pair sequence, and determining an optimal migration node in the node pair sequence; migrating the calculation task to the optimal migration node; according to the method, dual condition fusion of migration node judgment is realized, and the service continuity of the calculation task in the dynamic migration process is guaranteed.

Description

Technical Field

[0001] The present invention relates to the dynamic migration of edge tasks, and specifically to a method for dynamically migrating edge computing tasks that supports service continuity. Background Art

[0002] With the wide deployment of edge computing technology, the low-latency and high-bandwidth collaboration between terminal devices and edge nodes has become a key support for ensuring service quality. During actual operation, edge nodes may experience a decline in operating performance due to factors such as resource load fluctuations, network state deterioration, or hardware degradation, affecting the stable execution of computing tasks and causing service interruptions or response delays. Therefore, how to achieve the dynamic migration of computing tasks while ensuring service continuity is the core technical challenge in edge computing systems.

[0003] Existing edge computing task migration solutions mainly determine migration nodes based on a single metric. One type of method ranks candidate nodes according to their operating quality and selects nodes with higher scores for task migration. Although such methods can improve computing performance to a certain extent, they are insufficient in evaluating the relative relationship between the operating state of candidate nodes and the original degraded nodes, and may migrate tasks to nodes with seemingly high quality but still close to the degraded area in terms of state, affecting the effectiveness of migration. Another type of method determines based on the state difference between nodes and degraded nodes. For example, a distance metric is constructed using communication delay, network packet loss rate, operating anomaly metrics, etc., and the node with the farthest state distance is selected as the migration target. Such methods can avoid the degraded area to a certain extent, but without a comprehensive measurement of node computing power and service quality, they may lead to task migration to resource-constrained nodes, instead reducing the overall service performance.

[0004] In summary, the existing technologies still have the following problems in edge computing task migration: lack of unified consideration of "node operating quality" and "node state difference", making it difficult to balance the dual goals of service performance and degradation avoidance; especially in the case of inconsistent sorting results, there is a lack of effective decision-making basis to determine the optimal migration node. Summary of the Invention

[0005] In view of the deficiencies of the prior art, the present invention provides a method for dynamically migrating edge computing tasks that supports service continuity, which realizes the screening of the optimal migration node under dual constraints through the determination of the optimal migration node by double-sequence fusion alignment; at the same time, the technical problems raised in the background art are solved by introducing a shared node index number and the fallback of the third candidate node.

[0006] To achieve the above objectives, the present invention is realized through the following technical solutions: A method for dynamically migrating edge computing tasks that supports service continuity, comprising: S1. Anchor the original edge node where the computing task runs, and mark M first candidate nodes to which the computing task is to be migrated; S2. Assign unique node index numbers to the M first candidate nodes and the original edge node, and obtain the static attribute vectors corresponding to the node index numbers; S3. Screen Q second candidate nodes from the M first candidate nodes according to the static attribute vectors corresponding to the node index numbers; S4. Concatenate the static attribute vectors of the second candidate nodes with the normalized features of the computing task as the combined input of the parameter regression model to predict several dynamic operation parameters of the computing task on the Q second candidate nodes; S5. Calculate the dynamic operation quality of the Q second candidate nodes according to the predicted several dynamic operation parameters, and calculate the Euclidean distance between the second candidate nodes and the original edge node; S6. Sort the Q second candidate nodes in descending order according to the Euclidean distance and the dynamic operation quality respectively to generate a distance node sequence and a quality node sequence; Among them, both the distance node sequence and the quality node sequence are sorted in descending order, and each second candidate node is assigned a descending number that decreases in order from front to back; S7. Align the distance node sequence and the quality node sequence based on the descending numbers to generate a node pair sequence, and determine the optimal migration node in the node pair sequence; Among them, the node pair sequence includes Q binary node pairs arranged based on the descending numbers, and each binary node pair is formed by pairing two second candidate nodes from the distance node sequence and the quality node sequence; S8. Migrate the computing task to the optimal migration node.

[0007] In some specific embodiments, marking the M first candidate nodes to which the computing task is to be migrated includes: S1-1. Obtain the dynamic operation quality of the computing task on the original edge node; S1-2. Determine whether the original edge node belongs to a degraded node according to the dynamic operation quality; S1-3. If it belongs to a degraded node, mark the M first candidate nodes to which the computing task is to be migrated; otherwise, continue to run the computing task on the original edge node.

[0008] In some specific embodiments, obtaining the dynamic operation quality of the computing task on the original edge node includes: S1-1-1. Slide and collect several dynamic operation parameters of the computing task on the original edge node within a continuous time window; S1-1-2. Calculate the dynamic operation quality of the computing task within the time window according to the several dynamic operation parameters of the original edge node.

[0009] In some of the specific embodiments, assigning unique node index numbers to M first candidate nodes and the original edge node includes: S2-1. Marking several edge nodes that have a network connection with the original edge node; S2-2. Calculating the spatial distance between several edge nodes and the original edge node; S2-3. Marking the edge nodes with a spatial distance lower than the threshold as first candidate nodes; S2-4. Assigning unique node index numbers to the first candidate nodes and the original edge node according to the spatial distance between the first candidate nodes and the original edge node.

[0010] In some of the specific embodiments, screening Q second candidate nodes from the M first candidate nodes according to the static attribute vectors corresponding to the node index numbers includes: S3-1. Performing K-nearest neighbor clustering on the static attribute vectors of the M first candidate nodes to obtain K static clustering clusters; S3-2. Performing average pooling on several static attribute vectors within the K static clustering clusters to generate K cluster center vectors corresponding to the static clustering clusters; S3-3. Obtaining the static attribute vector of the original edge node and calculating the first similarity between this static attribute vector and the K cluster center vectors; S3-4. Sorting the K static clustering clusters corresponding to the cluster center vectors according to the K first similarities to generate a clustering cluster sequence; S3-5. Starting from the static clustering cluster corresponding to the maximum first similarity, intercepting the first Q similar clustering clusters in the clustering cluster sequence; S3-6. Extracting the cluster center vectors of the first Q similar clustering clusters from the K cluster center vectors, and in each similar clustering cluster, anchoring the static attribute vector closest to its cluster center vector; S3-7. Obtaining the closest static attribute vector and its node index number, and screening Q second candidate nodes based on the node index number; In some of the specific embodiments, calculating the Euclidean distance between the second candidate node and the original edge node includes: S5-1. Marking the timestamp when the original edge node is determined to be a degraded node on the time axis; S5-2. Marking the dynamic operating parameters of the original edge node at this timestamp as abnormal operating parameters; S5-3. Performing feature standardization on several abnormal operating parameters of the original edge node and splicing them into an abnormal operating vector of the original edge node; S5-4. Perform feature standardization on several dynamic operation parameters of the Q second candidate nodes, and splice them into the dynamic operation vectors of the Q second candidate nodes; S5-5. Calculate the Q Euclidean distances between the dynamic operation vectors and the abnormal operation vectors.

[0011] In some specific embodiments, determining the optimal migration node in the node pair sequence includes: S7-1. In the node pair sequence, determine whether there is a binary node pair sharing a node index number; S7-2. If there is a binary node pair sharing a node index number, mark the second candidate node of this binary node pair as the optimal migration node and terminate the search; otherwise, traverse the node pair sequence backward in descending order of the binary node pair numbers; S7-3. If there is no binary node pair sharing a node index number, define the two second candidate nodes in the binary node pair corresponding to the largest descending order number as the third candidate nodes; S7-4. Determine the optimal migration node based on the third candidate nodes.

[0012] In some specific embodiments, determining whether there is a binary node pair sharing a node index number in the node pair sequence includes: S7-1-1. In the node pair sequence, associate the node index numbers of the two second candidate nodes of each binary node pair; S7-1-2. Search in sequence in descending order of the binary node pair numbers, and compare the node index numbers of the two second candidate nodes; S7-1-3. If the node index numbers of the two second candidate nodes are the same, determine the corresponding binary node pair as the binary node pair sharing a node index number.

[0013] In some specific embodiments, finding the optimal migration node based on the third candidate nodes includes: S7-4-1. Obtain the node index numbers of the third candidate nodes, and extract the static attribute vectors of the two third candidate nodes based on the node index numbers; S7-4-2. Calculate the second similarity between the static attribute vectors of the two third candidate nodes and the original edge node vector; S7-4-3. Determine the second candidate node with the larger second similarity as the optimal migration node.

[0014] The present invention provides an edge computing task dynamic migration method supporting service continuity, which has the following beneficial effects: In summary, the present invention constructs a quality node sequence and a distance node sequence by respectively sorting the dynamic operation quality of the second candidate nodes and their Euclidean distances from the abnormal operation states of the original edge nodes in descending order, and generates a node pair sequence based on this. On this basis, if there is a second candidate node with a shared node index number in the node pair, it can be directly determined as the optimal migration node; if not, the static attribute similarity of the two second candidate nodes corresponding to the largest descending order number in the node pair sequence is used to supplement the determination path, so that even in the case of inconsistent sorting results, the migration target with the static attributes closest to the original node can still be selected. The present invention realizes the dual-condition fusion of migration node determination and ensures the service continuity during the dynamic migration process of computing tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 is a schematic flowchart of a method for dynamically migrating edge computing tasks that supports service continuity according to the present invention; Figure 2 is a schematic flowchart of the screening process of the second candidate nodes described in the present invention; Figure 3 is a schematic flowchart of the determination process of the optimal migration node described in the present invention; Figure 4 is a schematic flowchart of the definition process of the third candidate nodes described in the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0016] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts fall within the protection scope of the present invention.

[0017] Please refer to Figures 1 to 4 , the present invention provides a method for dynamically migrating edge computing tasks that supports service continuity, including the following steps: S1. Anchor the original edge node where the computing task runs, and mark M first candidate nodes to which the computing task is to be migrated.

[0018] Exemplarily, step S1 specifically includes: S1-1. Obtain the dynamic operation quality of the computing task on the original edge node; S1-2. Determine whether the original edge node belongs to a degraded node according to the dynamic operation quality; S1-3. If it belongs to a degraded node, mark M first candidate nodes to which the computing task is to be migrated; otherwise, continue to run the computing task on the original edge node.

[0019] Specifically, the dynamic operation quality can be compared with the dynamic operation quality threshold; if the dynamic operation quality is lower than the dynamic operation quality threshold, the original edge node is determined to be a degraded node, and the computing task is marked as a task to be migrated.

[0020] Furthermore, step S1-1 specifically includes: S1-1-1. Slide and collect several dynamic operation parameters of the computing task on the original edge node within a continuous time window; Among them, the dynamic operation parameters include: average response time, average network delay, network packet loss rate, CPU usage rate, and memory usage rate; Exemplarily, the edge node refers to a computing resource unit deployed near the data source or the end user side, with basic computing, storage, and communication capabilities, and is an edge computing device capable of independently carrying out partial task processing. Such as: edge servers, intelligent gateways, industrial control terminals, etc.

[0021] Furthermore, the dynamic operation parameters can reflect the dynamic operation and network status of the edge computing device from aspects such as task execution efficiency, network communication quality, and resource occupancy level.

[0022] Specifically, the time window slides with a unit time step to ensure that each slide triggers real-time update of the dynamic parameters, that is, every time the time window slides by a unit time step, parameters such as average response time and average network delay need to be recollected and calculated.

[0023] S1-1-2. Calculate the dynamic operation quality of the computing task within the time window according to several dynamic operation parameters on the original edge node; The calculation formula for the dynamic operation quality is:

[0024] Among them, is the dynamic operation quality of the original edge node, and its value range is , representing the dynamic operation of the task; is the network sub-item, is the resource sub-item; is the average response time, used to represent the time-consuming of task processing; is the average network delay, used to represent the communication time-consuming between the original edge nodes; is the packet loss rate, used to reflect the instability degree of the network; is a very small positive number (such as ), used to prevent the denominator from being 0; is the CPU usage rate, representing the CPU resource occupancy of the original edge node; Indicates the memory usage rate, representing the memory occupancy of the original edge node; Is the network sub-item adjustment parameter, used to control the sensitivity of the network sub-item, Is the resource sub-item adjustment parameter, used to control the sensitivity of the resource sub-item; Specifically, Is used to comprehensively measure the task delay and network quality; among them, the product Indicates that when the original edge node has high response time and high network latency simultaneously during task processing, the penalty should be increased; further, the denominator Represents the packet loss rate When it rises, the denominator decreases, and the network sub-item rises, thus simulating the phenomenon that high packet loss rate leads to performance degradation. Is used to reflect the collaborative utility of the resource bottleneck of the original edge node; among them, the product Means that only when both resources are highly occupied simultaneously, the resource sub-item will be significantly amplified, thus generating a strong quality penalty. For example, if both are 0.8, the product is 0.64; while if one is 0.8 and the other is only 0.1, the product is only 0.08. The former is 8 times the latter, reflecting the sensitivity of the system to collaborative bottlenecks. Among them, adding 1 to the product is used to ensure that the resource sub-item ≥ 1, avoiding negative exponents.

[0025] This calculation formula comprehensively evaluates the dynamic operation impact of the average response time, average network delay, network packet loss rate, CPU usage rate, and memory usage rate on the original edge node through an exponential decay structure. The function structure of this calculation formula has a non-linear amplification characteristic, which can significantly convert the increase of the above parameters into the decrease of the dynamic operation quality Q.

[0026] The method further includes: S2. Assign unique node index numbers to M first candidate nodes and the original edge node, and obtain the static attribute vectors corresponding to the node index numbers.

[0027] Specifically, the static attribute vector is composed of several static attribute characteristics of the first candidate nodes and the original edge node, including: geographical location information, computing power, network access bandwidth, etc., and remains relatively unchanged during task execution.

[0028] Exemplarily, step S2 includes: S2-1. Mark several edge nodes that have network connections with the original edge node; S2-2. Calculate the spatial distance between several edge nodes and the original edge node; S2-3. Mark the edge nodes with spatial distance lower than the threshold as the first candidate nodes; S2-4. Assign unique node index numbers to the first candidate nodes and the original edge nodes according to the spatial distance between the first candidate nodes and the original edge nodes.

[0029] Among them, the node index numbers are encoded according to the spatial distance. For example, it can be set that the closer the distance, the more forward (or smaller) the index value, so as to facilitate sorting, mapping, and alignment in subsequent operations.

[0030] It should be noted that: the spatial distance between the original edge node and itself is defined as a null value; when assigning node index numbers to all participating nodes, ascending coding of the spatial distance can be adopted to ensure that the node index can reflect the physical topology structure at the same time.

[0031] The method further includes: S3. Select Q second candidate nodes from the M first candidate nodes according to the static attribute vectors corresponding to the node index numbers.

[0032] Exemplarily, step S3 specifically includes: S3-1. Perform K-nearest neighbor clustering on the static attribute vectors of the M first candidate nodes to obtain K static clustering clusters; It should be noted that each static clustering cluster in this embodiment contains multiple static attribute vectors, and the static attribute vector is a multi-dimensional vector composed of static features generated after outlier processing and normalization of multiple static attribute parameters of the corresponding edge node.

[0033] Therefore, the "static features" described in this embodiment are all standard vectors with unified structure and consistent dimension, which can accurately reflect the attribute characteristics of each node in static dimensions such as computing power, network access, and geographical location.

[0034] S3-2. Perform average pooling on several static attribute vectors in the K static clustering clusters to generate K cluster center vectors corresponding to the static clustering clusters; It should be noted that the static attribute vectors in each static clustering cluster have N-dimensional features; since in K-nearest neighbor clustering, the static attribute vectors usually show high consistency in each dimension, they can be regarded as close to each other in the multi-dimensional feature space. However, in order to generate a representative vector for each static clustering cluster, it is still necessary to uniformly represent the static attribute vectors within the cluster. Because the vectors within the clustering cluster are relatively close in each dimension, average pooling operations can be performed on them to generate representative cluster center vectors for generally characterizing the overall static attributes of the clustering cluster.

[0035] S3-3. Obtain the static attribute vector of the original edge node and calculate the first similarity between this static attribute vector and the K cluster center vectors; Specifically, the first similarity can be selected from a variety of vector similarity metrics, such as cosine similarity, the negative value of Euclidean distance, the reciprocal of Manhattan distance, etc.

[0036] S3-4. Sort the K static clustering clusters corresponding to the cluster center vectors according to the K first similarities to generate a clustering cluster sequence. S3-5. Starting from the static clustering cluster corresponding to the maximum first similarity, intercept the first Q similar clustering clusters in the clustering cluster sequence. Specifically, in this embodiment, the value range of Q is negatively correlated with the number of clustering clusters K. Given that the larger the number of clustering clusters K, the more discrete the distribution of node static features, and the more likely there are clustering clusters that highly match the static features of the original edge nodes. Therefore, the value range of Q can be appropriately reduced.

[0037] Exemplarily, Q can be set by the following empirical rules: If K ≤ 10, then Q = 0.4K; If 10 < K < 30, then Q = 0.2K; If K > 30, then Q = 0.1K; S3-6. Extract the cluster center vectors of the first Q similar clustering clusters from the K cluster center vectors, and in each similar clustering cluster, anchor the static attribute vector closest to its cluster center vector. Specifically, the closest static vector refers to the static attribute vector with the smallest Euclidean distance from the cluster center vector in the similar clustering cluster.

[0038] It should be noted that the cluster center vector is the average pooling result of the clustering center, so it may not strictly correspond to a certain real edge node. Therefore, the closest static attribute vector is needed to represent the real edge node.

[0039] S3-7. Obtain the closest static attribute vectors and their node indexes, and screen Q second candidate nodes based on the node indexes. Specifically, each closest static attribute vector is uniquely associated with a first candidate node through its corresponding node index. Thus, the first candidate node corresponding to the closest static attribute vector can be clearly selected as the second candidate node, that is, a mapping path of "cluster center → closest static attribute vector → node index → second candidate node" is formed.

[0040] The method further includes: S4. Concatenate the static attribute vectors of the second candidate nodes with the normalized features of the computing task as the joint input of the parameter regression model to predict several dynamic operation parameters of the computing task on the Q second candidate nodes.

[0041] It should be noted that the standardized features of the computing tasks include: task type, input data scale, number of threads, response time constraint, interaction frequency, etc. After normalization or standardization processing, they have a unified dimension and scale; then, the standardized features of the computing tasks are concatenated with the static attribute vectors of each second candidate node to form the combined input of the parameter model.

[0042] Furthermore, the parameter regression model can select common regression prediction models such as support vector regression (SVR) or multi-layer perceptron (MLP). This regression prediction model is offline trained through historical training samples.

[0043] Exemplarily, the historical training samples are constructed in the following way: First, the historical computing tasks are standardized according to preset semantics or target features; Then, the standardized features of each historical task are concatenated with the static attribute vectors corresponding to the edge nodes (i.e., the second candidate nodes) where they have actually run, generating a joint feature representation of task-node, which is used to characterize the matching relationship between the computing task and the edge node in terms of static attributes.

[0044] Next, taking this joint feature representation as the input, and several dynamic operation parameters (such as response time, network delay, packet loss rate, CPU and memory usage rate) collected during the actual operation process of the historical computing task on the corresponding node as multi-objective labels, a supervised learning sample of the regression model is formed; Then, the model performance is optimized through loss functions such as mean square error (MSE), and strategies such as cross-validation are used to control the training error and generalization ability. Finally, based on several historical training samples, the mapping relationship between the computing task and node attributes to the dynamic operation parameters is fitted iteratively.

[0045] The method further includes: S5. Calculate the dynamic operation quality of Q second candidate nodes according to the predicted several dynamic operation parameters, and calculate the Euclidean distance between the second candidate nodes and the original edge node.

[0046] Exemplarily, step S5 specifically includes: S5-1. Mark the timestamp when the original edge node is determined to be a degraded node on the time axis; S5-2. Mark the dynamic operation parameters of the original edge node at this timestamp as abnormal operation parameters; It should be noted that "a number of abnormal operation parameters" refers to the set of dynamic operation parameters collected by the edge node at the moment when the computing task is determined to be a task to be migrated. This set has a clear timestamp (such as the central timestamp of the time window or the front-end and back-end timestamps). Since the computing task corresponding to this timestamp has been determined to be a task to be migrated, therefore, the dynamic operation parameters of this timestamp are uniformly marked as abnormal operation parameters. As for the specific reason for the abnormality, it may be caused by a single parameter abnormality or the combined effect of multiple parameters.

[0047] S5-3. Standardize the features of a number of abnormal operation parameters of the original edge node and splice them into an abnormal operation vector of the original edge node; S5-4. Standardize the features of a number of dynamic operation parameters of Q second candidate nodes and splice them into dynamic operation vectors of Q second candidate nodes; S5-5. Calculate Q Euclidean distances between the dynamic operation vectors and the abnormal operation vectors.

[0048] In this embodiment, through the unified standardization processing of the abnormal operation parameters and the operation parameters of the second candidate nodes, and calculating their Euclidean distances, a quantitative representation of the operation state difference is realized.

[0049] The method further includes: S6. Sort the Q second candidate nodes in descending order according to the Euclidean distance and the dynamic operation quality respectively to generate a distance node sequence and a quality node sequence.

[0050] Among them, both the distance node sequence and the quality node sequence are sorted in descending order, and a descending number is assigned to each second candidate node in sequence from front to back.

[0051] In the distance node sequence, after sorting the Euclidean distance values from large to small, a descending number is assigned to each second candidate node in sequence, and the numbers range from Q to 1. In the quality node sequence, after sorting the dynamic operation quality from large to small, a descending number is also assigned to the nodes in sequence, and the numbers range from Q to 1.

[0052] The method further includes: S7. Align the distance node sequence and the quality node sequence based on the descending numbers to generate a node pair sequence, and determine the optimal migration node in the node pair sequence.

[0053] Among them, the node pair sequence includes Q binary node pairs arranged based on the descending numbers, and each binary node pair is composed of two second candidate nodes from the distance node sequence and the quality node sequence paired.

[0054] Exemplarily, step S7 specifically includes: S7-1. Determine whether there is a binary node pair sharing a node index number in the node pair sequence; S7-3. If there is a binary node pair sharing a node index number, mark the second candidate node of this binary node pair as the optimal migration node and terminate the search; otherwise, traverse the node pair sequence backward in descending order of the binary node pairs; S7-4. If there is no binary node pair sharing a node index number, define the two second candidate nodes in the binary node pair corresponding to the largest descending order number as the third candidate nodes; Specifically, the largest descending order number refers to the number ranked most forward in the node pair sequence, that is, the combined node pair where the two nodes rank first (descending order number is Q) in the distance node sequence and the quality node sequence respectively. This node pair performs optimally in each of the two evaluation dimensions. Although it does not hit the same node, it can be used as a representative node with high operating quality and excellent operating status.

[0055] S7-4. Determine the optimal migration node based on the third candidate nodes.

[0056] In this embodiment, by setting the matching path for sharing the node index number and the fallback comparison path based on the third candidate nodes, a dual-path determination of the optimal migration node is achieved.

[0057] Further, step S7-2 specifically includes: S7-2-1. In the node pair sequence, associate the node index numbers of the two second candidate nodes of each binary node pair; S7-2-2. Sequentially search in descending order of the binary node pairs and compare the node index numbers of the two second candidate nodes; S7-2-3. If the node index numbers of the two second candidate nodes are the same, determine the corresponding binary node pair as the binary node pair sharing the node index number.

[0058] Specifically, if the node index numbers are the same, that is, the corresponding second candidate nodes in the distance node sequence and the quality node sequence point to the same edge node, it indicates that this second candidate node ranks optimally in both dimensions of the operating state (long Euclidean distance) and the operating quality (good performance), and can be regarded as the optimal migration node with double hits.

[0059] Exemplarily, step S7-4 specifically includes: S7-4-1. Obtain the node index numbers of the third candidate nodes and extract the static attribute vectors of the two third candidate nodes based on the node index numbers; S7-4-2. Calculate the second similarity between the static attribute vectors of the two third candidate nodes and the original edge node vector; Exemplarily, the second similarity preferably adopts the same similarity calculation strategy as the first similarity.

[0060] S7-4-3. Determine the second candidate node with a larger second similarity as the optimal migration node.

[0061] In this embodiment, by calculating the second similarity between the static attributes of the third candidate node and the original edge node, in the case where no consensus is reached in the sorting, a supplementary decision based on similarity is provided. It can effectively avoid finding the node to be migrated that is closest to the original edge node even in the case of lacking a shared node index number.

[0062] The method further includes: S8. Migrate the computing task to the optimal migration node.

[0063] Specifically, task migration can be implemented by using existing edge computing task migration technologies. For example, by saving and restoring the task context, encapsulating and transmitting the running state, reconstructing the environment of the target node, etc., the computing task can be smoothly migrated from the original node to the optimal migration node to ensure the service continuity of task execution.

[0064] In this embodiment, by aligning the distance node sequence and the quality node sequence based on the descending order number, constructing a node pair sequence, and sequentially searching for a binary node pair with a shared node index number, a dual-dimensional fusion judgment of the running state and the running quality is achieved; when a shared node exists, it can be directly determined as the optimal migration node; if not hit, further screening is performed through the static similarity of the third candidate node to ensure the accurate determination of the optimal migration node during task migration.

[0065] In summary, the present invention proposes a method for determining the optimal migration node based on double sequence alignment. Specifically, first, based on the first similarity between the static attribute vector of the original edge node and the static attribute vectors of several first candidate nodes, Q second candidate nodes with the static attributes closest to the original edge node are screened out through static feature clustering and cluster center proximity matching. Then, a parameter regression model is used to predict the dynamic operation parameters of these Q second candidate nodes under the current computing task, so as to calculate their dynamic operation quality and the Euclidean distance from the abnormal operation state respectively. Subsequently, by descendingly numbering the dynamic operation quality of the second candidate node and the Euclidean distance from the abnormal node respectively, two quality node sequences and distance node sequences are constructed; then, based on the unified descending order number, a node pair sequence is generated by alignment. On this basis, the optimal migration node is determined through the shared node index number and the third candidate node.

[0066] Further, if there is a second candidate node with a shared index number in the node pair, it indicates that the second candidate node performs optimally in both the operation state and operation quality dimensions and can be directly used as the optimal migration node. If the shared node is not hit, the similarity comparison of the static attributes of the end node corresponding to the largest descending number in the node pair sequence is used to supplement the node determination path, ensuring that even if the sorting is inconsistent, the optimal migration target with the static features closest to the original node can be selected.

[0067] The present invention gets rid of the dependence on a single sorting, avoids the conflict that the "optimal quality" and "farthest state" nodes cannot be uniformly determined, and effectively improves the accuracy and service continuity of edge computing tasks during task migration.

[0068] The above embodiments can be implemented in whole or in part by software, hardware, firmware, or any other combination. When implemented using software, the above embodiments can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, the processes or functions described in the embodiments of the present application are generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center by wire (such as infrared, wireless, microwave, etc.).

[0069] The computer-readable storage medium can be any available medium that the computer can access or a data storage device such as a server or data center that contains one or more collections of available media. The available media can be magnetic media (such as floppy disks, hard disks, magnetic tapes), optical media (such as DVD ), or semiconductor media. The semiconductor media can be a solid-state drive.

[0070] In several embodiments provided in the present application, it should be understood that the disclosed systems, devices, and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the displayed or discussed mutual coupling, direct coupling, or communication connection can be through some interfaces, and the indirect coupling or communication connection of the device or unit can be in an electrical, mechanical, or other form.

[0071] As described above, it is only the specific implementation manner of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present application can easily think of changes or substitutions, which should all be covered within the protection scope of the present application.

Claims

1. A dynamic migration method for edge computing tasks that supports service continuity, characterized in that, Including: S1. Anchor the original edge node where the computing task runs, and mark M first candidate nodes to which the computing task is to be migrated; S2. Assign unique node index numbers to the M first candidate nodes and the original edge node, and obtain the static attribute vectors corresponding to the node index numbers; S3. Screen Q second candidate nodes from the M first candidate nodes according to the static attribute vectors corresponding to the node index numbers; S4. Concatenate the static attribute vectors of the second candidate nodes with the normalized features of the computing task as the combined input of the parameter regression model to predict several dynamic operation parameters of the computing task on the Q second candidate nodes; S5. Calculate the dynamic operation quality of the Q second candidate nodes according to the predicted several dynamic operation parameters, and calculate the Euclidean distance between the second candidate nodes and the original edge node; S6. Respectively sort the Q second candidate nodes in descending order according to the Euclidean distance and the dynamic operation quality to generate a distance node sequence and a quality node sequence; Wherein, both the distance node sequence and the quality node sequence are sorted in descending order, and each second candidate node is assigned a descending number that decreases sequentially from front to back; S7. Align the distance node sequence and the quality node sequence based on the descending numbers to generate a node pair sequence, and determine the optimal migration node in the node pair sequence; Wherein, the node pair sequence includes Q binary node pairs arranged based on the descending numbers, and each binary node pair is paired by two second candidate nodes from the distance node sequence and the quality node sequence; S8. Migrate the computing task to the optimal migration node.

2. The dynamic migration method of edge computing tasks supporting service continuity according to claim 1, wherein Marking the M first candidate nodes to which the computing task is to be migrated includes: S1-1. Obtain the dynamic operation quality of the computing task on the original edge node; S1-2. Determine whether the original edge node belongs to a degenerate node according to the dynamic operation quality; S1-3. If it belongs to a degenerate node, mark the M first candidate nodes to which the computing task is to be migrated; otherwise, continue to run the computing task on the original edge node.

3. The dynamic migration method of edge computing tasks supporting service continuity according to claim 2, wherein Obtaining the dynamic operation quality of the computing task on the original edge node includes: S1-1-1. Slide and collect several dynamic operation parameters of the computing task on the original edge node within a continuous time window; S1-1-2. Calculate the dynamic operation quality of the computing task within the time window according to the several dynamic operation parameters of the original edge node.

4. A method for dynamically migrating edge computing tasks to support service continuity according to claim 1, characterized in that Assigning unique node index numbers to the M first candidate nodes and the original edge node includes: S2-1. Mark several edge nodes having a network connection with the original edge node; S2-2. Calculate the spatial distance between several edge nodes and the original edge node; S2-3. Mark the edge nodes with a spatial distance lower than the threshold as the first candidate nodes; S2-4. Assign unique node index numbers to the first candidate nodes and the original edge node according to the spatial distance between the first candidate nodes and the original edge node.

5. A method for dynamically migrating edge computing tasks that support service continuity according to claim 1, characterized in that, Screening Q second candidate nodes from the M first candidate nodes according to the static attribute vectors corresponding to the node index numbers includes: S3-1. Perform K-nearest neighbor clustering on the static attribute vectors of the M first candidate nodes to obtain K static clustering clusters; S3-2. Perform average pooling on a number of static attribute vectors within K static clustering clusters to generate K cluster center vectors corresponding to the static clustering clusters; S3-3. Obtain the static attribute vector of the original edge node and calculate the first similarity between this static attribute vector and the K cluster center vectors; S3-4. According to the K first similarities, sort the K static clustering clusters corresponding to the cluster center vectors to generate a clustering cluster sequence; S3-5. Starting from the static clustering cluster corresponding to the maximum first similarity, intercept the first Q similar clustering clusters in the clustering cluster sequence; S3-6. Extract the cluster center vectors of the first Q similar clustering clusters from the K cluster center vectors, and in each similar clustering cluster, anchor the static attribute vector closest to its cluster center vector; S3-7. Obtain the closest static attribute vector and its node index number, and screen Q second candidate nodes based on the node index number.

6. A method for dynamically migrating edge computing tasks to support service continuity according to claim 1, characterized in that Calculate the Euclidean distance between the second candidate nodes and the original edge node, including: S5-1. Mark the timestamp when the original edge node is determined to be a degraded node on the time axis; S5-2. Mark the dynamic operating parameters of the original edge node at this timestamp as abnormal operating parameters; S5-3. Perform feature standardization on a number of abnormal operating parameters of the original edge node and splice them into an abnormal operating vector of the original edge node; S5-4. Perform feature standardization on a number of dynamic operating parameters of the Q second candidate nodes and splice them into dynamic operating vectors of the Q second candidate nodes; S5-5. Calculate the Q Euclidean distances between the dynamic operating vectors and the abnormal operating vectors.

7. A method for dynamically migrating edge computing tasks that support service continuity according to claim 1, characterized in that, Determine the optimal migration node in the node pair sequence, including: S7-1. In the node pair sequence, determine whether there is a binary node pair sharing the node index number; S7-2. If there is a binary node pair sharing the node index number, mark the second candidate node of this binary node pair as the optimal migration node and terminate the search; otherwise, traverse the node pair sequence backward along the descending order number of the binary node pairs; S7-3. If there is no binary node pair sharing the node index number, define the two second candidate nodes in the binary node pair corresponding to the maximum descending order number as the third candidate nodes; S7-4. Determine the optimal migration node based on the third candidate nodes.

8. A method for dynamically migrating edge computing tasks that support service continuity according to claim 1, characterized in that, In the node pair sequence, determine whether there is a binary node pair sharing the node index number, including: S7-1-1. In the node pair sequence, associate the node index numbers of the two second candidate nodes of each binary node pair; S7-1-2. Sequentially search along the descending order number of the binary node pairs and compare the node index numbers of the two second candidate nodes; S7-1-3. If the node index numbers of the two second candidate nodes are the same, determine the corresponding binary node pair as the binary node pair sharing the node index number.

9. A method for dynamically migrating edge computing tasks that support service continuity according to claim 1, characterized in that Search for the optimal migration node based on the third candidate nodes, including: S7-4-1. Obtain the node index numbers of the third candidate nodes and extract the static attribute vectors of the two third candidate nodes based on the node index numbers; S7-4-2. Calculate the second similarity between the static attribute vectors of the two third candidate nodes and the original edge node vector; S7-4-3. Determine the second candidate node with the second largest similarity as the optimal migration node.

Images