A remote sensing product production task scheduling method based on a particle swarm algorithm

Abstract

The application provides a remote sensing product production task scheduling method based on a particle swarm algorithm. The method comprises the following steps: step 1, giving task set information and production node set information; step 2, constructing a pre-execution time ETC matrix and a resource allocation CAL matrix of the task; step 3, encoding the particles by using an indirect encoding mode, establishing a particle swarm with a scale of N, and setting a maximum iteration number; step 4, selecting a typical linear decreasing strategy as an inertia weight algorithm; step 5, calculating the fitness value of each particle according to the pre-execution time ETC matrix and the resource allocation CAL matrix; step 6, updating the speed and position of the particle; step 7, updating the global optimal solution; step 8, judging whether convergence or the maximum iteration number is reached, if yes, outputting the global optimal solution at this time, and allocating the task to the corresponding production node for production according to the global optimal solution, and if not, returning to step 5 to continue the next search process.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of task resource scheduling, and particularly relates to a remote sensing product production task scheduling method based on a particle swarm algorithm. BACKGROUND

[0002] With the rapid development of computer technology and remote sensing technology, remote sensing data has increased rapidly, and spectral resolution has been continuously improved. The processing and management of massive remote sensing data have attracted more and more attention.

[0003] At present, cluster management is a relatively effective solution for massive remote sensing data. However, with the increase in the number of tasks, limited computing resources and unstable networks often lead to slow task allocation, load imbalance in heterogeneous systems, and other problems. Therefore, designing a suitable task scheduling method is one of the core problems of research. The task scheduling algorithm in the existing system tends to allocate all tasks to one node, ignoring the CPU and memory of the node and the amount of tasks already on the node. As the amount of tasks increases, the node load becomes unbalanced. In addition, based on the characteristics of orders and tasks in high-performance cluster systems, the acquisition of the ETC matrix is not well solved in the previous scheduling algorithm strategy based on the particle swarm. Secondly, the number of particle swarms and the number of iterations are often fixed values, without considering the size of the task amount. SUMMARY

[0004] In view of the problem that the existing traditional task scheduling method cannot flexibly adjust the scheduling strategy according to the size of the remote sensing data task amount, resulting in low scheduling efficiency, the present application provides a remote sensing product production task scheduling method based on a particle swarm algorithm.

[0005] The remote sensing product production task scheduling method based on the particle swarm algorithm provided by the present application comprises the following steps:

[0006] Step 1: Given the task set information and the production node set information, the task set refers to a set composed of remote sensing product production tasks that need to be allocated, and the production node set information refers to a set composed of production nodes for executing remote sensing product production tasks;

[0007] Step 2: Construct the task pre-execution time ETC matrix and the resource allocation CAL matrix according to the task set and the production node set; wherein, ETC ij represents the element in the ith row and the jth column of the ETC matrix, which is used to represent the task pre-execution time of task i on production node j; CAL ij represents the element in the ith row and the jth column of the CAL matrix, which is used to represent the resources consumed when task i is allocated to production node j for execution;

[0008] Step 3: encode the particles by using indirect encoding method, establish a particle group with a scale of N, construct a task allocation DIS matrix according to the particle position, and set a maximum iteration number; wherein, one particle represents a task allocation result, and the length of the encoding is the number of tasks; the larger the task quantity is, the larger the scale of the particle group and the maximum iteration number are; wherein, DIS ij = 1 indicates that task i is allocated to production node j, and DIS ij = 0 indicates that no task is allocated to production node j;

[0009] Step 4: select a typical linear decreasing strategy as the inertia weight algorithm;

[0010] Step 5: calculate the fitness value of each particle according to the pre-execution time ETC matrix, resource allocation CAL matrix and task allocation DIS matrix according to a set fitness function formula;

[0011] Step 6: update the speed and position of the particle;

[0012] Step 7: update the global optimal solution;

[0013] Step 8: judge whether to converge or reach the set maximum iteration number, if yes, output the global optimal solution at this time, and allocate the tasks to the corresponding production nodes for production according to the global optimal solution; if not, return to step 5 to continue the next search process.

[0014] Further, if the optimal fitness value in the continuous multiple search processes is unchanged, it is considered that the particle swarm algorithm may fall into local optimum, at this time, the Naive Acceptance rule is executed to jump out of the local optimum; the Naive Acceptance rule refers to: when the fitness value of the new global optimal solution is better than the fitness value of the current global optimal solution, the new global optimal solution is accepted; otherwise, a random number p is generated, if p ≥ 0.5, the new global optimal solution is accepted; wherein, the determination method of the new global optimal solution is: in the set iteration range, the suboptimal solution is selected as the new global optimal solution.

[0015] Further, if the optimal fitness value in the continuous multiple search processes is unchanged, it is considered that the particle swarm algorithm may fall into local optimum, at this time, the task exchange rule is executed to jump out of the local optimum; the task exchange rule specifically includes:

[0016] The current global optimal solution is expressed as a task allocation matrix, the task completion time of each production node is determined according to the corresponding relationship between the task allocation matrix and the ETC matrix, the task completion time matrix of all production nodes is constructed, and the maximum task completion time T max and the minimum task completion time T min, and the tasks in the production nodes R max with the minimum task pre-execution time of the tasks Task in the production nodes R min ; wherein R max refers to the production node corresponding to T max , and R min refers to the production node corresponding to T min ; the task completion time of the production node is equal to the sum of the task execution times of all tasks thereon;

[0017] Suppose that the task Task is allocated to the production node R min , if the task completion time of the production node R min is less than T max at this time, the task Task is allocated to the production node R min , and the task allocation matrix and the task completion time matrix are updated; if the task completion time of the production node R min is not less than T max at this time, the second smallest task completion time is further determined from the values of the task completion time matrix, and is taken as the new minimum task completion time T min to re-perform task exchange, if all the minimum task completion times are traversed and task exchange cannot be performed, the second largest task completion time is further determined from the values of the task completion time matrix, and is taken as the new maximum task completion time T max to re-perform task exchange.

[0018] Further, in step 1, set the task set Task = {T1, T2, …, T m}, wherein the information of the task T i is represented as: T i = {TaskID, OrderID, taskType}, i = 1, 2, …, m; wherein TaskID, OrderID, tasktype respectively represent task ID, order ID corresponding to the task, and task type.

[0019] Set the production node set Node = {N1, N2, …, N n}, wherein each production node N j includes node IP and production capacity information, j = 1, 2, …, n.

[0020] Further, the production capacity information includes CPU computing power, RAM memory size, disk storage space, and network bandwidth.

[0021] Further, in the entire task scheduling process, the same task cannot be allocated to more than two production nodes for execution.

[0022] Further, the production nodes are parallel, and the relationship between different tasks on the same production node is non-preemptive.

[0023] Further, step 2 further comprises: dynamically updating the task pre-execution time of different tasks on different production nodes according to historical data, so as to dynamically update the ETC matrix.

[0024] Further, in step 5, the fitness function formula is:

[0025]

[0026] Wherein, β1 and β2 represent the pre-set adjustment coefficients and β1 + β2 = 1, if β1 > β2, it means that more attention is paid to the time to be spent in the scheduling process, if β1 < β2, it means that more attention is paid to the resources to be consumed in the scheduling process.

[0027] The beneficial effects of the present application are:

[0028] The remote sensing product production task scheduling method based on the particle swarm algorithm provided by the present application combines the characteristics of remote sensing product production tasks, constructs an ETC matrix and a CAL matrix, and then updates the fitness value of each particle according to the two matrices, so that the particle swarm algorithm is more suitable for solving the task scheduling problem in remote sensing products; and the size of the particle swarm and the maximum number of iterations are dynamically set according to the size of the task, which can reasonably improve the efficiency of processing tasks; at the same time, the ETC matrix is dynamically updated according to the historical data of the system, so that the forward feedback in the particle swarm algorithm is realized, the final task scheduling result is more in line with the actual demand, the total execution time of the task is shortened, and the system execution efficiency is improved. BRIEF DESCRIPTION OF DRAWINGS

[0029] Figure 1 Fig. 1 is a flowchart of a remote sensing product production task scheduling method based on a particle swarm algorithm provided by an embodiment of the present application;

[0030] Figure 2 Fig. 2 is another flowchart of a remote sensing product production task scheduling method based on a particle swarm algorithm provided by an embodiment of the present application;

[0031] Figure 3 Fig. 3 is a third flowchart of a remote sensing product production task scheduling method based on a particle swarm algorithm provided by an embodiment of the present application. DETAILED DESCRIPTION

[0032] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of the embodiments of this invention will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0033] Example 1

[0034] like Figure 1 As shown, this embodiment of the invention provides a method for scheduling remote sensing product production tasks based on particle swarm optimization, including:

[0035] S101: Given task set information and production node set information, the task set refers to a set consisting of remote sensing product production tasks that need to be assigned, and the production node set information refers to a set consisting of production nodes used to perform remote sensing product production tasks;

[0036] As one possible implementation, a task set is defined as Task = {T1, T2, ..., T...} m}, in which task T i The information is represented as: T i = {TaskID, OrderID, taskType}, i = 1, 2, ..., m; where TaskID, OrderID, and taskType represent the task ID, the order ID corresponding to the task, and the task type, respectively.

[0037] Define the set of production nodes as Node = {N1, N2, ..., N} n}, where each production node N j Each includes node IP and production capacity information, j = 1, 2, ..., n.

[0038] Specifically, the production capacity information includes CPU computing power, RAM memory size, disk storage space, and network bandwidth. Among these, CPU computing power is related to three factors: the number of cores, core frequency, and core clock cycle capability. The CPU computing power can be calculated based on these three factors.

[0039] S102: Construct the pre-execution time (ETC) matrix and resource allocation (CAL) matrix for the task; where ETC... ij The element in the i-th row and j-th column of the ETC matrix represents the pre-execution time of task i on production node j; CAL ijrepresents the element in the i-th row and j-th column of the CAL matrix, represents the resource consumed when task i is assigned to production node j for execution (for example, the CPU computing power occupied, the RAM memory size and the network bandwidth for executing the task); t = 1, 2, …, m, j = 1, 2, …, n; m represents the number of tasks, and n represents the number of production nodes.

[0040] The form of the constructed ETC matrix is as follows:

[0041]

[0042] The form of the constructed CAL matrix is as follows:

[0043]

[0044] Preferably, in order to make the values of the elements in the ETC matrix more reasonable, in the embodiment, the task pre-execution time of different tasks on different production nodes is dynamically updated according to historical data, so that the ETC matrix is dynamically updated, the fitness value of subsequent updates is more reliable, and finally a more accurate global optimal solution, that is, a task scheduling result, can be obtained.

[0045] S103: encode the particles in an indirect encoding mode, establish a particle swarm with a scale of N, construct a task allocation DIS matrix according to the particle positions, and set the number of iterations; wherein one particle represents a task allocation result, and the length of the encoding is the number of tasks; the larger the number of tasks, the larger the scale of the particle swarm and the number of iterations set; wherein DIS ij = 1 indicates that task i is allocated to production node j, and DIS ij = 0 indicates that no task is allocated to production node j.

[0046] Specifically, the basic parameters of the particle swarm algorithm are initialized, and the first generation of particles is initialized according to the number of remote sensing product tasks and the number of production nodes. According to the characteristics of the remote sensing product production environment, the indirect encoding mode is selected in this step. For example, it is assumed that the number of tasks m = 5 and the number of production nodes n = 3, and if one of the task allocation results is that the five tasks are allocated to node 1, node 2, node 1, node 3 and node 1 respectively, the encoding of one of the particles is 12131.

[0047] The form of the task allocation matrix constructed according to the particle positions is as follows:

[0048]

[0049] Correspondingly, the particles also need to be decoded, and the purpose is to obtain the task sequence allocated to each production node. For example, according to the above encoding, it can be decoded that the task sequence allocated to production node 1 is {1, 3, 5}.

[0050] In addition, in this step, the size of the particle swarm and the number of iterations are dynamically determined according to the size of the task amount. For example, when an order is received, if the task amount in the order is small, a small population can be selected and the number of iterations can be reduced, so as to reduce the waste of system resources and improve the resource utilization rate. If the task amount in the order is large, a large population can be selected and the number of iterations can be increased to meet the corresponding demand. By dynamically selecting the number of particle populations and the number of iterations, the efficiency of processing tasks can be reasonably improved.

[0051] S104: Select a typical linear decreasing strategy as an inertia weight algorithm;

[0052] Specifically, the formula of the typical linear decreasing strategy is as follows:

[0053]

[0054] When the inertia weight ω is large, the global convergence ability is strong, and when the inertia weight ω is small, the local convergence ability is strong. ω max ωmax represents a given maximum inertia weight value, which is usually 0.9; ω min ωmin represents a given minimum inertia weight value, which is usually 0.4; t max tmax represents a given maximum number of iterations, and t represents the current number of iterations.

[0055] S105: Calculate the fitness value of each particle according to the pre-execution time ETC matrix, the resource allocation CAL matrix, and the task allocation DIS matrix according to a set fitness function formula;

[0056] Specifically, by constantly updating the fitness value, the particle swarm eventually converges in the direction of the minimum comprehensive cost. In the embodiment of the application, the fitness function formula is:

[0057]

[0058] Wherein, β1 and β2 represent pre-given adjustment coefficients and β1 + β2 = 1, if β1 > β2, it means that more attention is paid to the time to be spent in the scheduling process, and if β1 < β2, it means that more attention is paid to the resources to be consumed in the scheduling process.

[0059] S106: Update the speed and position of the particle;

[0060] Specifically, in the embodiment of the application, the original speed and position update formula of the particle swarm algorithm is used to update the speed and position of the particle

[0061] v i (t+1) = ωv i (t) + C1rand1(Pbesti (t)-X i (t))+C2rand2(Gbest i (t)-X i (t))

[0062] x i (t+1)=X i (t)+v i (t+1)

[0063] where i={1,2,…,N}, N is the total number of particles, t is the current iteration number, ω is the inertia weight, v i ={v1,v2,…,v d} represents the flying speed of particle i, Pbest i ={p1,p2,…,p d} is the current individual optimal position of particle i, Gbest i ={g1,g2,…,g d} represents the optimal position on the path experienced by all particles, X i ={x1,x2,...,x n} represents the current position of particle i, C1, C2 are learning factors, usually 2, rand1, rand2 are random numbers between 0 and 1; d represents the dimension of the search space. In the embodiment of the application, d=m.

[0064] S107: updating the global optimal solution;

[0065] Specifically, the optimal solution Pbest i of each particle in the particle swarm is updated; the global optimal solution Gbest i of the particle swarm is updated.

[0066] S108: judging whether to converge or reach the set maximum iteration number, if yes, outputting the global optimal solution at this time and distributing the task to the corresponding production node for production according to the global optimal solution; if no, returning to step S105 to continue the next search process.

[0067] Embodiment 2

[0068] The particle swarm algorithm is used for task scheduling, and there may be a case that the optimal fitness value in the continuous multiple search processes is unchanged, at this time, it is considered that the particle swarm algorithm falls into local optimum, and for this case, on the basis of the above embodiment 1, the embodiment of the application further provides the following steps:

[0069] When the particle swarm algorithm falls into local optimum, the Naive Acceptance rule is executed to jump out of the local optimum. The Naive Acceptance rule refers to: when the fitness value of the new global optimal solution is better than the fitness value of the current global optimal solution, the new global optimal solution is accepted; otherwise, a random number p is generated, and if p is greater than or equal to 0.5, the new global optimal solution is accepted; wherein, the determination manner of the new global optimal solution is: within the set iteration range, the suboptimal solution is selected as the new global optimal solution.

[0070] Specifically, there is a current optimal fitness value in the algorithm iteration process. When it is detected that the fitness value does not change after multiple iterations, it is judged that the algorithm may fall into local optimum, and the Naive Acceptance rule is used to jump out of the local optimum. If the update is judged every time, the particle update may be updated in the non-optimal direction. In the update process, in order to avoid the result being worse than the previous result in the subsequent update, the fitness value which does not change after multiple iterations is saved. If the value without using the acceptance rule is better than the value using the acceptance rule, the original particle swarm algorithm result can also be used. Finally, when the algorithm ends, all the saved results are obtained, and the optimal solution is found from them.

[0071] For example, 200 iterations have been performed, but the fitness value does not change during 100 to 200 iterations. It is judged that the algorithm may fall into local optimum, and then the solution which is only second to the optimal value in the current set iteration range is found, and the acceptance rule is used.

[0072] The overall flowchart of the embodiment of the application is shown in Figure 2 .

[0073] Embodiment 3

[0074] On the basis of the above embodiment 1, the difference from embodiment 2 is that, in order to solve the problem that the algorithm falls into local optimum, the embodiment of the application further provides another solution, which specifically includes the following steps:

[0075] If the optimal fitness value in the continuous multiple search processes does not change, it is considered that the particle swarm algorithm falls into local optimum, and the task exchange rule is executed to jump out of the local optimum; the task exchange rule specifically includes:

[0076] The current global optimal solution is represented as a task allocation matrix. According to the corresponding relationship between the task allocation matrix and the ETC matrix, the task completion time of each production node is determined, and the task completion time matrix of all production nodes is formed, and the maximum task completion time T max , the minimum task completion time T min , and the distribution of the production node R are further determined.max the task in the production node R min with the minimum task pre-execution time; wherein R max refers to the production node corresponding to T max , R min refers to the production node corresponding to T min ; the task completion time of the production node is equal to the sum of the task execution times of all tasks thereon;

[0077] Suppose the task Task is assigned to the production node R min , if the task completion time of the production node R min is less than T max , the task Task is assigned to the production node R min , and the corresponding ETC matrix and task assignment matrix are updated; if the task completion time of the production node R min is not less than T max , the second smallest task completion time is further determined from the values in the task completion time matrix, and is taken as the new minimum task completion time T min , so as to re-perform task exchange; if all the minimum task completion times are traversed and task exchange cannot be performed, the second largest task completion time is further determined from the values in the task completion time matrix, and is taken as the new maximum task completion time T max , so as to re-perform task exchange.

[0078] Specifically, according to the predefined task exchange principle, if the task in the production node R max can be assigned to the production node R min , T max and T min are also taken as the maximum resource completion time and the minimum resource completion time of the next task exchange, and the same is true for the subsequent task exchanges. If the task in the production node R max cannot be assigned to the production node R min , the second smallest T min is found for the next task exchange, and the same is true for the subsequent task exchanges. If all the tasks are traversed and task assignment cannot be performed, the second largest resource completion time in the matrix is taken as T max for the next round of task exchange, and T min is still taken as the minimum resource completion time, until all the tasks are assigned. The overall flowchart of the embodiment of the present application is shown in Figure 3 .

[0079] In order to further explain the task exchange rule in the embodiment of the present application, an example is as follows:

[0080] Given the ETC matrix and the task assignment matrix DIS, as follows:

[0081]

[0082]

[0083] From the task allocation matrix DIS and the ETC matrix, it is known that there are 4 tasks and 4 production nodes, and the 4 tasks are allocated on the 4 production nodes respectively, and the corresponding relationship can obtain the value of the task completion time matrix, as shown in Table 1:

[0084] Table 1 Value of task completion time matrix

[0085] R1 R2 R3 R4 1 3 5 4

[0086] In Table 1, R1 to R4 represent 4 production nodes respectively. From Table 1, it is known that the node with the maximum task completion time is R3 node, T max is 5, and the node with the minimum task completion time is R1 node, T min is 1. In combination with the ETC matrix, it is known that the task allocated to the node R3 has the minimum execution time in R1, which is Task3, and the execution time of Task3 in R1 is 1. If Task3 is exchanged to the node R1, the task completion time in R1 is 1+1=2<5, and the task exchange can be performed. The task allocation matrix DIS and the task completion time matrix are updated (as shown in Table 2):

[0087]

[0088] Table 2 Value of task completion time matrix

[0089] R1 R2 R3 R4 2 3 0 4

[0090] From Table 2, it is known that the node with the maximum task completion time is R4 node, T max is 4, and the node with the minimum task completion time is R3 node, T min is 0. In combination with the ETC matrix, it is known that the task allocated to the node R4 has the minimum execution time in R3, which is Task4, and the execution time of Task4 in R3 is 1. If Task4 is exchanged to the node R3, the task completion time in R3 is 1+0=1<4, and the task exchange can be performed. The task allocation matrix DIS and the task completion time matrix are updated (as shown in Table 3):

[0091]

[0092] Table 3 Value of task completion time matrix

[0093] R1 R2 R3 R4 2 3 1 0

[0094] From table 3, the node with the maximum task completion time is the R2 node, T max is 3, and the node with the minimum task completion time is the R4 node, T min 0. In combination with the ETC matrix, it can be known that the task allocated to the R2 node is Task2, which has the minimum execution time in the R4 node, and the execution time of Task2 in the R4 node is 2. If Task2 is exchanged to the R4 node, the task completion time in the R4 node is 0+2=2<3, and the task exchange can be performed. The task allocation matrix and the task completion time matrix are updated (as shown in table 4):

[0095]

[0096] Table 4: Values of the task completion time matrix

[0097] R1 R2 R3 R4 2 0 1 2

[0098] From table 4, T max is 2, T min is 0, and 0+2=2 according to the task exchange principle, and the task exchange is ended.

[0099] It should be noted that in the entire task scheduling process, the same task cannot be allocated to more than two production nodes for execution. The production nodes are in parallel production; and the relationship between different tasks on the same production node is non-preemptive.

[0100] The task scheduling method of the present application can effectively reduce the execution time of the task and improve the efficiency of the system. The production node processes tile data in a shared folder, reduces the read and write of the file and the communication interaction of the production node, and the expansion of the node does not affect the running efficiency, so that the normal work of the remote sensing cluster and the running of the platform can be ensured.

[0101] Finally, it should be noted that: the above examples are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing examples, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing examples, or make equivalent replacement for part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for scheduling remote sensing product production tasks based on a particle swarm algorithm, characterized in that, Comprise: Step 1: given task set information and production node set information, the task set refers to a set consisting of remote sensing product production tasks that need to be assigned, and the production node set information refers to a set consisting of production nodes for executing remote sensing product production tasks; parallel production between various production nodes; the relationship between different tasks on the same production node is non-preemptive; Step 2: Construct the pre-execution time of tasks according to the task set and the production node set Matrix and resource allocation Matrix; wherein, represents an element in the i-th row and j-th column of the matrix, used to represent the task pre-execution time of task i on production node j; represents an element in the i-th row and j-th column of the matrix, used to represent the resource consumption required when task i is executed on production node j; Step 3: encode the particles by indirect encoding method, establish a particle group with scale N, and construct task allocation according to particle position Matrix, in the whole task scheduling process, the same task cannot be allocated to more than two production nodes for execution, and the maximum iteration number is set; wherein, a particle represents a task allocation result, and the length of the code is the number of tasks; the larger the task quantity is, the larger the scale of the particle group and the maximum iteration number set are; wherein, Indicates that the task is allocated to the production node , Indicates that the production node j has no allocated task; Step 4: select a typical linear decreasing strategy as the inertia weight algorithm; Step 5: According to the pre-execution time Matrix, resource allocation Matrix and task allocation DIS matrix calculates the fitness value of each particle according to the set fitness function formula; If the optimal fitness value in the continuous search process does not change, it is considered that the particle swarm algorithm falls into local optimum, at this time the task exchange rule is executed to jump out of the local optimum; the task exchange rule specifically includes: The current global optimal solution is represented as a task allocation matrix. Based on the correspondence between the task allocation matrix and the ETC matrix, the task completion time of each production node is determined. The task completion times of all production nodes are then used to construct a task completion time matrix, from which the maximum task completion time is further determined. Minimum task completion time and allocated to production nodes The task in the production node The task with the minimum pre-execution time; among them, It refers to and The corresponding production node, It refers to and The corresponding production node; the task completion time of the production node is equal to the sum of the task execution times of all tasks on it; Suppose task Task is assigned to production node , if the task completion time of production node is less than , task Task is assigned to production node , and the task assignment matrix and the task completion time matrix are updated; if the task completion time of production node is not less than , the second smallest task completion time is further determined from the values of the task completion time matrix, and is taken as the new minimum task completion time , and the task exchange is re-performed; if all the minimum task completion times cannot perform task exchange, the second largest task completion time is further determined from the values of the task completion time matrix, and is taken as the new maximum task completion time , and the task exchange is re-performed; Step 6: update the speed and position of the particle; Step 7: update the global optimal solution; Step 8: judge whether to converge or reach the set maximum iteration number, if yes, output the global optimal solution at this time, and distribute the tasks to the corresponding production nodes for production according to the global optimal solution; if not, return to step 5 to continue the next search process. 2.The method of claim 1, wherein, If the optimal fitness value in the continuous search process does not change, it is considered that the particle swarm algorithm falls into local optimum, at this time the Naive Acceptance rule is executed to jump out of the local optimum; the Naive Acceptance rule refers to: when the fitness value of the new global optimal solution is better than the fitness value of the current global optimal solution, the new global optimal solution is accepted; otherwise, a random number p is generated, if p >= 0.5, the new global optimal solution is accepted; wherein, the determination method of the new global optimal solution is: within the set iteration range, select the suboptimal solution as the new global optimal solution.

3. The method according to claim 1, wherein, In step 1, a task set Task = { , , …, } is set, wherein the information of a task is represented as: , ; wherein , , respectively represent a task ID, an order ID corresponding to the task, and a task type;​​​​ Set the production node set Node={ }, wherein each production node includes node IP and production capacity information, .

4. The method according to claim 3, wherein, The production capacity information includes CPU operation capacity, RAM memory size, disk storage space and network bandwidth.

5. The method of claim 1, wherein, In step 2, it also includes: dynamically updating the task pre-execution time of different tasks in different production nodes according to historical data, thereby dynamically updating the matrix. matrix.

6. The method of claim 1, wherein, In step 5, the fitness function formula is: wherein and denotes a pre-defined adjustment factor and denotes that during the scheduling process more emphasis is put on the time to be spent, if denotes that during the scheduling process more emphasis is put on the resources to be consumed, if denotes that during the scheduling process more emphasis is put on the resources to be consumed, if

Images