A satellite edge computing offloading method and device based on dynamic time-varying

Abstract

The present invention discloses a satellite edge computing offloading method and device based on dynamic time-varying properties. The method comprises: before each task offloading starts, the task source extracts the characteristic information of the task through a task structure analyzer; the task source sends the characteristic information to the nearest cluster head satellite, so that the cluster head satellite determines the target satellite for task offloading according to the characteristic information of the task; the task source offloads the task to the target satellite for execution, and receives the task execution result of the target satellite. Aiming at the dynamic time-varying characteristics of the satellite network, the present invention realizes the reliable execution of tasks in dynamic time-varying satellites by reasonably selecting the execution satellite nodes and judging the possible signal loss situations.

Description

Technical Field

[0001] The embodiments of the present invention relate to the field of satellite collaborative computing technology, and in particular to a satellite edge computing offloading method and device based on dynamic time-varying properties. Background Art

[0002] In recent years, with the development of the Internet of Things and artificial intelligence technology, a large number of computing scenarios (such as smart cameras, autonomous driving, etc.) have emerged, and the amount of computing requests generated by clients has increased exponentially. Although cloud computing centers have super-large-scale computing capabilities, cloud computing requires centralized processing in the cloud. Massive requests and data uploads to the cloud center will face the risk of network delay and jitter. In addition, some terminal devices may be in a complex edge environment and far away from the cloud center, and the data upload efficiency is seriously affected by the network signal environment. Therefore, low-orbit satellites have gradually become a hot spot in edge computing research due to their wide network coverage and strong computing power. Some studies use satellites as network relays to expand the communication distance between user devices and edge computing servers. Some studies also further reduce access delays by offloading computing tasks to satellites. However, in these studies, each satellite only participates in the task as an independent individual, ignoring the collaborative computing method, so there are certain limitations.

[0003] First, the satellite's computing resources and memory capacity are insufficient, resulting in a large number of tasks that cannot be executed on the satellite side. The tasks need to be sent to the ground cloud center platform for processing, causing the problem of satellite downlink bandwidth overload. Secondly, the population distribution on the earth's surface is uneven, and the edge device task requests are concentrated in the local area of ​​the surface. Therefore, a large number of satellites are idle most of the time, and the satellite resource utilization rate is low. Finally, there are a large number of satellites in orbit, and the system design is very complicated with a single satellite as the minimum unit for resource management and task allocation. A large number of studies have shown that satellite collaborative edge computing has the following advantages: tasks can be divided into multiple subtasks and executed in parallel in multiple satellites, which speeds up task processing. Satellite clusters can form a resource pool, dynamically share hardware resources between satellites, and receive more tasks. Using satellite clusters as the minimum unit for task allocation can greatly reduce the scale of the problem solution space and improve the efficiency of solving.

[0004] However, collaborative computing of on-orbit satellites often faces challenges such as limited hardware resources, real-time constraints of tasks, and dynamic time-varying nature of satellites.

[0005] 1. Unloading a large number of tasks at a concentrated time is prone to hardware resource limitations

[0006] When a satellite cluster receives a large number of task requests from devices within the communication range within a certain period of time, improper offloading strategy settings can easily generate a large amount of hardware resource fragmentation in the cluster, causing more tasks to wait for resource release, which delays the task completion time and results in low task execution efficiency.

[0007] 2. Task real-time requirements and constraints

[0008] Tasks often have real-time requirements. The mission loads of satellites in orbit vary, which may cause the waiting time of the same task to vary greatly among different satellites in the satellite cluster. Therefore, it is particularly important to establish a mission load analysis at the satellite node.

[0009] 3. Satellite movement causes loss of connection with the device

[0010] The geographical location of edge devices such as satellites has a strong time-varying characteristic, which leads to problems such as unstable relative positions between different nodes, highly dynamic and time-varying network topology, and unstable connections. Summary of the invention

[0011] In order to solve the problems of the prior art, the present invention provides a satellite edge computing offloading method and device based on dynamic time variability. The method allows nodes to dynamically fall behind and reconnect, which can simultaneously meet the real-time requirements of the task and achieve resource optimization.

[0012] In a first aspect, an embodiment of the present invention provides a satellite edge computing offloading method based on dynamic time variability, including:

[0013] S110, before each task offloading starts, the task source extracts the characteristic information of the task through the task structure analyzer;

[0014] S120, the task source sends the characteristic information to the nearest cluster head satellite, so that the cluster head satellite determines the target satellite for task offloading according to the characteristic information of the task;

[0015] S130: The task source unloads the task to the target satellite for execution, and receives the task execution result of the target satellite.

[0016] Optionally, the cluster head satellite determines the target satellite for task offloading according to the characteristic information of the task, including:

[0017] Determine the matching degree between each satellite node and the task according to the task application resource amount and the remaining resource amount of each satellite node;

[0018] Introducing neural network similarity modeling analysis technology to predict the execution time of tasks on different satellite nodes;

[0019] Determine the communication distance of the task at different satellite nodes according to the distance between the cluster head satellite and each satellite node and the task source;

[0020] A target satellite for task offloading is determined according to the matching degree, the predicted execution time, and the communication distance.

[0021] Optionally, the matching degree between each satellite node and the task is determined according to the task application resource amount and the remaining resource amount of each satellite node, including:

[0022] Determine the matching degree between each satellite node and the task according to the proportion of the sum of the task application resource amount and the resources used by each satellite node in the current satellite, and the variance between the task application resource amount and the remaining resources of each satellite node;

[0023] The greater the difference between the proportion and the variance, the higher the matching degree.

[0024] Optionally, neural network similarity modeling analysis technology is introduced to predict the execution time of tasks on different satellite nodes, including:

[0025] When the cluster head satellite receives the characteristic information, it combines the performance characterization network of each satellite node in the cluster to predict the running time of the task on each satellite;

[0026] When there are no idle resources available for execution in the satellites within the cluster, the current task is first added to the task waiting queue, and the greedy algorithm is used to calculate the estimated waiting time for the task to wait for resource release in each satellite node.

[0027] Optionally, determining a target satellite for task offloading according to the matching degree, the predicted execution time, and the communication distance includes:

[0028] Setting weights corresponding to the matching degree, the predicted execution time, and the communication distance, respectively, according to task characteristics;

[0029] The target satellite for task offloading is determined according to the matching degree, the predicted execution time, and the sum of the products of the communication distance and the corresponding weight.

[0030] Optionally, the method further includes:

[0031] Predict the communication distance between the target satellite and the mission source during the mission execution;

[0032] When it is predicted that the target satellite will be out of the communication range of the task source before the task is completed, the task source sends a task unloading signal to the target satellite and determines the target migration satellite from other satellite nodes in the cluster again;

[0033] The unloading on the target satellite is transferred to the target migration satellite and continued.

[0034] In a second aspect, an embodiment of the present invention further provides a satellite edge computing unloading device based on dynamic time variability, including:

[0035] The feature information extraction module is used for the task source to extract the feature information of the task through the task structure analyzer before each task offloading starts;

[0036] The task scheduling module is used to send the characteristic information to the nearest cluster head satellite through the task source, so that the cluster head satellite can determine the target satellite for task offloading according to the characteristic information of the task;

[0037] The task execution module is used to offload the task to the target satellite for execution through the task source, and receive the task execution result of the target satellite.

[0038] Optionally, the task scheduling module is specifically used to:

[0039] Determine the matching degree between each satellite node and the task according to the task application resource amount and the remaining resource amount of each satellite node;

[0040] Introducing neural network similarity modeling analysis technology to predict the execution time of tasks on different satellite nodes;

[0041] Determine the communication distance of the task at different satellite nodes according to the distance between the cluster head satellite and each satellite node and the task source;

[0042] A target satellite for task offloading is determined according to the matching degree, the predicted execution time, and the communication distance.

[0043] Optionally, the device further includes a task migration module, configured to execute:

[0044] Predict the communication distance between the target satellite and the mission source during the mission execution;

[0045] When it is predicted that the target satellite will be out of the communication range of the task source before the task is completed, the task source sends a task unloading signal to the target satellite and determines the target migration satellite from other satellite nodes in the cluster again;

[0046] The unloading on the target satellite is transferred to the target migration satellite and continued.

[0047] The beneficial effects of the present invention include:

[0048] 1. The present invention predicts the running time of satellite-borne missions on different satellites after analyzing the mission characteristics, and uses a greedy algorithm to predict the execution time of the satellite-borne missions on different satellites, providing an evaluation index for satellite selection, thereby avoiding the satellite-borne missions from waiting for resources to be released and causing a long waiting time.

[0049] 2. The present invention designs a satellite scoring strategy based on the modeling of the remaining resources of each satellite in the satellite cluster. The purpose is to reduce the fragmentation of computing resources and prevent uneven resource allocation from causing the satellite's resource gap to be unable to adapt to subsequent conventional tasks, thereby reducing the task waiting time for resource release and improving task offloading efficiency.

[0050] 3. The present invention maintains the execution satellite within the communication range of the task source device. The target execution satellite first transmits back information such as the running speed and the current longitude and latitude. The task source device determines the specific departure time of the target execution satellite, and then sends an unloading command to the satellite that is about to leave the communication range to migrate its task to other satellites to ensure that the task can continue to be executed. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 An overall architecture diagram of a satellite edge computing offloading method based on dynamic time variability provided by an embodiment of the present invention;

[0052] Figure 2 A flow chart of a satellite edge computing offloading method based on dynamic time variability provided by an embodiment of the present invention;

[0053] Figure 3 A schematic diagram of task migration provided by an embodiment of the present invention;

[0054] Figure 4 A schematic diagram of the communication distance between a task source and a satellite node provided in an embodiment of the present invention;

[0055] Figure 5 A schematic diagram of task waiting time under different offloading strategies provided by an embodiment of the present invention;

[0056] Figure 6 A schematic diagram of the total execution time of tasks under different offloading strategies provided by an embodiment of the present invention;

[0057] Figure 7 A schematic diagram of resource utilization under different unloading strategies provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0058] The present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments. It is to be understood that the specific embodiments described herein are only used to explain the present invention, rather than to limit the present invention. It should also be noted that, for ease of description, only parts related to the present invention, rather than all structures, are shown in the accompanying drawings.

[0059] Unlike the static scenarios of ground edge computing, satellite missions are mainly affected by the dynamic time-varying characteristics of satellite networks. Satellites are always in high-speed motion, and the communication coverage area between cloud-edge devices may change, which may cause some specific tasks to be interrupted or data lost. Therefore, this application will propose a satellite edge computing offloading method (DTV) based on dynamic time-varying joint resource optimization for the dynamic time-varying characteristics of satellite networks.

[0060] like Figure 1 As shown, the present invention is mainly implemented by a task scheduling module and a task migration module. This application fully considers the challenges of task unloading and execution of satellites in dynamic operation. The task scheduling module reasonably selects the execution satellite node, and the task migration module provides a guarantee for possible signal loss. Through the close combination of these two modules, the task can be reliably executed in a dynamic and time-varying satellite.

[0061] Example

[0062] See also Figure 2 , Figure 2 A flowchart of a satellite edge computing offloading method based on dynamic time-varying performance provided by an embodiment of the present invention specifically includes the following steps:

[0063] S110 , before each task offloading starts, the task source extracts feature information of the task through a task structure analyzer.

[0064] The task source is the equipment that needs to offload the task, such as an aircraft. During the task execution, the task source may also be in a real-time motion state.

[0065] The characteristic information of a task may include the type of task, such as tasks may be divided into different types such as computation-intensive tasks and communication-intensive tasks, and may also include resources required to execute the task.

[0066] S120: The task source sends the characteristic information to the nearest cluster head satellite, so that the cluster head satellite can determine the target satellite for task offloading according to the characteristic information of the task.

[0067] In this embodiment, the target satellite node for executing the current task is reasonably selected by utilizing the task scheduling module. According to the characteristics of the task, the task can be divided into computing-intensive, communication-intensive and other tasks. Different tasks require different attributes of resources. If the resource overload on a satellite node is serious, it will cause an unbalanced state of resource distribution. Therefore, subsequent tasks may not be able to fully utilize the idle resources of the satellite node, resulting in more resource fragments. In addition, considering that some tasks will cause major problems if they fail to return results before the deadline. Therefore, the hard real-time and soft real-time characteristics of the task should also be used as important reference indicators for container task scheduling.

[0068] Therefore, this embodiment first scores the matching degree between the task application resource amount and the node remaining resource amount in order to solve the resource fragmentation problem; then introduces the neural network similarity modeling and analysis technology to predict the execution time of the task at different nodes; finally, since the relative position of the satellite and the terminal device is constantly changing, if the best execution satellite cannot be selected, the possibility of data packet loss during transmission will increase, and the transmission delay will increase significantly, so the communication nodes between the task source and the satellite must be calculated.

[0069] Specifically, S120 includes:

[0070] S121. Determine the matching degree between each satellite node and the task according to the task application resource amount and the remaining resource amount of each satellite node.

[0071] In the satellite edge computing scenario, task execution usually requires the participation of multi-dimensional resources. The shortage of a certain resource may affect the final execution effect of the task, so the task needs to wait for the resource to be released before it can be executed normally. The resources required for task execution are mainly divided into computing resources and storage resources. Usually, resource scheduling starts from occupying more satellites, which can improve the utilization rate of satellites. However, this scheduling strategy will increase the possibility of resource fragmentation in satellites. This resource fragmentation will result in the remaining resources in the satellite cluster being sufficient for task delivery, but a single satellite cannot meet the resource application conditions alone.

[0072] Therefore, the resource optimization modeling in this embodiment will start from optimizing resource fragmentation, and by reducing the possibility of single satellite resource fragmentation, more satellite resources can be utilized in a distributed manner, thereby reducing the waiting delay caused by waiting for resource release.

[0073] In terms of data and algorithms, there are the following definitions and explanations:

[0074] Definition 1.1: Assuming that satellite k has hardware resources R, the hardware resources on the satellite can be defined as:

[0075] R k = {r cpu , rmen , r gpu}

[0076] R is the resource set of satellite k, including the resources of CPU, MEM, and GPU.

[0077] Definition 1.2: Assuming that the current mission applies for satellite k resources rq, the adaptation value Fit in satellite k can be defined as:

[0078]

[0079] l ik represents the usage of resource i in satellite node k, where rq mem 、rq cpu 、rq gpu The calculation units are MB, number of cores, and number of items, respectively. var represents the variance function.

[0080] According to the above formula, the satellite node with the highest resource adaptability can be found. The satellite node with the highest adaptability should meet the following two evaluation indicators:

[0081] (1) The higher the proportion of the sum of the resources requested by the mission and the resources already used in the current satellite, the higher the satellite suitability. That is, the mission is preferentially sent to the satellite with the most suitable resources to reduce satellite resource fragmentation.

[0082] (2) The smaller the variance between the resources requested by the task and the remaining resources, the higher the satellite adaptability. This can make the remaining resources of the satellite more balanced after the task is assigned, avoiding the problem of resource gap that causes the subsequent normal tasks to fail to be issued.

[0083] S122. Introduce neural network similarity modeling analysis technology to predict the execution time of tasks on different satellite nodes.

[0084] In the satellite offloading computing scenario, due to the limited number of hardware resources, the task may need to increase the waiting time due to waiting for the release of resources, resulting in low efficiency of the entire task execution. In the general ground computing offloading scenario, the cluster head satellite may always monitor the resource situation of the entire cluster, and immediately unload the task to a satellite after the resources are released. However, due to the bandwidth size of the inter-satellite link, this method will produce a large transmission delay, which will eventually increase the total execution time of the task. Therefore, if it is possible to predict when a satellite will release the hardware resources that meet the task operation, the task can be unloaded to the satellite in advance. At this time, the transmission delay can be masked by the waiting delay, thereby reducing the total execution time of the task.

[0085] After extracting the task features, the task source device sends the feature information to the cluster head satellite, thus avoiding the increase of communication delay by transmitting the entire task. After receiving the feature information, the cluster head satellite combines the performance characterization network of the satellites in the cluster to predict the running time T of the task on each satellite. e When a satellite may not have free resources to execute and needs to join the task waiting queue first, this embodiment uses a greedy algorithm to calculate the estimated waiting time that the task needs to wait for resource release in each satellite.

[0086] Specifically, the optimal unloading subproblem with the least waiting time is expressed as T new , the expected deadline set of each satellite mission is W = {T1, T2, T3…T N}. Since the deadlines in W are sorted from early to late, T new The size of depends on the latest expected deadline in W that satisfies the release of resources for task dispatch. That is, minimize T new It can be expressed as:

[0087] A n =A(T i ...T n )

[0088]

[0089] Among them, A n Represents the combined set of tasks in the waiting queue. new Indicates the shortest deadline for task issuance and task prediction time T e sum.

[0090] This embodiment predicts the running time of the satellite-borne mission on different satellites after analyzing the mission characteristics, and uses a greedy algorithm to predict the execution time of the satellite-borne mission on different satellites, providing an evaluation index for satellite selection, thereby avoiding the satellite-borne mission from waiting for too long due to waiting for resource release.

[0091] S123, determining the communication distance of the task in different satellite nodes according to the distances between the cluster head satellite and each satellite node and the task source.

[0092] The satellite collaborative edge computing processing flow usually includes five steps: uploading task parameters to the cluster head satellite, the cluster head satellite selects a suitable execution satellite, the task source unloads the task to the execution satellite, the execution satellite calculates the task, and the calculation results are transmitted back to the task source. In the above five processes, the task source unloads the task to the execution satellite and the calculation results are transmitted back. There will be a certain transmission delay. When the task size is consistent, selecting the execution satellite with the shortest communication distance can effectively reduce the transmission delay. Therefore, in addition to considering the influencing factors of resource occupancy and task time, this embodiment also introduces the communication delay factor into the scheduling modeling analysis. The former reduces the task waiting delay, while the latter reduces the task transmission delay.

[0093] Specifically, assuming that the angle between the line connecting the cluster head satellite and the execution satellite and the line connecting the cluster head satellite and the task source is α, the distances between the cluster head satellite and the execution satellite and the task source are d s and d d , so the communication distance can be defined as:

[0094] d=d s 2 +d d 2 -2d s d d cosα.

[0095] In summary, the embodiments of the present invention determine the target satellite corresponding to the task from three perspectives: resource optimization, waiting time, and communication distance. However, there are various tasks in the system, so the focus of analysis for different tasks should be different. For example, for the situation where there are many large resources and few small resources, scheduling focuses on resource optimization. For real-time tasks, scheduling focuses on waiting delay. For communication-intensive tasks, scheduling focuses on communication distance. Therefore, the emphasis of relevant indicators should be adjusted according to the different types of tasks, see S124 for further details.

[0096] S124. Determine a target satellite for task offloading according to the matching degree, the predicted execution time, and the communication distance.

[0097] Specifically, S124 includes: setting weights corresponding to the matching degree, the predicted execution time and the communication distance according to the mission characteristics; determining the target satellite for task offloading according to the sum of the products of the matching degree, the predicted execution time and the communication distance and the corresponding weights.

[0098] In this embodiment, the scheduling score Score of each satellite is used to determine the target satellite, which can be specifically expressed as follows:

[0099] Score=w1Fit+w2T e +w3d

[0100] Among them, w1, w2, and w3 represent the weights of matching degree, predicted execution time, and communication distance, respectively.

[0101] For example, w1 = 0.5, w2 = 0.2, and w3 = 0.3. The above values ​​are mainly based on the consideration that in the scenario of satellite edge computing offloading, most of the light computing tasks are completed locally. Therefore, the tasks offloaded to the satellite are basically heavy computing tasks that cannot be completed independently locally, which require more computing resources and have a longer task transmission delay.

[0102] In summary, this embodiment comprehensively considers three indicators: network transmission delay, container task resource fragmentation, and task running time, calculates the matching degree of each node, and finally selects the task execution node with the highest matching degree.

[0103] S130: The task source unloads the task to the target satellite for execution, and receives the task execution result of the target satellite.

[0104] In this embodiment, when the task execution is completed within the communication range, the task is terminated after the calculation task result is successfully transmitted back.

[0105] Based on the above embodiment, the method further includes the following steps performed by the task migration module:

[0106] Predict the communication distance between the target satellite and the mission source during the mission execution;

[0107] When it is predicted that the target satellite will be out of the communication range of the task source before the task is completed, the task source sends a task unloading signal to the target satellite and determines the target migration satellite from other satellite nodes in the cluster again;

[0108] The unloading on the target satellite is transferred to the target migration satellite and continued.

[0109] The constant movement of satellites in orbit causes changes in the satellite network. After a satellite has been operating in a cluster for a period of time, it may be out of the communication range of the end-side device, causing the end-side device to lose connection with the satellite. Therefore, before the satellite flies out of the communication range of the end-side device, some tasks should be migrated to other satellites in the cluster according to the task type to ensure that the tasks can continue to be executed in the cluster. The task migration process is as follows: Figure 3 shown.

[0110] In the satellite edge computing scenario, not only are the satellites in constant operation, but the task sources (aircraft, etc.) may also be in dynamic flight. Therefore, the task sources and edge computing satellite nodes need to be in the communication area to maintain a communicative geometric relationship.

[0111] like Figure 4 As shown, R erepresents the radius of the Earth, H d and H s They represent the flight altitudes of the mission source and the satellite respectively. Communication can only be achieved when the geocentric angle between the satellite and the mission source is less than α. The expression of α is:

[0112]

[0113] Assuming that the clockwise flight is the positive direction, the angular velocities of the satellite and the mission source are ω1 and ω2 respectively. Then the satellite departure time can be expressed as:

[0114]

[0115] Where α0 represents the geocentric angle between the satellite and the mission source at the initial connection.

[0116] Therefore, after t hours, the task source will fly out of the communication range with the satellite. The task migration module will send a task migration command to the satellite before it flies out of the cluster based on the satellite departure prediction result. The task can be divided into local tasks and remote tasks according to whether it is executed only in the current area. After receiving the command, the satellite can suspend the execution of the local task and migrate the task to the cluster head satellite. The satellite migration task can distribute the task to each node through the above-mentioned task scheduling module. The remote task can store the task results and transmit the task data back after it can communicate with the task source again. The present invention maintains the execution satellite within the communication range of the task source device to ensure that the task can continue to be executed.

[0117] An embodiment of the present invention further provides a satellite edge computing unloading device based on dynamic time variability, the device comprising:

[0118] The feature information extraction module is used for the task source to extract the feature information of the task through the task structure analyzer before each task offloading starts;

[0119] The task scheduling module is used to send the characteristic information to the nearest cluster head satellite through the task source, so that the cluster head satellite can determine the target satellite for task offloading according to the characteristic information of the task;

[0120] The task execution module is used to offload the task to the target satellite for execution through the task source, and receive the task execution result of the target satellite.

[0121] The task scheduling module is specifically used for:

[0122] Determine the matching degree between each satellite node and the task according to the task application resource amount and the remaining resource amount of each satellite node;

[0123] Introducing neural network similarity modeling analysis technology to predict the execution time of tasks on different satellite nodes;

[0124] Determine the communication distance of the task at different satellite nodes according to the distance between the cluster head satellite and each satellite node and the task source;

[0125] A target satellite for task offloading is determined according to the matching degree, the predicted execution time, and the communication distance.

[0126] Specifically, the cluster head satellite determines the target satellite for task offloading according to the characteristic information of the task, including:

[0127] Determine the matching degree between each satellite node and the task according to the task application resource amount and the remaining resource amount of each satellite node;

[0128] Introducing neural network similarity modeling analysis technology to predict the execution time of tasks on different satellite nodes;

[0129] Determine the communication distance of the task at different satellite nodes according to the distance between the cluster head satellite and each satellite node and the task source;

[0130] A target satellite for task offloading is determined according to the matching degree, the predicted execution time, and the communication distance.

[0131] According to the amount of resources requested by the task and the remaining amount of resources of each satellite node, the matching degree between each satellite node and the task is determined, including:

[0132] Determine the matching degree between each satellite node and the task according to the proportion of the sum of the task application resource amount and the resources used by each satellite node in the current satellite, and the variance between the task application resource amount and the remaining resources of each satellite node;

[0133] The greater the difference between the proportion and the variance, the higher the matching degree.

[0134] The neural network similarity modeling analysis technology is introduced to predict the execution time of tasks on different satellite nodes, including:

[0135] When the cluster head satellite receives the characteristic information, it combines the performance characterization network of each satellite node in the cluster to predict the running time of the task on each satellite;

[0136] When there are no idle resources available for execution in the satellites within the cluster, the current task is first added to the task waiting queue, and the greedy algorithm is used to calculate the estimated waiting time for the task to wait for resource release in each satellite node.

[0137] Optionally, determining a target satellite for task offloading according to the matching degree, the predicted execution time, and the communication distance includes:

[0138] Setting weights corresponding to the matching degree, the predicted execution time, and the communication distance, respectively, according to task characteristics;

[0139] The target satellite for task offloading is determined according to the matching degree, the predicted execution time, and the sum of the products of the communication distance and the corresponding weight.

[0140] Furthermore, the device also includes a task migration module, which is used to execute:

[0141] Predict the communication distance between the target satellite and the mission source during the mission execution;

[0142] When it is predicted that the target satellite will be out of the communication range of the task source before the task is completed, the task source sends a task unloading signal to the target satellite and determines the target migration satellite from other satellite nodes in the cluster again;

[0143] The unloading on the target satellite is transferred to the target migration satellite and continued.

[0144] A satellite edge computing unloading device based on dynamic time variability provided by an embodiment of the present invention can execute a satellite edge computing unloading method based on dynamic time variability provided by any embodiment of the present invention, and has the corresponding functional modules and beneficial effects of the execution method, which will not be repeated here.

[0145] Implementing the verification part

[0146] 1. Experimental setup

[0147] 1.1 The specific settings of the test platform server are as follows Table 1

[0148] Table 1 Specific parameters of the test server

[0149]

[0150] According to the intersatellite link communication rules, the satellites with the closest communication distance are selected to form a cluster, so five satellites form a cluster. Two clusters AB are set up respectively, with a total of ten servers for simulating satellites. The hardware settings of each server are the same.

[0151] 1.2 Dataset

[0152] The large-scale production cluster data of Alibaba's artificial intelligence platform was selected as the test data task set. The test data set tracked and recorded the data in the large-scale production cluster of Alibaba's artificial intelligence platform (a total of 1,800 servers) from July to August 2020. Its tasks are mainly AI tasks of Alibaba's machine learning platform. This dataset has been published in the NSDI'22 paper "MLaaS in the Wild: Workload Analysis and Scheduling in Large-Scale Heterogeneous GPU Clusters".

[0153] First, the unreasonable data in the data set is cleaned to meet the hardware platform settings set up for this experiment. It mainly includes the following aspects.

[0154] (1) The CPU and Mem values ​​are normalized to the maximum and minimum values. The number of CPU cores is normalized to 0 to 16 cores, and the memory size is normalized to 0 to 32000 Mi.

[0155] (2) Since the maximum number of GPUs set on the server in this experiment is 4, the number of GPUs in the dataset is normalized to 0 to 4.

[0156] (3) Filter data with missing information.

[0157] 1.3 Comparison Benchmarks

[0158] This experiment verifies the performance of DTV through simulation results, simulating the satellite edge computing scenario of a dual-star cluster with a single task source, and sets up the following three offloading schemes for comparison.

[0159] (1) Non-Cooperative Random Offloading (NC-RO): All tasks are randomly selected to be sent to satellite nodes.

[0160] (2) Cooperative Random Offloading (C-RO): The task source randomly selects a suitable cluster, and the cluster head satellite randomly selects a satellite node to send it. This strategy is different from (1) in that the task can only be executed in the selected cluster.

[0161] (3) Cooperative Greedy Offloading (C-GO): The task source selects the cluster with the most abundant resources each time, and the cluster head satellite selects the most suitable target satellite for execution based on the abundance of resources.

[0162] 2. Experimental results and conclusions

[0163] In the simulation experiment, there are two clusters with a total of 10 simulated satellites to support task offloading. A comparative study is conducted through experiments on the task waiting execution time and the total task running time under different offloading schemes. Figure 5 and Figure 6 It can be seen that compared with the non-cooperative offloading scheme, the cooperative offloading scheme has obvious advantages in task waiting time and task running time. In addition, although the C-GO scheme selects the satellite node with the most surplus resources to offload each time, it does not take into account the cost of transmission delay and resource waiting delay, so it takes more time than the DTV scheme. DTV predicts the task running time and masks the transmission delay by waiting time, so the total task execution time is lower.

[0164] Figure 7 The resource utilization data under different offloading strategies are shown. Among them, the C-RO strategy only selects execution satellites in the selected cluster, and the number of executable satellites to be selected is significantly less than that of the NC-RO strategy, so the resource utilization is not as good as the NC-RO strategy. In addition, the C-GO strategy tends to select execution satellites with greater computing power, which will generate more resource fragments and lead to lower resource utilization. DTV uses a resource optimization strategy to offload tasks. Because it generates less resource fragments, the resource utilization is higher than other strategies.

[0165] Note that the above are only preferred embodiments of the present invention and the technical principles used. Those skilled in the art will understand that the present invention is not limited to the specific embodiments described herein, and that various obvious changes, readjustments and substitutions can be made by those skilled in the art without departing from the scope of protection of the present invention. Therefore, although the present invention has been described in more detail through the above embodiments, the present invention is not limited to the above embodiments, and may include more other equivalent embodiments without departing from the concept of the present invention, and the scope of the present invention is determined by the scope of the appended claims.

Claims

1. A satellite edge computing offloading method based on dynamic time-varying properties, characterized in that: include: S110, before each task offloading starts, the task source extracts the characteristic information of the task through the task structure analyzer; S120, the task source sends the characteristic information to the nearest cluster head satellite, so that the cluster head satellite determines the target satellite for task offloading according to the characteristic information of the task; S130, the task source unloads the task to the target satellite for execution, and receives the task execution result of the target satellite; The cluster head satellite determines the target satellite for task offloading according to the characteristic information of the task, including: Determine the matching degree between each satellite node and the task according to the task application resource amount and the remaining resource amount of each satellite node; Introducing neural network similarity modeling analysis technology to predict the execution time of tasks on different satellite nodes; Determine the communication distance of the task at different satellite nodes according to the distance between the cluster head satellite and each satellite node and the task source; Determine a target satellite for task offloading according to the matching degree, the predicted execution time, and the communication distance; According to the amount of resources requested by the task and the remaining amount of resources of each satellite node, the matching degree between each satellite node and the task is determined, including: Determine the matching degree between each satellite node and the task according to the proportion of the sum of the task application resource amount and the resources used by each satellite node in the current satellite, and the variance between the task application resource amount and the remaining resources of each satellite node; The greater the difference between the proportion and the variance, the higher the matching degree; The neural network similarity modeling analysis technology is introduced to predict the execution time of tasks on different satellite nodes, including: When the cluster head satellite receives the characteristic information, it combines the performance characterization network of each satellite node in the cluster to predict the running time of the task on each satellite; When there are no idle resources available for execution in the satellites within the cluster, the current task is first added to the task waiting queue, and the greedy algorithm is used to calculate the estimated waiting time for the task to wait for resource release in each satellite node.

2. The method according to claim 1, characterized in that Determining a target satellite for task offloading according to the matching degree, the predicted execution time, and the communication distance includes: Setting weights corresponding to the matching degree, the predicted execution time, and the communication distance, respectively, according to task characteristics; The target satellite for task offloading is determined according to the matching degree, the predicted execution time, and the sum of the products of the communication distance and the corresponding weight.

3. The method according to claim 1, characterized in that The method further includes: Predict the communication distance between the target satellite and the mission source during the mission execution; When it is predicted that the target satellite will be out of the communication range of the task source before the task is completed, the task source sends a task unloading signal to the target satellite and determines the target migration satellite from other satellite nodes in the cluster again; The unloading on the target satellite is transferred to the target migration satellite and continued.

4. A satellite edge computing unloading device based on dynamic time-varying properties, characterized in that: include: The feature information extraction module is used for the task source to extract the feature information of the task through the task structure analyzer before each task offloading starts; The task scheduling module is used to send the characteristic information to the nearest cluster head satellite through the task source, so that the cluster head satellite can determine the target satellite for task offloading according to the characteristic information of the task; A task execution module, used for offloading the task to the target satellite for execution through the task source, and receiving the task execution result of the target satellite; The task scheduling module is specifically used for: Determine the matching degree between each satellite node and the task according to the task application resource amount and the remaining resource amount of each satellite node; Introducing neural network similarity modeling analysis technology to predict the execution time of tasks on different satellite nodes; Determine the communication distance of the task at different satellite nodes according to the distance between the cluster head satellite and each satellite node and the task source; Determine a target satellite for task offloading according to the matching degree, the predicted execution time, and the communication distance; According to the amount of resources requested by the task and the remaining amount of resources of each satellite node, the matching degree between each satellite node and the task is determined, including: Determine the matching degree between each satellite node and the task according to the proportion of the sum of the task application resource amount and the resources used by each satellite node in the current satellite, and the variance between the task application resource amount and the remaining resources of each satellite node; The greater the difference between the proportion and the variance, the higher the matching degree; The neural network similarity modeling analysis technology is introduced to predict the execution time of tasks on different satellite nodes, including: When the cluster head satellite receives the characteristic information, it combines the performance characterization network of each satellite node in the cluster to predict the running time of the task on each satellite; When there are no idle resources available for execution in the satellites within the cluster, the current task is first added to the task waiting queue, and the greedy algorithm is used to calculate the estimated waiting time for the task to wait for resource release in each satellite node.

5. The device according to claim 4, characterized in that The device also includes a task migration module, which is used to execute: Predict the communication distance between the target satellite and the mission source during the mission execution; When it is predicted that the target satellite will be out of the communication range of the task source before the task is completed, the task source sends a task unloading signal to the target satellite and determines the target migration satellite from other satellite nodes in the cluster again; The unloading on the target satellite is transferred to the target migration satellite and continued.

Images