Service function chain migration method based on fault prediction

Abstract

The present invention relates to a service function chain migration method based on fault prediction. First, a long short-term memory network is used to predict hardware faults. Simultaneously, a simulated annealing particle swarm algorithm is used to optimize the number of neurons and the time window length of each layer of LSTM, and a fault prediction method based on an improved LSTM, namely, a PMILSTM method, is proposed. Secondly, chaos, reverse learning, dynamic weight factors, and mutation operations are introduced into a sparrow search algorithm, and an improved sparrow search algorithm, namely, an ISSA, is proposed. Finally, the improved sparrow search algorithm is applied to an SFC migration method, and a service function chain migration method based on an ISSA, namely, an MMISSA method, is proposed. The MMISSA method simultaneously considers the migration server nodes of all VNFs on the faulty node, and the fitness function simultaneously considers the migration overhead and then the migration time. Therefore, the MMISSA method effectively improves the migration success rate and reduces the migration overhead and migration time.

Description

Technical Field

[0001] The present invention belongs to the field of network function virtualization, and relates to a service function chain migration method based on fault prediction, including an improved long short-term memory network, an improved sparrow search algorithm, and a service function chain migration method based on the improved sparrow search algorithm. Background Art

[0002] The literature "J.Shen, J.Wan, S.Lim, L.Yu.Random-forest-based failure prediction for hard disk drives" proposes a random forest failure prediction method, namely the RFBP method, to address the problem that hardware is prone to failure. The literature "A.Santo, A.Galli, M.Gravina, V.Moscato, G.Sperli.Deep learning for HDD health assessment: an application based on LSTM" proposes a failure prediction method based on long short-term memory network, namely the LSTMP method, to address the problem that hardware is prone to failure. Literature 3 "J.Xue, B.Shen.A novel swarm intelligence optimization approach: sparrow search algorithm" proposes a sparrow search algorithm (SSA). Compared with other swarm intelligence optimization algorithms, the sparrow optimization algorithm has stronger global optimization capabilities and faster search speed. However, in the later stages of iteration, as the population size decreases, the sparrow search algorithm is also prone to fall into local optimality. The paper "S.Zhou, H.Xie, C.Zhang, Y.Hua, W.Zhang, Q.Chen, G.Gu, X.Sui. Wavefront-shaping focusing based on a modified sparrow search algorithm" introduces crossover mutation operation into the sparrow search algorithm, namely MSSA, thereby improving the global optimization ability of the sparrow search algorithm. The paper "Y.Zhu, N.Yousefi. Optimal parameter identification of PEMFCstacks using adaptive sparrow search algorithm" introduces adaptive learning factors into the position update functions of the discoverer, follower, and sentinel, namely ASSA, thereby improving the global optimization ability of the sparrow search algorithm. The paper "D.Zhao, G.Sun, D.Liao, S.Xu, V.Chang. Mobile-aware service function chain migration in cloud-fog computing" proposes a two-stage SFC method, namely the SFCTSM method, for the migration problem of the service function chain (SFC).The SFCTSM method first migrates the minimum number of virtual network functions (VNFs) to quickly restore services. Then other VNFs are migrated to reduce the underlying resource overhead. The document "B.Yi, X.Wang, M.Huang, K.Li. Design and implementation of network-aware VNF migration mechanism" proposes a node-aware VNF migration method, namely NAVMM, to address the migration problem of the service function chain. The NAVMM method prioritizes migrating VNFs with the lowest migration overhead and repeats this process until there are no overloaded nodes. However, the following problems still exist:

[0003] (1) In order to better migrate and protect the corresponding services for hardware failures, the performance of the fault prediction method still needs to be further improved.

[0004] (2) The Sparrow Algorithm is prone to fall into local optimality in the later stages of iteration. In order to better apply the Sparrow Algorithm to practical problems, the global optimization capability of the Sparrow Algorithm still needs to be further improved.

[0005] (3) For server nodes that are about to experience hardware failure, all VNFs deployed on this server node need to be migrated. These VNFs may belong to different SFCs, which will make the SFC migration problem more complicated. In order to use the underlying resources more reasonably and improve the migration performance of the corresponding SFCs, the migration nodes of these VNFs need to be considered simultaneously. The existing SFC migration method only considers the migration node of one VNF at a time. For hardware failures, the existing migration method will reduce the migration success rate and resource utilization. Summary of the invention

[0006] Technical issues to be solved

[0007] In order to effectively predict hardware failures and migrate service function chains in advance, the present invention provides a service function chain migration method based on failure prediction.

[0008] Technical Solution

[0009] A service function chain migration method based on fault prediction, characterized by the following steps:

[0010] Step 1: Use LSTM neural network to predict hardware failures, and use SAPSO algorithm to optimize the number of neurons and time window length of each LSTM layer, namely PMILSM method; determine whether the VNF on the server node needs to be migrated based on the prediction results, that is, if the service node is about to fail, all VNFs carried by the server node need to be migrated;

[0011] Step 2: Obtain a set of candidate migration server nodes according to CPU resource constraints, storage resource constraints, and carrying capacity constraints; use an improved sparrow search algorithm to simultaneously select the optimal migration server node from the set of candidate migration server nodes for the VNF to be migrated.

[0012] The PMILSM method in step 1 is specifically as follows:

[0013] Input: Initialize the number of iterations Ni and learning factors α1, α2, initialize the time window length and the number of LSTM neurons in each layer as particles, SMART attribute value

[0014] Output: Fault prediction results

[0015] (1) Train the LSTM model and calculate the initial fitness value;

[0016] (2) Use the SAPSO algorithm to update the global optimal position of the particle;

[0017] (3) Obtain the optimal time window length and number of LSTM neurons in each layer;

[0018] (4) Update the LSTM model using the parameters obtained in step (3);

[0019] (5) Use training data to train the LSTM model and threshold value;

[0020] (6) Input the SAMRT attribute value and predict the hardware drive failure through the trained LSTM model.

[0021] The improved sparrow search algorithm in step 2:

[0022] Step 21: Initialize parameters and set the number of VNFs carried by the failed server node as the dimension of the search space;

[0023] Step 22: Use the position update functions of producers, followers and sentinels to update the positions of producers, followers and sentinels, and calculate the fitness values;

[0024] Step 23: Determine whether the iteration stop condition is met. If not, continue the iteration; if so, output the optimal position, which is the migration server node of these VNFs.

[0025] A computer system, characterized in that it comprises: one or more processors, and a computer-readable storage medium for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the above-mentioned method.

[0026] A computer-readable storage medium, characterized in that it stores computer-executable instructions, which are used to implement the above method when executed

[0027] Beneficial Effects

[0028] The present invention provides a service function chain migration method based on fault prediction. First, a long short-term memory network (LSTM) is used to predict hardware faults. Simultaneously, a simulated annealing particle swarm algorithm is used to optimize the number of neurons and the time window length of each layer of LSTM, and a fault prediction method based on an improved LSTM, namely, the PMILSTM method, is proposed. Secondly, chaos, reverse learning, dynamic weight factors, and mutation operations are introduced into the sparrow search algorithm (SSA), and an improved sparrow search algorithm, namely, ISSA, is proposed. Finally, the improved sparrow search algorithm is applied to the SFC migration method, and a service function chain migration method based on ISSA, namely, the MMISSA method, is proposed. The MMISSA method simultaneously considers the migration server nodes of all VNFs on the faulty node, and the fitness function simultaneously considers the migration overhead and then the migration time. Therefore, the MMISSA method effectively improves the migration success rate and reduces the migration overhead and migration time.

[0029] 1. First, the simulated annealing particle swarm algorithm is used to optimize the number of neurons in each layer of LSTM and the length of the time window. The improved LSTM network is used to predict hardware failures, which improves the accuracy of prediction.

[0030] 2. Secondly, chaos, reverse learning, dynamic weight factors and mutation operations are introduced into SSA to improve the global optimization ability of SSA.

[0031] 3. Finally, ISSA is applied to SFC migration, and the migration server nodes of all VNFs deployed on the failed node are considered, which effectively improves the migration success rate and reduces the migration overhead and migration time. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] The drawings are only for the purpose of illustrating particular embodiments and are not to be considered limiting of the present invention. Like reference symbols denote like components throughout the drawings.

[0033] Figure 1 It is a schematic diagram of SFC migration provided by the present invention.

[0034] Figure 2 It is a structural diagram of the LSTM unit provided in the present invention.

[0035] Figure 3 It is a flow chart of the SAPSO algorithm provided by the present invention.

[0036] Figure 4 It is a flow chart of the sparrow search algorithm provided by the present invention.

[0037] Figure 5 It is a flow chart of the improved sparrow search algorithm in the method of the present invention.

[0038] Figure 6 This is a comparison result diagram of fault detection rates in the method of the present invention.

[0039] Figure 7 This is a comparison result diagram of fault false alarm rate in the method of the present invention.

[0040] Figure 8 Graphs showing convergence comparison results in the method of the present invention: (a) F1 test function; (b) F2 test function; (c) F3 test function.

[0041] Fig. 9 The comparison results of the migration success rate in the method of the present invention are shown in FIG. (a) λ p =0.05; (b)λ p =0.1; (c)λ p =0.15.

[0042] Fig.10 The average migration cost comparison result diagram in the method of the present invention is as follows: (a) λ p =0.05; (b)λ p =0.1; (c)λ p =0.15.

[0043] Fig.11 This is a comparison result diagram of average migration time in the method of the present invention;

[0044] Fig.12 It is a flow chart of the method of the present invention. DETAILED DESCRIPTION

[0045] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0046] The present invention provides a service function chain migration method based on ISSA (i.e., MMISSA method), which includes the following parts:

[0047] Part I: Establishing the network model for SFC migration.

[0048] Part 2: Design of fault prediction method based on improved LSTM (i.e. PMILSTM).

[0049] Part 3: Description of the sparrow search algorithm.

[0050] Part 4: Design of improved sparrow search algorithm (ie ISSA).

[0051] Part 5: Design of service function chain migration method based on ISSA (ie MMISSA).

[0052] Part 6: Verify the effectiveness of the proposed method.

[0053] The following examples refer to Figures 1 to 11 The specific implementation process is:

[0054] 1. Establish a network model for SFC migration

[0055] Physical network: The present invention selects multiple infrastructure providers to provide hardware services as the application scenario, and the carrying capacity of general server nodes may vary. The physical network can be represented by a weighted undirected graph G s =(V s ,E s ) indicates that V s =V s,s ∪V s,f , V s,s = {v s,i |i=1,2,...,|V s,s |}.

[0056] SFC request: An SFC request can be represented by a weighted undirected G g = {N g ,L g ,S g ,D g}, where N g ={f j |j=1,2,...,|N g |}. The service flow flows into a switch node, passes through the virtual network functions in a given order, and then flows out from another switch node.

[0057] SFC deployment: When an SFC request arrives, the service provider instantiates each virtual network function of the service function chain in a predetermined order. s,i Can carry VNFf j ,x(v s,i ,f j )=1, otherwise, x(v s,i ,f j )=0.

[0058] SFC migration: When the server node v s,i When a failure occurs, all VNFs it carries need to be migrated. During the migration process, for each VNF, the migration server node must meet its CPU resource and storage resource requirements. At the same time, the underlying link must also meet the bandwidth resource requirements of the corresponding virtual link.

[0059] like Figure 1 As shown in Figure 1. VNFs 1, 2, and 3 in SFC1 are deployed on server nodes 2, 3, and 7, respectively. The black dashed line represents the real route of SFC1. Assuming that server node 3 fails, VNF2 of SFC1 needs to be migrated to other server nodes. VNF2 is migrated to server node 8, and the available CPU and storage resources of service node 8 meet the resource requirements of VNF2. The gray dashed line represents the reconstructed route of SFC1. The underlying link e 2,8 and e 8,7 The available bandwidth of virtual link l 1,2 and l 2,3 The commonly used symbols and their meanings in the present invention are shown in Table 1.

[0060] Table 1 Common symbols and their meanings

[0061]

[0062]

[0063] 2. Design of fault prediction method based on improved LSTM

[0064] LSTM neural network deletes / adds information through forget gate, input gate and output gate. Its repeating module has four network layers, while the standard recurrent neural network has only one network layer. The LSTM unit structure is as follows Figure 2 As shown, the first step is to decide from the unit state M t-1 Which information should be discarded or retained in the process? This is done through the forget gate. As shown in formula (1), the output of the forget gate is r t The second step is to decide what information to store in the cell state, which is done through the input gate, as shown in formula (2). The output of the input gate is i t . Candidate unit information Update according to formula (3). The third step is to update the unit state M t-1 and candidate unit information To update the unit state M t , as shown in formula (4). The next step is to decide which state features to output, which is done through the output gate, as shown in formula (5). The output of the output gate is O t The final output result is composed of output information Ot and the cell state M t Determine, as shown in formula (6).

[0065] Symbol W rk , W ik , W mk and W ok Indicates the previous output k t-1 The weighting factor of . Symbol W ry , W iy , W my and W oy Indicates input y t The weighting factor of . Symbol b r , b i , b o and b m Represents the forget gate, input gate, output gate, and bias of the candidate unit. The symbols σ and tanh represent activation functions, and the symbol * represents dot product.

[0066] r t =σ(W rk k t-1 +W ry y t -b r ) (1)

[0067] i t =σ(W ik k t-1 +W iy y t +b i ) (2)

[0068]

[0069]

[0070] O t =σ(W ok k t-1 +W oy y t +b o ) (5)

[0071] k t =O t *tanh(M t ) (6)

[0072] Hardware usually deteriorates gradually rather than suddenly. LSTM neural networks have better performance in memorizing long-term information. Therefore, the present invention uses LSTM neural networks to predict hardware failures. For LSTM neural networks, the time window length and the number of neurons in each layer of LSTM are not fixed. The time window length and the number of neurons in each layer of LSTM have a great impact on the prediction accuracy. The simulated annealing particle swarm (SAPSO) algorithm has the advantages of high precision, simple implementation and short analysis time. Therefore, the present invention uses the SAPSO algorithm to optimize the number of neurons in each layer of LSTM and the time window length. The flowchart of the SAPSO algorithm is as follows: Figure 3 As shown in the figure, for the optimization problem, one particle represents one solution. In each iteration, the velocity vector and position vector of the particle are updated according to formula (7) and (8) respectively. In order to jump out of the local optimal solution, the SAPSO algorithm performs simulated annealing according to formula (9). The probability of accepting a worse solution is calculated according to formula (10). Finally, the global optimal position of all particles is updated according to the Metropolis criterion.

[0073] ve i (j+1)=ωve i (j)+α1β1(pbest(j)-x i (j))+α2β2(gbest(j)-x i (j)) (7)

[0074] x i (j+1)=x i (j)+ve i (j+1) (8)

[0075] The symbol ve i (j) and x i (j) represents the position vector and velocity vector of particle i in the jth iteration respectively. Symbols α1 and α2 represent the local learning factor and the global learning factor respectively. Symbols β1 and β2 are random numbers between 0 and 1. In each iteration, symbol pbest represents the local optimal position of the particle, and symbol gbest represents the global optimal position of all particles. Symbol ω represents the weight factor.

[0076]

[0077]

[0078] The symbol T(j) represents the temperature at the jth iteration, and the symbol ε represents the cooling coefficient. s,g (j) represents the fitness value of the global optimal position at the jth iteration, symbol F s,p (j) represents the fitness value of the optimal position after the jth iteration.

[0079] In order to improve the accuracy of predicting hardware failures, the present invention adopts the SAPSO algorithm to optimize the number of neurons in each layer of LSTM and the length of the time window, and proposes a fault prediction method based on an improved LSTM, namely the PMILSTM method. For the convenience of analysis, the present invention mainly considers hardware driver failures. SMART technology has been verified and has good performance in hardware driver failure prediction. Therefore, the present invention uses SMART attributes to predict hardware driver failures. The main steps of the PMILSM method are shown in Algorithm 1, and the number of LSTM neurons in each layer and the length of the time window are used as particles in the search space. The mean square error between the predicted SMART attribute value and the true SMART attribute value is used as the fitness value.

[0080]

[0081] 3. Sparrow search algorithm

[0082] The sparrow search algorithm (SSA) divides the population into producers, followers, and guards. Producers have higher fitness and mainly provide foraging areas and directions for sparrows. Followers forage in the area provided by producers, and some followers even monitor producers to compete for food. Guards are responsible for detecting predators and early warning. Once the guards detect predators, they will warn other sparrows.

[0083] The flowchart of the sparrow algorithm is as follows Figure 4 As shown, initialize the position of the sparrow, using the set s i =(s i,1 ,s i,2 ,...,s i,D ) indicates that sparrows i The fitness value is expressed as F(s i ). The symbols NN, NP, NS, and NC represent the number of populations, producers, followers, and sentinels, respectively. The symbol q max represents the maximum number of iterations, and the symbol D represents the search space. The upper and lower limits of the search space are respectively represented as ub = {ub j |j=1,2,...,D} and lb={lb j |j=1,2,...,D}. The producer's position update function is shown in formula (11).

[0084]

[0085] The symbols AT and ST represent the alarm value and safety threshold, respectively. The symbol q represents the current iteration number, and the symbol μ is a random number. The symbol Q is a random number that follows a normal distribution. The symbol I represents a 1×D matrix, in which each element is 1. When AT<ST, it means that the producer has not received a danger warning and will expand its foraging range; otherwise, the producer will move to a safer area to forage according to the normal distribution.

[0086] The position update function of the follower is shown in formula (12).

[0087]

[0088] The symbol G worst and P best Represent the worst position of the sparrow and the current best position of the producer. Symbol A + represents a 1×D matrix where each element is either 1 or -1 and A+=A T ·(A·A T ) -1 When i>NN / 2, follower i will fly to other areas to forage; otherwise, follower i will forage near the producer's optimal position.

[0089] The position update function of the sentinel is shown in formula (13).

[0090]

[0091] The symbol G best represents the current best position of the sparrow, and the symbol γ is a random number that follows a normal distribution. The symbol λ is a random number that represents the step size control factor, and the symbol is a constant approaching 0.

[0092] 4. Improved sparrow search algorithm

[0093] The sparrow search algorithm is prone to fall into local optimum in the later iteration period due to the decrease in population size. Therefore, the present invention proposes an improved sparrow search algorithm, namely the ISSA algorithm. In order to improve the global optimization capability of SSA, ISSA introduces chaos, reverse learning, dynamic weighting factors and mutation operations into SSA. The algorithm flow chart of ISSA is shown in the figure. Figure 5 In order to improve the global search capability at the beginning of iteration, ISSA uses the logistic map in the chaotic map to initialize the position of the population. The logistic map is shown in formula (14).

[0094] s(d+1)=τs(d)(1-s(d))s(d)∈(0,1),τ∈(0,4] (14)

[0095] The symbol τ represents the control parameter

[0096] Initialize the population positions as follows:

[0097] Using formula (15), the initial position s of the sparrow is ij (i=1,2,...,NN;j=1,2,...,D)normalized to s ij (0)(i=1,2,...,NN;j=1,2,...,D)to satisfy the chaotic conditions.

[0098] s ij (0)=(s ij -lb j ) / (ub j -lb j ) (15)

[0099] Using logistic mapping to get s ij (d+1), as shown in formula (16).

[0100] s ij (d+1)=τs ij (d)(1-s ij (d)) (16)

[0101] According to formula (17), s ij (d+1) is transformed from the original domain [0,1] to the new domain s ij (i=1,2,...,NN; j=1,2,...,D).

[0102] s ij =lb j +s ij (d+1)(ub j -lb j ) (17)

[0103] In order to further improve the global search capability of the initial population, ISSA uses reverse learning to obtain a new population, as shown in formula (28). The sparrows of the two populations are sorted according to their fitness values, and NN sparrows with high fitness values ​​are selected as the initial population.

[0104] s' i,j =ub j +lb j -s i,j (18)

[0105] In each iteration, in order to avoid falling into the local optimum, ISSA uses reverse learning to optimize the current best position and the global best position. ISSA obtains the reverse position of the current producer's best position and the sparrow's global best position through formula (19), and then obtains the new producer's current best position and the sparrow's global best position by comparing the fitness values, as shown in formulas (20) and (21).

[0106]

[0107]

[0108]

[0109] The producer moves to the global optimal position from the beginning of the iteration, resulting in insufficient search range and easy to fall into the local optimum. Therefore, ISSA introduces the global optimal position of the last sparrow into the producer's position update function. In this way, the producer's position is not only affected by the producer's position in the last iteration, but also by the last global optimal position, which can avoid falling into the local optimum. In addition, ISSA adds a dynamic weighting factor ω s It is introduced into the producer's position update function. The dynamic weighting factor is larger at the beginning of the iteration, which can perform better global search. In the later stage of the iteration, the weighting factor is adaptively reduced to better perform local search and improve convergence speed. The calculation method of the weighting factor is shown in formula (22).

[0110]

[0111] The improved producer position update function is shown in formula (23).

[0112]

[0113] Among them, the symbol Indicates the global optimal position of the previous generation of sparrows.

[0114] In the later stages of iteration, SSA is prone to fall into a local optimum due to the decrease in population. Therefore, ISSA introduces a mutation operation to increase the diversity of the population. In order not to affect the convergence of the algorithm, only one individual is mutated in each iteration, as shown in formula (24).

[0115] s a =lb+(ub-lb)*rand(1,D) (24)

[0116] The symbol s a represents a random individual in each iteration, and the symbol * represents dot product.

[0117] 5. Design of service function chain migration method based on ISSA

[0118] In order to better solve the service function chain migration problem caused by hardware failure. The present invention proposes a service function chain migration method based on ISSA, namely the MMISSA method. The MMISSA method migrates the SFC according to the prediction results of the PMILSTM method. If the prediction results show that a server node will fail, all VNFs deployed on the server node need to be migrated. The MMISSA method uses an improved sparrow search algorithm to migrate these VNFs and corresponding virtual links. The dimension D of the search space is the same as the number of VNFs carried by the server node that is about to fail. Therefore, the MMISSA method can simultaneously find migration server nodes for all VNFs deployed on the server node that is about to fail. In order to reduce the average migration overhead and the average migration time, the fitness function simultaneously considers the average migration overhead and the average migration time, as shown in formula (25).

[0119]

[0120] The symbol C av and They represent the average migration cost and average migration time, symbol Q C and Q T They represent the influencing factors of average migration cost and average migration time respectively.

[0121] Flowchart as Fig.12 As shown, first use PMILSTM to predict the failure of each server node. If the prediction result shows that a server node is about to fail, determine whether this server node carries VNF. If it carries VNF, these VNFs need to be migrated. Obtain a set of candidate migration server nodes based on CPU resource constraints, storage resource constraints, and carrying capacity constraints. Use the ISSA algorithm to simultaneously select the optimal migration server node for these VNFs in the set of candidate migration server nodes. Initialize the parameters and set the number of VNFs carried by this server node that is about to fail as the dimension of the search space. Use the position update function of producers, followers, and sentinels to update the positions of producers, followers, and sentinels, and calculate the fitness value. Determine whether the iteration stop condition is met. If not, continue the iteration. If it is met, output the optimal position, which is the migration server node for these VNFs.

[0122] The main steps of the MMISSA method are shown in Algorithm 2. The MMISSA method first determines whether the VNF on the server node needs to be migrated based on the prediction results of the PMILSTM method. If the service node is about to fail, all VNFs carried by the server node need to be migrated. For each VNF that needs to be migrated, the MMISSA method selects a set of candidate server nodes based on CPU resource constraints, storage resource constraints, and carrying capacity constraints, and uses these sets as the search space of the sparrow (Line 3–10). Symbol V m,i VNFf i The set of candidate migration server nodes. The MMISSA method simultaneously searches for migration server nodes for all VNFs deployed on the server nodes that are about to fail, so that the overall migration performance of the corresponding SFC is better. The use of logistic mapping and reverse learning improves the global search capability at the beginning of the iteration (Line13–14). The positions of the producer, follower, and sentinel are updated according to formulas (23), (12), and (13), respectively (Line17–18). Reverse learning is used to optimize the current best position of the producer and the global best position of the sparrow (Line20). Each iteration randomly performs a mutation operation on an individual (Line21). If the termination condition is met, the iteration stops. Output the global optimal position, which is the migration result of the corresponding SFC.

[0123]

[0124]

[0125] 6. Performance evaluation and analysis

[0126] The experimental simulation of the present invention is divided into three parts. The first part verifies the performance of the PMILSTM method, the second part verifies the performance of the ISSA algorithm, and the third part verifies the performance of the MMISSA method. The present invention uses MATLAB to perform experimental simulation, each simulation experiment is performed 100 times, and the average result is taken as the final result.

[0127] 6.1 Fault prediction performance verification and discussion

[0128] The present invention mainly considers hardware drive failures and uses real data sets from data centers for experiments. These data have been used in the literature "C.Xu, G.Wang, X.Liu, D.Guo, T.Liu. Health status assessment and failure prediction for hard drives with recurrent neural networks" and "B.Zhu, G.Wang, X.Liu, D.Hu. Proactive drive failure prediction for large scale storage systems". 70% of the data are used as training data and 30% of the data are used as test data. As shown in Table 2, the data set retains 10 attributes from the 23 attributes in SMART. Each sample has 12 features, including 10 attributes and the original data of the reallocatedsectors count and current pending sector count attributes. The description of the three prediction methods is shown in Table 3.

[0129] Table 2 Attributes selected from SMART

[0130]

[0131] Table 3 Description of three prediction methods

[0132]

[0133]

[0134] Figure 6 and Figure 7 The fault detection rate and fault false alarm rate of the three methods are shown. The PMILSTM method has the highest fault detection rate and the lowest fault false alarm rate. Hardware drivers usually deteriorate gradually rather than suddenly. LSTM neural networks have better performance in memorizing long-term information. Therefore, the performance of the PMILSTM method is better than that of the RFBP method. The PMILSTM method uses the SAPSO algorithm to optimize the number of LSTM neurons in each layer and the length of the time window. Therefore, its performance is better than that of the LSTMP method.

[0135] 6.2 Optimization performance verification and discussion

[0136] The superiority of the ISSA algorithm is verified by simulation experiments using three test functions. The descriptions of the three test functions are shown in Table 4. The descriptions of the three optimization algorithms are shown in Table 5. The maximum number of iterations is set to 500, the population size is set to 50, and the safety threshold is set to 0.8.

[0137] Table 4 Three test functions

[0138]

[0139] Table 5 Description of three optimization algorithms

[0140]

[0141] Figure 8 The convergence results of the three optimization algorithms are shown. The ISSA algorithm has the strongest optimization ability and the fastest convergence speed. The ISSA algorithm uses logistic mapping and reverse learning to initialize the population, so the global optimization ability in the initial stage is improved. In each iteration, in order to avoid falling into the local optimum, the ISSA algorithm uses reverse learning to optimize the current optimal position of the producer and the global optimal position of the sparrow. In each iteration, an individual is randomly mutated to increase the diversity of the population. In addition, introducing the dynamic weighting factor and the global optimal position of the previous generation into the producer's position update function can effectively avoid falling into the local optimum. Therefore, the performance of the ISSA algorithm is better than that of the MSSA algorithm and the ASSA algorithm.

[0142] 6.3 Migration Performance Verification and Discussion

[0143] The underlying network has 200 server nodes. The CPU and storage resources of the server nodes follow a uniform distribution of [30–70]. The SMART attributes of the server nodes are the same as those in Experiment 1. The bandwidth resources of the underlying links follow a uniform distribution of [30–50]. The number of VNFs in the SFC follows a random distribution of [5–8]. The CPU and storage resource requirements of the VNFs follow a uniform distribution of [15–30]. The bandwidth requirements of the virtual links follow a uniform distribution of [10–20]. The arrival rate of SFC requests follows a Poisson distribution. The description of the three migration methods is shown in Table 6.

[0144] Table 6 Description of three migration methods

[0145]

[0146] Fig. 9 The figure shows the migration success rates of the three migration methods under different SFC request arrival rates. Fig.10As can be seen from the figure, the MMISSA method has the highest migration success rate. The MMISSA method uses the ISSA algorithm to migrate VNFs. The ISSA algorithm has a higher global optimization capability, so the MMISSA method has a higher global optimization capability. In addition, the MMISSA method simultaneously searches for migration server nodes for all VNFs on the server node that is about to fail, which can effectively improve the overall migration performance of the corresponding SFC. Therefore, the MMISSA method has the highest migration success rate among the three methods.

[0147] Fig.10 The average migration overhead of the three migration methods under different SFC request arrival rates is shown. Fig.10 As can be seen from the figure, the MMISSA method has the lowest average migration overhead. The MMISSA method uses the ISSA algorithm to migrate VNFs and simultaneously searches for migration server nodes for all VNFs hosted by the failing server node. In addition, the fitness function includes the average migration overhead. Therefore, the MMISSA method has the lowest average migration overhead among the three methods.

[0148] Fig.11 The average migration time of the three migration methods under different SFC request arrival rates is shown. Fig.11 As can be seen from the figure, the MMISSA method has the lowest average migration time. The MMISSA method simultaneously searches for migration server nodes for all VNFs hosted by the failing server node. In addition, the fitness function includes the average migration time. Therefore, the MMISSA method has the lowest average migration time among the three methods.

[0149] The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should be included in the protection scope of the present invention.

Claims

1. A service function chain migration method based on fault prediction, characterized in that Here are the steps: Step 1: Use LSTM neural network to predict hardware failures, and use SAPSO algorithm to optimize the number of neurons and time window length of each LSTM layer, namely PMILSM method; determine whether the VNF on the server node needs to be migrated based on the prediction results. If the server node is about to fail, all VNFs carried by the server node need to be migrated; The PMILSM method is specifically as follows: Input: Initialize the number of iterations Ni and learning factors α1, α2, initialize the time window length and the number of LSTM neurons in each layer as particles; the input value also includes the SMART attribute value; Output: fault prediction results; (1) Train the LSTM model and calculate the initial fitness value; (2) Use the SAPSO algorithm to update the global optimal position of the particle; (3) Obtain the optimal time window length and number of LSTM neurons in each layer; (4) Update the LSTM model using the parameters obtained in step (3); (5) Use training data to train the LSTM model and threshold value; (6) Input the SAMRT attribute value and predict the hardware drive failure through the trained LSTM model; Step 2: Obtain a set of candidate migration server nodes according to CPU resource constraints, storage resource constraints, and carrying capacity constraints; use an improved sparrow search algorithm to simultaneously select the best migration server node from the set of candidate migration server nodes for the VNF to be migrated; The improved sparrow search algorithm: Step 21: Initialize parameters and set the number of VNFs carried by the failed server node as the dimension of the search space; Step 22: Use the position update functions of producers, followers and sentinels to update the positions of producers, followers and sentinels, and calculate the fitness values; Step 23: Determine whether the iteration stop condition is met. If not, continue the iteration; if so, output the optimal position, which is the migration server node of these VNFs.

2. A computer system, characterized in that include: One or more processors, and a computer-readable storage medium for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors are enabled to implement the method of claim 1.

3. A computer-readable storage medium, characterized in that Computer executable instructions are stored, and when the instructions are executed, they are used to implement the method of claim 1.

Images