CN113608830A

CN113608830A - VNF migration method and device based on fault prediction

Info

Publication number: CN113608830A
Application number: CN202110791840.4A
Authority: CN
Inventors: 张�浩; 刘川; 刘世栋; 胡博; 杨超; 徐思雅; 邵苏杰; 童日明; 雷承昊; 贺文晨
Original assignee: State Grid Corp of China SGCC; Beijing University of Posts and Telecommunications; Global Energy Interconnection Research Institute; Information and Telecommunication Branch of State Grid Liaoning Electric Power Co Ltd
Current assignee: State Grid Corp of China SGCC; Beijing University of Posts and Telecommunications; Global Energy Interconnection Research Institute; Information and Telecommunication Branch of State Grid Liaoning Electric Power Co Ltd
Priority date: 2021-07-13
Filing date: 2021-07-13
Publication date: 2021-11-05

Abstract

The invention provides a VNF migration method and a VNF migration device based on fault prediction, wherein the VNF migration method comprises the following steps: predicting the fault condition of a physical node in an edge network at a future moment based on a physical node fault prediction model of a BP neural network; and constructing a VNF instance migration optimization model based on the prediction result, the node cost and the link cost determined by the fault prediction model, and migrating the VNF instance on the fault node to a normal node based on the VNF instance migration optimization model. According to the method and the device, before the VNF instances actually fail, the failure conditions of the nodes in the network at the future moment can be predicted based on the physical node failure prediction model of the BP neural network, so that the VNF instances on the failed nodes are migrated to normal nodes in advance, and the continuous normal operation of the service is guaranteed.

Description

VNF migration method and device based on fault prediction

Technical Field

The present invention relates to the field of communications technologies, and in particular, to a VNF migration method and apparatus based on fault prediction.

Background

The mobile edge computing MEC is used as a technology for enhancing and expanding computing capacity, and resources required by business are provided by installing small-sized limited-resource cloud infrastructure on an edge network, so that a business data processing loop is short, service response is fast, and cloud pressure is effectively relieved.

In a mobile edge network scenario, a virtual resource allocation mechanism is adopted to provide resources required by services in combination with NFV and edge computing technologies. The NFV adopts a more flexible network configuration mode, and replaces dedicated hardware with a software instance (VNF instance) running in a virtualization environment, so that the distribution of network resources is more extensible and elastic, and a more efficient and flexible management and running mechanism is provided for network functions, thereby greatly reducing the overall cost.

High availability is crucial to an edge network, most of currently proposed virtual resource failure protection mechanisms have a hysteresis problem, that is, migration of a VNF instance is performed after a failure occurs, which may cause service interruption and bring a bad use experience to a user. In addition, some current solutions related to VNF instance migration, such as integer linear programming and heuristic algorithms, have either huge solution search space or need to manually consider optimal VNF resource size and location, and these shortcomings limit their application in complex networks. Therefore, today with strict service availability requirements and increasingly complex networks, how to more effectively implement virtual resource failure protection and VNF instance migration is an urgent problem to be solved.

Disclosure of Invention

To solve the problems in the prior art, embodiments of the present invention provide a VNF migration method and apparatus based on failure prediction.

In a first aspect, an embodiment of the present invention provides a VNF migration method based on failure prediction, including:

predicting the fault condition of a physical node in an edge network at a future moment based on a physical node fault prediction model of a BP neural network;

and constructing a VNF instance migration optimization model based on the prediction result, the node cost and the link cost determined by the fault prediction model, and migrating the VNF instance on the fault node to a normal node based on the VNF instance migration optimization model.

Further, migrating the VNF instance on the failed node to the normal node based on the VNF instance migration optimization model specifically includes:

the VNF migration scheme of the VNF instance is determined based on a deep reinforcement learning VNF migration algorithm, and the VNF instance on the fault node is migrated to the normal node; wherein the VNF migration algorithm is a solution algorithm for cost and priority determination.

Further, the physical node fault prediction model based on the BP neural network is as follows: and taking SMART characteristic data of the physical node hard disk as input data and a fault prediction result corresponding to the SMART characteristic data of the physical node hard disk as output data, and training the output data based on a deep learning algorithm to obtain the model.

Further, still include:

determining a misclassification cost by using a first relation model; taking the misclassification cost as an evaluation index of the fault prediction model; wherein the first relationship model is:

Cos t_mis＝Cos t₁*Num_FP+Cos t₂*Num_FN

wherein, Cos t_misDenotes the misclassification cost, Num_FPNumber of samples, Num, predicted as the fault is true normal_FNRepresents the number of samples predicted to be normal and truly faulty, Cos t₁Represents the loss, Cos t, caused by misclassification of a normal hard disk as a failed hard disk₂Indicating the loss caused by misclassification of a failed hard disk as a normal hard disk.

In a second aspect, an embodiment of the present invention provides a VNF migration apparatus based on failure prediction, including:

the prediction module is used for predicting the fault condition of the physical nodes in the edge network at the future time based on a physical node fault prediction model of the BP neural network;

and the migration module is used for constructing a VNF instance migration optimization model based on the prediction result, the node cost and the link cost determined by the fault prediction model, and migrating the VNF instance on the fault node to a normal node based on the VNF instance migration optimization model.

Further, when the migration module executes migration of the VNF instance on the failed node to the normal node based on the VNF instance migration optimization model, the migration module is specifically configured to:

Further, the BP neural network-based physical node fault prediction model in the prediction module is: and taking SMART characteristic data of the physical node hard disk as input data and a fault prediction result corresponding to the SMART characteristic data of the physical node hard disk as output data, and training the output data based on a deep learning algorithm to obtain the model.

Further, the prediction module is further configured to:

Cos t_mis＝Cos t₁*Num_FP+Cos t₂*Num_FN

In a third aspect, an embodiment of the present invention further provides an electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor, where the processor executes the computer program to implement the steps of the VNF migration method based on failure prediction according to the first aspect.

In a fourth aspect, the present invention further provides a non-transitory computer-readable storage medium, on which a computer program is stored, which, when being executed by a processor, implements the steps of the failure prediction based VNF migration method according to the first aspect.

According to the technical scheme, the fault prediction-based VNF migration method and the fault prediction-based VNF migration device predict the fault condition of the physical nodes in the edge network at the future time through the BP neural network-based physical node fault prediction model; and constructing a VNF instance migration optimization model based on the prediction result, the node cost and the link cost determined by the fault prediction model, and migrating the VNF instance on the fault node to a normal node based on the VNF instance migration optimization model. According to the method and the device, before the VNF instances actually fail, the failure conditions of the nodes in the network at the future moment can be predicted based on the physical node failure prediction model of the BP neural network, so that the VNF instances on the failed nodes are migrated to normal nodes in advance, and the continuous normal operation of the service is guaranteed.

Drawings

In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below, and it is obvious that the drawings in the following description are some embodiments of the present invention, and for those skilled in the art, other drawings can be obtained according to these drawings without creative efforts.

Fig. 1 is a schematic flowchart of a VNF migration method based on failure prediction according to an embodiment of the present invention;

fig. 2 is a schematic diagram of a BP neural network-based physical node fault prediction model according to an embodiment of the present invention;

fig. 3 is a schematic diagram illustrating a relationship between a deployment cost of a k-type VNF instance on a node v and a resource occupancy rate according to an embodiment of the present invention;

FIG. 4 is a schematic diagram illustrating an example migration of a VNF according to an embodiment of the present invention;

fig. 5 is a schematic diagram of an example migration model of a DQN-based VNF according to an embodiment of the present invention;

fig. 6 is a schematic diagram of a training process of the DQN algorithm according to an embodiment of the present invention;

figure 7 is a flowchart of a VNF instance migration algorithm based on priority according to an embodiment of the present invention;

FIG. 8 is a diagram illustrating a comparison of BP and SVM provided by an embodiment of the present invention over TPR;

FIG. 9 is a diagram illustrating comparison between BP and SVM provided by an embodiment of the present invention on FPR;

FIG. 10 is a diagram illustrating comparison between BP and SVM provided by another embodiment of the present invention on FPR;

fig. 11 is a schematic diagram illustrating a variation of the return value of the DQN algorithm according to the training round in accordance with an embodiment of the present invention;

fig. 12 is a schematic diagram illustrating SFC acceptance ratio comparison between a service migration algorithm for failure prediction and a service migration algorithm based on failure triggering according to an embodiment of the present invention;

FIG. 13 is a diagram illustrating a comparison of migration costs of three algorithms provided in accordance with an embodiment of the present invention;

fig. 14 is a schematic structural diagram of a VNF migration apparatus based on failure prediction according to an embodiment of the present invention;

fig. 15 is a schematic physical structure diagram of an electronic device according to an embodiment of the present invention.

Detailed Description

In order to make the objects, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are some embodiments, but not all embodiments, of the present invention. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention. The VNF migration method based on failure prediction provided by the present invention will be explained and explained in detail by specific embodiments.

Fig. 1 is a schematic flowchart of a VNF migration method based on failure prediction according to an embodiment of the present invention; as shown in fig. 1, the method includes:

step 101: and predicting the fault condition of the physical nodes in the edge network at the future time based on a physical node fault prediction model of the BP neural network.

Step 102: and constructing a VNF instance migration optimization model based on the prediction result, the node cost and the link cost determined by the fault prediction model, and migrating the VNF instance on the fault node to a normal node based on the VNF instance migration optimization model.

In this embodiment, it can be understood that, in the VNF migration method based on failure prediction provided in the embodiment of the present invention, a physical node failure prediction model based on a BP neural network is first established, and a failure condition of a physical node in an edge network at a future time is predicted, that is, a failure node is determined. And then, planning a VNF instance migration optimization model based on the physical node fault prediction model result and considering the node cost and the link cost, and migrating the VNF instance on the fault node to a normal node, namely a non-fault node, so as to realize VNF migration.

In this embodiment, for model establishment, it should be noted that, the physical node fault prediction model based on the BP neural network: a large number of physical hard drives are used in the mobile edge network, carrying VNF instances supplied by the NFV provider. The hard disk failure is one of the most important causes of the failure of the general-purpose server, and in the embodiment, the hard disk failure is taken as a physical node failure, and a physical node failure prediction model (PNFP-BP) based on a BP neural network is proposed, as shown in fig. 2. Specifically, the BP neural network adopts a typical three-layer feedforward network structure, an input layer, a hidden layer and an output layer. SMART characteristic data x of hard disk is received by input layer_iThe hidden layer is used for abstracting the characteristics of input data to another dimensional space to show more abstract characteristics, the characteristics can be better linearly divided, and the output layer outputs results

. The input layer of hard disk SMART data can be from a public data set in a real production environment of Blackblaze company, the initial data set is 23-dimensional, and the redundant features are removed through feature selection to obtain 19-dimensional data, wherein the 19-dimensional data comprises the number of sectors migrated by the hard disk, the internal temperature of the hard disk and the like. The active functions of the hidden layer and the output layer are both Sigmoid functions, and the output layer is

The value range of (1) is 0-1. For a binary classification task, each sample contains two labels, one is the true case of the sample, and the other is the predicted result of the classification algorithm. The prediction results of the binary classification problem can be classified into the following four categories according to the situation: 1. true example (True P)ositive, TP): the output result of the model is 1, and the actual result is 1; 2. false Positive (FP): the output result of the model is 1, and the actual result is 0; 3. true Negative (TN): the output result of the model is 0, and the actual result is 0; 4. false Negative (FN): the output result of the model is 0, and the actual result is 1; in this embodiment, a failed hard disk is defined as a positive example, and a normally operating hard disk is defined as a negative example. Using two indexes, True Positive Rate (TPR) and False Positive Rate (FPR), the calculation formulas are shown in formula (1) and formula (2).

In addition to TPR and FPR, the misclassification cost Cos t is also presented_misCos t as an index for comprehensively measuring the performance of a fault prediction model_misIs calculated as shown in equation (3).

Cos t_mis＝Cos t₁*Num_FP+Cos t₂*Num_FN (3)

Wherein, Num_FPNumber of samples, Num, predicted as the fault is true normal_FNIndicating the number of samples predicted to be normal and truly faulty. Cos t₁The loss caused by misclassification of a normal hard disk as a failed hard disk, namely, the cost loss caused by unnecessary VNF instance migration. Cos t₂Is a loss caused by misclassification of a failed hard disk as a normal hard disk. Cos t₁And Cos t₂Is set by the service provider, and Cos t is defined in the embodiment₁And Cos t₂In a ratio of 1: 10.

Modeling the mobile edge physical network with an undirected graph G ═ V, L, where V and L represent the set of nodes and links in the edge physical network, respectively. In the network, u, V epsilon V is used for representing two physical nodes, and uv epsilon L represents a connection physical nodeThe links of points u and v. For any node V e V in the edge network, one or more general servers can be connected, and the processing capacity of the general servers can be expressed as C_vI.e. the maximum capacity that the node can provide for the VNF. Similarly, there is a bandwidth capacity B for any link/e L_lI.e., the amount of available bandwidth. Remaining node capacity C of physical network node v at time t_v(t) indicates that the link residual bandwidth is B_l(t) represents.

VNF instance types that NFV service providers can deploy and operate are total | K |, denoted as K ═ VNF₁,VNF₂,...,VNF_|K|}. For a VNF of type K ∈ K, C_kIndicating the processing power that this example has. The interval of the fault prediction model FNFP-BP prediction time based on the BP neural network is delta t_fau-preOn the order of hours. By gamma_vThe v utilization rate of the node is represented, the calculation method is shown as the formula (4),

C_v(t) represents the remaining processing capacity of node v at time t. Binary variable r for failure or not of physical node_vIs represented by_vEqual to 0 node failure, r_vNode 1 is normal. In this embodiment, the failure condition of the VNF instance deployed on the node is consistent with that of the physical node. The set of failed and normal nodes are denoted F and W, respectively. v-k_MRepresenting the number of k-type VNF instances on the node v, and using binary variables for the fault condition of the mth k-type VNF instance on the node v

Which is represented by the formula (5),

is shown at time t₃Whether to migrate the mth VNF instance of type k on node v onto node v',

representing the n-th SFC affected by the m-th VNF instance of type k on the failed node v.

The VNF instance deployment cost unit price of the node v is calculated by equation (6),

VNF instance deployment cost unit price and node resource occupancy rate gamma_vThe relationship of (c) is shown in fig. 3. When gamma is_v≤U₀In time, k-type VNF instance deployment cost and resource occupancy γ_vIn a linear relationship, γ_v＞U₀In time, the deployment cost and the occupancy rate are in an exponential relationship, and the size of the deployment cost can be represented by the area of the rectangle in fig. 3.

Assuming that a total of I service chains provide services in the network at the predicted time t of the FNFP-BP model, SFC_iRepresents the ith service chain, and the routing path is

the residual bandwidth of the adjacent two-node link at the time t is B_e(t), the set of paths between two physical nodes u and v is Path_uvResidual bandwidth of path B_pathAnd (t) is the minimum residual bandwidth in the path coverage link, as shown in equation (7).

B_path(t)＝Min B_e(t),e∈path,path∈Path_uv (7)

A VNF instance migration diagram is shown in fig. 4. When r is_vWhen the time is 0, the result of prediction of the FNFP-BP model at time t is that the node v may fail, and therefore r to which the VNF instance deployed on the node needs to be migrated is required_v1 on the physical node. In addition, the embodiment considers the QoS requirements of different services when performing service chain traffic routing, divides the service types into three classes,namely EF, AF and BE class traffic. 1. EF characteristics: the reliability is high, and the time delay requirement is high; 2. AF characteristics: the bandwidth requirement is high, the time delay requirement is high, and the reliability requirement is not high; 3. BE is characterized in that: the reliability requirement is high, the time delay requirement is not high, and the bandwidth requirement is very low; a specific service request can be represented by a six-tuple: SFC (SFC)_cpu,SFC_band,SFC_delay,r_H,b_H,t_L) Wherein SFC_cpu,SFC_band,SFC_delayRespectively representing the CPU requirements, bandwidth requirements, and latency requirements of the service. By comparing with each parameter threshold value corresponding to service requirement, whether the SFC is required to be high in reliability, high in bandwidth and low in time delay can be judged, and binary variable r is respectively used for comparison results_H、b_H、t_LIt means that if it is 1, otherwise it is 0. Thus, the parameters for three types of traffic are available as follows: 1. and (4) EF type service: r is_H＝1，b_H＝0,t _L1 is ═ 1; 2. AF service r_H＝0，b_H＝1，t _L1 is ═ 1; 3. BE service r_H＝1，b_H＝0，t _L0. The calculation formula defining the service priority is defined as formula (8), and the traffic routing is performed according to the service priority from high to low.

QoS(SFC)＝r_Ht_L+(1-r_Ht_L)(P_C*SFC_cpu+P_B*SFC_band+P_D*SFC_delay) (8)

Wherein r is_Ht_LIndicating whether the service belongs to the EF class service, if r_Ht_LIf it is 1, then it is EF type service, and its priority is highest, and its qos (sr) is 1; if r is_Ht_LIf it is 0, then not EF type service, and according to every parameter of service making priority calculation, P_cpu、P_B、P_delayRespectively, the weight factors of CPU computing power, link bandwidth, SR request time delay, SFC_cpu、SFC_band、SFC_delayNormalization processing is respectively carried out, and the following conditions are met:

P_cpu+P_B+P_delay＝1 (9)

the calculated range is 0 < QoS (SFC) < 1. And then, carrying out routing planning on the service chain flow according to the descending order of the QoS (SFC) value, wherein the higher the QoS (SFC) value is, the higher the priority is, and the smaller the QoS (SFC) value is, the lower the priority is. The VNF instances are migrated with a certain priority, and the calculation formula is shown as a formula (10),

wherein the content of the first and second substances,

indicating whether the mth k-type VNF instance on the failed node v affects the service chain SFC_i，QoS(SFC_i) For service chain SFC_iThe priority of (c) is calculated by the formula (8).

When the VNF instance is copied and deployed, the VNF instance on the node v still provides service for the service function chain in the network, and it is ensured that the service is not interrupted. After the VNF instance is migrated, the original path cannot provide service for the traffic, and the path is re-planned to enable the traffic of the affected part to pass through the VNF instance again in sequence to complete the service chain request. The dashed arrows in the figure represent the VNF instance migration process, and the solid bold arrows represent the re-planned paths. Fig. 4 is a schematic diagram of VNF instance migration, and in actual migration, cost of different types of VNF instances deployed on different nodes and path cost caused by VNF instances deployed on different nodes need to be considered, which are mutually influenced, and operation cost of the NFV provider can be greatly reduced by reasonably selecting a migration target node for a VNF instance on a failed node. Note that, taking normal nodes adjacent to the failed node before and after as a start point and an end point, only performing path re-planning on the traffic between the start point and the end point, where the traffic passes through the end point, and as shown in the figure, only the part of the thick solid line is a re-planned path, and the path where the thick solid line and the thin solid line coexist is a path that remains unchanged, the change of the network state is effectively reduced, and frequent reconfiguration of devices in the network is avoided.

In this embodimentIn the above description, for an optimization problem model, that is, a VNF instance migration optimization model, it should be noted that in this embodiment, migration of a VNF instance on a failed node predicted by an FNFP-BP model to a normal node is considered, so as to ensure deployment cost Cos t during VNF instance migration_migPath cost Cos t due to different deployment scenarios_pathThe sum is minimal. Cos t_migIs shown in equation (11), and consists of the total cost of all VNF instance deployments on all failed nodes.

Cos t_pathThe calculation method (2) is shown as the formula (12), and consists of the total cost of path re-planning by all SFCs affected by the fault node.

Where β represents the cost of use per bandwidth, dist_p(u, v) denotes the length of the path between nodes u and v, which respectively denote the starting and ending points of the physical nodes before and after, i.e. the affected routes, adjacent to the failed node. Thus, the optimization objective can be expressed as,

Minimize Cos t_mig+Cos t_path (13)

the constraint condition is that,

the constraint (14) ensures that at time t₃The processing capacity of the mth k-type VNF instance on the failed node v does not exceed the remaining capacity of the target migration node. The constraint (15) ensures that it is affected by the mth k-type VNF instance on the failed node vThe service chain does not exceed the remaining bandwidth of the path when the path re-planning is carried out.

And the optimization model modeling of the VNF instance migration problem is completed, and the optimization target is that the VNF instance migration cost and the traffic path re-planning cost are minimum. In this section, the optimization problem described above is modeled as a markov decision process, and then a DQN-based VNF instance migration algorithm (VNFMA-DQN) is proposed. 1. Markov decision process: the VNF instance migration problem on this failed node is modeled as a markov decision process { S, a, R, P }, S, a, R, P, representing state space, action space, reward function, and transition probability, respectively. Each VNF instance to be migrated on a failed node is considered to be an agent in deep reinforcement learning. And the intelligent agent interacts with the network environment, continuously tries different migration schemes, acquires experience and finally obtains the optimal scheme for VNF instance migration.

(1) State space S: in the proposed model described above, there are | F | faulty nodes and | W | normal nodes in the network in common. The state space of each VNF instance may be represented by a one-dimensional vector S,

wherein w is the coordinate of the migration target node in the network,

for the number of service chains affected by the mth k-type VNF instance on the failed node f,

and selecting a routing path of the nth affected service chain according to a shortest path algorithm. Due to the limited bandwidth, the selection sequence is that the QoS values of all service chains are arranged from high to low, and the operation of high-priority services is guaranteed.

(2) An action space A: according to the state information of the network, namely the deployment prices of VNF instances of different nodes and the residual bandwidth conditions of different links, the agent needs to select a proper node for VNF migration. The present embodiment defines an action space a ═ { W ' }, W ' ∈ W, which means that the migration position of the VNF instance is turned from W to W '.

(3) The reward function R: the VNF instances migrate to different nodes, i.e. the agents are in different states and the NFV providers have different running costs. In the DQN-based VNF migration algorithm, the cost change migrated to different nodes is used as an immediate reward for reinforcement learning, including the deployment cost of different nodes and the link cost caused by deployment on different nodes, and is calculated as formula (17), where ρ is a weight coefficient. Denormal is penalized if the state is shifted in the correct direction, resulting in a positive immediate reward.

R(s,s')＝ρ(Cos t(s)-Cos t(s'))，ρ＞0 (17)

Cos t(s) is the cost of the agent in state s, calculated by equation (18),

wherein, dist_path(u_n,v_n) Starting point u for routing affected part of nth service chain_nAnd endpoint v_nThe shortest path length between.

Training a DQN-based VNF instance migration algorithm:

DRL setting: the DQN-based VNF instance migration model as shown in fig. 5, the DRL setup consists of two parts, the mobile edge network environment and the agent. The network environment consists of cloudlet server nodes in the mobile edge network and links between different nodes. And the agent is a VNF instance to be migrated on the fault node v, acquires the resource price of the server node and the link residual bandwidth through continuous interaction with the network environment, continuously performs migration attempt, and finally finds the optimal target node for VNF migration.

DRL training: the agent learns the optimal migration strategy of the VNF instance by acquiring experience through interaction with the network environment. The Q-learning algorithm uses a Q function, i.e., a motion cost function Q (S)_t,A_t) To indicate that the agent is in state S_tTake action A_tIs of great valueSmall, Q (S)_t,A_t) The update method (2) is as shown in the formula (19), and recorded in a Q table.

Q(S_t,A_t)＝Q(S_t,A_t)+lr(R_t+1+λmax_aQ(S_t+1,A_t)-Q(S_t,A_t)) (19)

Wherein lr is the learning rate of the Q function, and λ is the reward discount factor. After updating the Q table, the agent selects the action a with the maximum Q value according to the current state s, and the state is converted to s'. However, this is hardly feasible in this embodiment because the states are really too many and the way of using the table does not hold at all. Thus, the present embodiment uses deep Q learning to solve this problem. The deep Q learning uses a deep neural network Q (s, a; theta), which is a weight parameter of the neural network, to replace the Q table in the Q learning algorithm. By learning the samples, θ converges to an optimal value, and the value of the action is obtained by inputting the state and the action. However, there are some disadvantages in the practical application process: (1) reinforcement learning does not require a training set, it returns a reward value only through the environment. Meanwhile, it also has noise and time delay problems, so the reward values of multiple states are 0; (2) each sample of deep learning is independent, and the current state value of reinforcement learning depends on the subsequent state return value; (3) instability problems may arise when a value function is represented by a non-linear network. In order to improve the learning capability of DQN, two techniques, an empirical playback mechanism and a target network, are added in its iteration process, as shown in fig. 5. The empirical playback is to store data obtained by exploring the environment by the intelligent agent in a memory base and then update the parameters of the neural network in a random sampling mode. The sample data stored in the memory bank uses a quadruple < s_t,a_t,r_t,s_t+1Is > represents, s_tIs the current state, a_tFor the current action, r_tFor the prize value, s_t+1The state is the next moment. Its main contribution is to overcome the problem of data correlation and non-stationary distribution of empirical data. Typically, the variance of parameter updates increases due to the correlation between successive sample data, a mechanism that does soThe variance of the parameter update can be reduced. The random selection destroys the correlation between the empirical data, making the neural network update more efficient. In addition, because one sample can be used for multiple times, the utilization rate of the empirical data can be improved.

The target network targetNet for predicting Q reality has the same structure as the predicted Q estimation network evalNet but different weight parameters. The parameter in the Q-evaluation network is the latest parameter θ, the output is Q (s, a; θ), θ varies with each time step t. The parameter in the Q-target network is the parameter theta in the Q-evaluation network long ago^-The output is Q (s, a; theta)^-)，θ^-Fixed for a period of time, and assigned by θ after a number of time steps. The loss function is given by (20),

L(θ)＝(r+λmax aQ(s',a'；θ-)-Q(s,a；θ))² (20)

after the loss function is determined, the gradient is calculated using a back propagation algorithm and the weight parameters of the entire network are updated using a random gradient descent. In the training process, an epsilon greedy algorithm is used for realizing the balance between the exploration of the intelligent agent and the utilization of the existing experience, and the local optimal solution is avoided. The e greedy algorithm enables the agent to have the probability of e to explore the environment, and the probability of 1-e utilizes the existing experience. A flow chart of the DQN-based DRL training algorithm is shown in fig. 6. And after the training is finished, obtaining the optimal target node for VNF instance migration. In the application phase, the present embodiment proposes a VNF instance migration algorithm based on priority, and the algorithm flow is shown in fig. 7.

As can be seen from the above technical solutions, in the VNF migration method based on failure prediction provided in the embodiment of the present invention, the failure condition of the physical node in the edge network at a future time is predicted through the physical node failure prediction model based on the BP neural network; and constructing a VNF instance migration optimization model based on the prediction result, the node cost and the link cost determined by the fault prediction model, and migrating the VNF instance on the fault node to a normal node based on the VNF instance migration optimization model. According to the method and the device, before the VNF instances actually fail, the failure conditions of the nodes in the network at the future moment can be predicted based on the physical node failure prediction model of the BP neural network, so that the VNF instances on the failed nodes are migrated to normal nodes in advance, and the continuous normal operation of the service is guaranteed.

On the basis of the foregoing embodiment, in this embodiment, the migrating the VNF instance on the failed node to the normal node based on the VNF instance migration optimization model specifically includes:

As can be seen from the above technical solutions, in the failure prediction-based VNF migration method provided in the embodiments of the present invention, the VNF instance on the failed node is migrated to the normal node through the VNF migration algorithm based on deep reinforcement learning, and since the VNF migration algorithm is a solution algorithm oriented to cost and priority determination, it is possible to minimize the cost while ensuring that the service is performed normally.

On the basis of the foregoing embodiment, in this embodiment, the physical node fault prediction model based on the BP neural network is: and taking SMART characteristic data of the physical node hard disk as input data and a fault prediction result corresponding to the SMART characteristic data of the physical node hard disk as output data, and training the output data based on a deep learning algorithm to obtain the model.

According to the technical scheme, the VNF migration method based on the fault prediction provided by the embodiment of the invention comprises the steps of firstly establishing a physical node fault prediction model based on a BP neural network, and predicting whether a physical node in the network has a fault or not at a future time according to SMART data of a hard disk; and then, according to the failure prediction result and the cost and priority, providing a service migration algorithm based on deep reinforcement learning DQN, migrating the influenced VNF instances on the failure nodes to other normal nodes, and minimizing the cost.

On the basis of the above embodiment, in this embodiment, the method further includes:

Cos t_mis＝Cos t₁*Num_FP+Cos t₂*Num_FN

According to the technical scheme, the fault prediction-based VNF migration method provided by the embodiment of the invention designs the model evaluation index-the mis-classification cost (namely the mis-classification cost), and realizes the comprehensive evaluation of the fault prediction model.

In order to better understand the present invention, the following examples are further provided to illustrate the content of the present invention, but the present invention is not limited to the following examples.

In this embodiment, assuming that each piece of traffic traverses a service chain consisting of 2-4 VNFs chosen at random, the delay for deploying each VNF instance is 30 seconds. For the topology of the network, the number of nodes of the simulated network is 30, and the 30 nodes each have the capability of providing virtual computing resources, that is, VNF instances can be arranged on the nodes. The initial processing capacity of the node is randomly set to [2.4, 3.6 ]]Between Gbps, the initial bandwidth of the link is randomly set at [6, 10 ]]Gbps. For the

Is provided with

In this embodiment, for the prediction effect of the physical node fault prediction model based on the BP neural network, it should be noted that, in this embodiment, the BP neural network is built based on the MATLAB toolbox, the dimension of the input layer is consistent with the dimension of the sample data set and is 19 dimensions, the hidden layer and the output layer are sigmoid functions, the learning rate is 0.01, and the maximum iteration number is 500.

One important problem in fault prediction models is: samples collected on the fault hard disk and within a time period of being far from the fault time can be used as positive samples, and the effect of the whole classification model is obviously influenced. Four time points, 2 hours, 4 hours, 8 hours and 16 hours before the failure time, were selected for this example. Table 1 shows the predicted effect of the model after training with the failed hard disk positive sample at different time points.

TABLE 1 BP model prediction Effect

At different time points	TPR(％)	FPR(％)	Cos t _mis
				2 hours	93.54％	0.42％	856
4 hours	96.71％	0.77％	714
				8 hours	98.33％	1.09％	712
16 hours	99.45％	1.55％	875

As can be seen from Table 1, as the time point moves forward, both TPR and FPR of the BP fault prediction model are continuously increased, but the misclassification cost Cos t_misShows a decreasing-then-increasing process, at this time point of 8 hours Cos t_misThe lowest is reached. In this case, the TPR of the BP model was 98.33%, the FPR was 1.09%, and the Cos t was_misAt 712, this shows that although the model predicts the failure hard disk with higher and higher probability (TPR), the probability that the normal hard disk is mistaken for the failure hard disk is increased (FPR), and the two jointly affect the misclassification cost Cos t_mis。

Fig. 8 and 9 show a comparison of the TPR and FPR indices for the SVM algorithm and the BP algorithm of this embodiment. It can be seen that the trend of change of the TPR and the FPR predicted by the SVM algorithm is the same as that of the BP algorithm, and both the trend of change of the TPR and the FPR increase with the transition of the time point, and the TPR of the BP algorithm is always greater than that of the SVM algorithm, while the FPR is always smaller than that of the other. Therefore, the BP algorithm has better performance in fault node prediction. FIG. 10 shows two algorithms at Cos t_misIn contrast, it can be seen that the loss of the BP algorithm due to misclassification is significantly less than that of the SVM algorithm. Cost of loss Cos t of BP algorithm for the same test sample at time point 8_mis712 Cos t of SVM Algorithm_mis1950, Cos t of SVM Algorithm_misMore than twice as much as the BP algorithm. Therefore, the BP algorithm proposed by the present embodiment can significantly reduce the operation cost of the service provider.

In this embodiment, for VNF instance migration algorithm performance, it should be noted that, first, the convergence condition during DQN algorithm training is verified through simulation. Fig. 11 shows the variation of the proposed DQN algorithm with training rounds at learning rates of 0.001, 0.01, 0.05 and 0.1 respectively. As can be seen from fig. 11, in the training step of the algorithm, the learning rate affects the value of the learning reward, because the learning rate represents the learning step size for realizing the convergence of the reward function, a large learning rate may miss the global optimum of the learning process, and a small learning rate requires more steps to reach the global optimum. From the simulation results, a learning rate of 0.01 performs best in the simulation scenario, converges at a return value of 1, and its learning speed is acceptable, enabling the reward function to converge quickly.

Next, simulation verifies the advantages of the overall framework of VNF instance migration based on node failure prediction. Fig. 12 shows a comparison of SFC acceptance rates between a service migration algorithm based on failure prediction and a service migration algorithm based on failure triggering. The service migration algorithm based on the fault trigger means that when a physical node actually fails, a service provider migrates a VNF instance on the failed node to a normal node and replans a route to provide a service for a user. The method has the advantages that the migration is carried out only when the actual fault occurs, so that the normal node is not mistaken for the fault node to generate unnecessary migration, but the acceptance rate of the service chain is reduced, the user experience is influenced, and the profit of the service provider is reduced. As can be seen from fig. 12, in the statistical failure time, the SFC acceptance rate of the prediction-based method is always higher than that of the failure-trigger-based method, because the prediction method predicts the physical node to be failed and migrates the VNF instance on the failed node to the normal node in advance. When a fault actually occurs, the traffic can still provide service for the user only by passing through the target migration node, and the SFC acceptance rate is not 100% because the prediction model may have a situation that the fault node cannot be identified. After verifying the advantages of the failure prediction in terms of service migration, the performance of the proposed DQN-based VNF instance migration algorithm is verified next. There are two algorithms participating in the comparison, one is a failure node global migration algorithm OMA, which migrates all VNF instances on a failure node to the same target node. The other is the greedy algorithm GMA, which divides VNF instance migration into two steps, since node cost and link cost affect each other. The node with the lowest deployment cost is selected first and then links are greedy selected so that the link cost is also minimized. Fig. 13 shows a comparison of the migration costs of the three algorithms, and the migration cost of the algorithm VNFMA-DQN proposed in this embodiment is always smaller than that of the comparative algorithms OMA and GMA as the number of service chains increases, and the cost gap gradually increases. Taking the number of service chains as 150 as an example, the cost of the VNFMA-DQN algorithm is 3325, OMA and GMA are 4036 and 4211, respectively, and the VNFMA-DQN algorithm saves about 21.3% and 26.6% of the cost compared with OMA and GMA, respectively.

The physical node fault prediction model is established based on the BP neural network, the prediction precision is higher compared with methods such as an SVM (support vector machine), and the comprehensive evaluation of the design model can be realized by the provided misclassification cost evaluation index;

the method and the system simultaneously consider the node cost and the link cost to plan a VNF instance migration optimization model, realize global minimization of the cost, design a DQN-based solving algorithm facing the cost and the priority, ensure the quality of service of high-priority services, and realize intelligent migration of the VNF instance on a fault node.

Fig. 14 is a schematic structural diagram of a failure prediction-based VNF migration apparatus according to an embodiment of the present invention, as shown in fig. 14, the apparatus includes: a prediction module 201 and a migration module 202, wherein:

the prediction module 201 is configured to predict a fault condition of a physical node in an edge network at a future time based on a physical node fault prediction model of a BP neural network;

a migration module 202, configured to construct a VNF instance migration optimization model based on the prediction result determined by the failure prediction model, the node cost, and the link cost, and migrate the VNF instance on the failed node to the normal node based on the VNF instance migration optimization model.

The failure prediction-based VNF migration apparatus provided in the embodiment of the present invention may be specifically configured to execute the failure prediction-based VNF migration method in the foregoing embodiment, and the technical principle and the beneficial effect thereof are similar, and reference may be specifically made to the foregoing embodiment, and details are not described here.

Based on the same inventive concept, an embodiment of the present invention provides an electronic device, which specifically includes the following contents with reference to fig. 15: a processor 301, a communication interface 303, a memory 302, and a communication bus 304;

the processor 301, the communication interface 303 and the memory 302 complete mutual communication through the communication bus 304; the communication interface 303 is used for realizing information transmission between related devices such as modeling software, an intelligent manufacturing equipment module library and the like; the processor 301 is used for calling the computer program in the memory 302, and the processor executes the computer program to implement the method provided by the above method embodiments, for example, the processor executes the computer program to implement the following steps: predicting the fault condition of a physical node in an edge network at a future moment based on a physical node fault prediction model of a BP neural network; and constructing a VNF instance migration optimization model based on the prediction result, the node cost and the link cost determined by the fault prediction model, and migrating the VNF instance on the fault node to a normal node based on the VNF instance migration optimization model.

Based on the same inventive concept, yet another embodiment of the present invention further provides a non-transitory computer-readable storage medium, on which a computer program is stored, which, when being executed by a processor, is implemented to perform the methods provided by the above method embodiments, for example, predicting a fault condition of a physical node in an edge network at a future time based on a physical node fault prediction model of a BP neural network; and constructing a VNF instance migration optimization model based on the prediction result, the node cost and the link cost determined by the fault prediction model, and migrating the VNF instance on the fault node to a normal node based on the VNF instance migration optimization model.

The above-described embodiments of the apparatus are merely illustrative, and the units described as separate parts may or may not be physically separate, and the parts displayed as units may or may not be physical units, may be located in one place, or may be distributed on a plurality of network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of the present embodiment. One of ordinary skill in the art can understand and implement it without inventive effort.

Through the above description of the embodiments, those skilled in the art will clearly understand that each embodiment can be implemented by software plus a necessary general hardware platform, and certainly can also be implemented by hardware. With this understanding in mind, the above-described technical solutions may be embodied in the form of a software product, which can be stored in a computer-readable storage medium, such as ROM/RAM, magnetic disk, optical disk, etc., and includes instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute the methods of the various embodiments or some parts of the embodiments.

In addition, in the present invention, terms such as "first" and "second" are used for descriptive purposes only and are not to be construed as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one such feature. In the description of the present invention, "a plurality" means at least two, e.g., two, three, etc., unless specifically limited otherwise.

Moreover, in the present invention, relational terms such as first and second, and the like may be used solely to distinguish one entity or action from another entity or action without necessarily requiring or implying any actual such relationship or order between such entities or actions. Also, the terms "comprises," "comprising," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but may include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising an … …" does not exclude the presence of other identical elements in a process, method, article, or apparatus that comprises the element.

Furthermore, in the description herein, reference to the description of the term "one embodiment," "some embodiments," "an example," "a specific example," or "some examples," etc., means that a particular feature, structure, material, or characteristic described in connection with the embodiment or example is included in at least one embodiment or example of the invention. In this specification, the schematic representations of the terms used above are not necessarily intended to refer to the same embodiment or example. Furthermore, the particular features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples. Furthermore, various embodiments or examples and features of different embodiments or examples described in this specification can be combined and combined by one skilled in the art without contradiction.

Finally, it should be noted that: the above examples are only intended to illustrate the technical solution of the present invention, but not to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, it will be understood by those of ordinary skill in the art that: the technical solutions described in the foregoing embodiments may still be modified, or some technical features may be equivalently replaced; and such modifications or substitutions do not depart from the spirit and scope of the corresponding technical solutions of the embodiments of the present invention.

Claims

1. A failure prediction based VNF migration method is characterized by comprising the following steps:

2. The failure prediction-based VNF migration method according to claim 1, wherein migrating the VNF instance on the failed node to the normal node based on the VNF instance migration optimization model specifically includes:

3. The failure prediction-based VNF migration method of claim 1, wherein the BP neural network-based physical node failure prediction model is: and taking SMART characteristic data of the physical node hard disk as input data and a fault prediction result corresponding to the SMART characteristic data of the physical node hard disk as output data, and training the output data based on a deep learning algorithm to obtain the model.

4. The failure prediction based VNF migration method of claim 3, further comprising:

Cost_mis＝Cost₁*Num_FP+Cost₂*Num_FN

among them, Cost_misDenotes the misclassification cost, Num_FPNumber of samples, Num, predicted as the fault is true normal_FNIndicating the number of samples predicted to be normal and truly faulty, Cost₁Represents the loss caused by misclassification of a normal hard disk as a failed hard disk, Cost₂Indicating the loss caused by misclassification of a failed hard disk as a normal hard disk.

5. A VNF migration apparatus based on failure prediction, comprising:

6. The failure prediction-based VNF migration apparatus according to claim 5, wherein the migration module, when executing the migration of the VNF instance on the failed node to the normal node based on the VNF instance migration optimization model, is specifically configured to:

7. The failure prediction based VNF migration apparatus of claim 5, wherein the BP neural network based physical node failure prediction model in the prediction module is: and taking SMART characteristic data of the physical node hard disk as input data and a fault prediction result corresponding to the SMART characteristic data of the physical node hard disk as output data, and training the output data based on a deep learning algorithm to obtain the model.

8. The failure prediction based VNF migration apparatus of claim 7, wherein the prediction module is further configured to:

Cost_mis＝Cost₁*Num_FP+Cost₂*Num_FN

9. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processor implements the failure prediction based VNF migration method of any one of claims 1 to 4 when executing the program.

10. A non-transitory computer readable storage medium having stored thereon a computer program, which when executed by a processor implements the failure prediction based VNF migration method according to any one of claims 1 to 4.