A dynamic weight optimization load balancing algorithm for wafer surface defect detection
By using dynamic weight optimization load balancing algorithm and adaptive discrete quantum particle swarm algorithm in wafer surface defect detection, node weights are adjusted in real time, the load imbalance problem in heterogeneous clusters is solved, and the stability and resource utilization of the system are improved. Rate.
Patent Information
- Application Number
- CN202210827115.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-07-13
- Publication Date
- 2024-11-01
- Estimated Expiration
- 2042-07-13
AI Technical Summary
In heterogeneous cluster environments, existing load balancing algorithms are difficult to effectively regulate node load, resulting in waste of resources and performance degradation, especially when task uncertainty and node performance differences are large in wafer surface defect detection. The weight design method is not accurate and flexible enough, resulting in load imbalance.
Dynamic weight optimization load balancing algorithm is adopted, and the load balancing model and objective function are established, combined with the adaptive discrete quantum particle swarm algorithm (ADQPSO), the node weights are adjusted in real time, and the CPU utilization, memory utilization and disk IO utilization are carried out according to the CPU utilization, memory utilization and disk IO utilization. Optimize to ensure the stability of load balancing.
It realizes timely adjustments to the cluster in a short time, maintains high load balancing, avoids further load imbalance, improves system stability and resource utilization, and reduces the fluctuation range of load balancing.
Smart Images

Figure CN115116879B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of computer networks, and in particular to a dynamic weight optimization load balancing algorithm for wafer surface defect detection. Background Art
[0002] With the increasing expansion of wafer specifications and output, and the increasing refinement of grain structures, in order to ensure the real-time detection, the speed improvement at the image algorithm level is no longer sufficient to meet the requirements, and the automatic optical inspection system equipped with only a single computer for detection is gradually becoming unable to cope with the task. Therefore, the number of computing devices must be increased to improve computing power, thereby shortening the detection cycle.
[0003] Cluster computing technology can manage multiple independent computers in a unified mode and provide a single interface to the outside world. The outside world can use the entire cluster as a single computer, effectively avoiding the drawbacks of decentralized management of multiple computers. Cluster computing technology is highly flexible and can add or delete internal nodes without being constrained by performance configuration, thereby effectively breaking through the bottleneck of a single machine and improving the computing power and availability of the system. However, in some cases, due to production conditions, the models, performance and other parameters of the servers used for detection work may not be completely consistent, which means that the load of each machine is often unbalanced. Inappropriate task allocation schemes are prone to cause the "barrel effect", which prolongs the entire detection cycle due to the excessive load of a certain machine, resulting in low equipment utilization. Therefore, it is crucial to design a relevant solution to balance the load on each machine under the condition of different server resources and performance, and reasonably dispatch each wafer defect detection task to each server for execution, so that the resources on each server can be fully and effectively utilized to avoid no load and overload.
[0004] The core of the cluster system load balancing problem lies in how to more reasonably distribute tasks to each node in the cluster to avoid resource waste and cluster performance degradation caused by uneven task distribution. Especially in the case of heterogeneous cluster systems, the differentiation of nodes in the system is a key consideration for load balancing. An excellent load balancing strategy can average the load on each node when the performance of each node is different and the number of resources is different to improve the overall performance attributes of the cluster and make the best use of the computing resources within the cluster. Nowadays, the research on cluster load balancing technology by scholars at home and abroad can be divided into two main directions: theoretical level and application level. At the theoretical research level, the current research direction mainly focuses on the optimization and improvement of load balancing algorithms based on dynamic feedback strategies combined with artificial intelligence or machine learning technology. At the application research level, the current research direction mainly focuses on the targeted optimization of relevant algorithms in existing load balancing devices.
[0005] The load balancing scheduling algorithms widely used in the industry today often design weights for each node in the cluster, so that the scheduling algorithm can take into account the differences in resources and performance of different nodes in a heterogeneous cluster, and on this basis, perform load balancing more accurately. Therefore, the design of weights plays a key role in the load balancing algorithm. At present, in the design of weights for heterogeneous cluster load balancing, the manual setting method still occupies a dominant position, that is, the operator judges the differences in resources and performance of each node in the cluster, and then sets a quantitative weight for each node based on his own experience, or designs a calculation formula based on the differences of each node, and obtains the weight of the node according to the formula.
[0006] However, this artificial setting method is not only inaccurate, but also lacks flexibility. In the wafer surface defect detection environment, the tasks to be detected generated by the image acquisition module are not completely consistent. The differences in wafer model specifications, acquisition locations, and image detection algorithms will all cause disturbances to the nodes. There are uncertainties and differences in the tasks, so the operating status of each node will fluctuate to a greater or lesser extent than expected, and it is not static. As the system runs longer and longer, the load gradually concentrates on nodes with larger weights, and nodes with smaller weights become increasingly idle, and the system will gradually become unbalanced in the later stages.
[0007] The working principle of the polling algorithm is to distribute tasks to all nodes in the cluster in an inherent order, treat all nodes equally, and do not consider the differences in node performance configuration. It is simple to implement, but it is only applicable to the situation where all nodes in the cluster have similar performance. The load balancing performance on heterogeneous clusters with large performance differences is poor. The weighted polling algorithm evolved from the polling algorithm. While inheriting the advantages of the polling algorithm's simple implementation, it can also take into account the particularity of cluster heterogeneity and the balance of scheduling. It is one of the most widely used load balancing algorithms in heterogeneous cluster environments. The weighted polling algorithm takes node performance differences into consideration. In the algorithm, weights are set for each node according to its performance. According to the idea of "the more capable, the more work", the size of the weight is positively correlated with the performance. By making each node receive tasks with a corresponding weight ratio, the purpose of load balancing is achieved. However, the load balancing degree of the polling algorithm and the smooth weighted polling algorithm fluctuates greatly throughout the process, and increases over time, and the stability is poor. Summary of the invention
[0008] The purpose of the present invention is to provide a dynamic weight optimization load balancing algorithm for wafer surface defect detection, so as to achieve timely adjustment of the cluster in a short time and maintain cluster stability. To achieve the above purpose, the present invention adopts the following technical solutions:
[0009] The present invention discloses a dynamic weight optimization load balancing algorithm for wafer surface defect detection, which includes the following processes: a. establishing a load balancing model for a computer cluster used for wafer surface defect detection; b. constructing an objective function, which is designed as a linear function of the load balancing degree of various indicators after introducing influencing factors:
[0010] f(x)=aσ cpu +bσ mem +cσ io , where a, b, and c are the influencing factors of CPU utilization, memory utilization, and disk IO utilization, respectively, which are determined according to the degree of influence of each indicator on cluster load balancing, and the sum of a, b, and c is 1; σ cpu , σ mem , σ io They are the current load balancing degree of the three indicators: CPU utilization, memory utilization, and disk IO utilization.
[0011] c. Feedback the load status of each node to the master control node at a fixed period. The master control node calculates the load balance based on the current load status of each node and determines whether it exceeds the threshold. If so, proceed to process d. If not, continue to monitor the real-time load status.
[0012] d. Using a dynamic weight optimization algorithm to solve the new weight; including the following steps:
[0013] d1. Randomly initialize each particle in the population.
[0014] d2. Calculate the optimal, worst, optimal, worst and average optimal positions of the initial individual, the group, based on the initial population.
[0015] d3. Each particle in the population updates its current position according to the position update formula.
[0016] d4. Calculate the particle function value and update the individual optimal, worst, group optimal, worst and average optimal positions.
[0017] d5. Determine whether the algorithm has reached the termination condition, that is, whether the maximum number of iterations has been reached or whether the solution quality has met the requirements. If so, output the optimal solution as the weighted solution for this time; if not, return to step d3 to continue optimizing.
[0018] e: Replace the original weight with the obtained weight solution.
[0019] f: The master node continues to distribute tasks according to the new weights.
[0020] Furthermore, in step d2, the position update formula of the adaptive discrete quantum particle swarm algorithm is:
[0021]
[0022] Among them, ⊙ is the individual mutation operator. The individual mutation operation randomly selects a mutation point in the multidimensional particle position space and generates a random probability r in the range of (0,1). If r is less than the mutation probability, the value of the mutation point is replaced by any random integer in the range of the dimension, and the remaining elements are retained, thereby generating a new individual; if r is greater than or equal to the mutation probability, the original individual is maintained unchanged.
[0023] is the crossover selection operator. The crossover selection operation uses uniform crossover and roulette wheel selection to generate new individuals. The operator Each dimension of the individuals at both ends generates a random probability r in the range of (0,1) i , i is the dimension index, if r i If the crossover probability is less than , each element is exchanged with a certain probability. After the two individuals cross, two new individuals are generated. At this time, the roulette wheel is used to select one of the two individuals to become the final individual obtained by the crossover selection operation. If r i If the crossover probability is greater than or equal to , no crossover selection is performed, and the operator is directly set to The previous individual is taken as the final individual obtained by this operation. β is the expansion and contraction factor, and the value range of β is (0,1). x(t) is the current particle position, x(t+1) is the updated particle individual position, α and u are both random numbers uniformly distributed on (0,1), pbest i is the optimal position of the individual particle, gbest is the optimal position of the particle group, mbest is the average optimal position of the particle group, p1 is the crossover probability of the advantage guidance term, p2 is the crossover probability of the individual exploration term, and p3 is the crossover probability of the comprehensive term.
[0024] Preferably, the calculation formula of the expansion-contraction factor β is as follows:
[0025]
[0026] Among them, β min With β max are the minimum and maximum values of the expansion-contraction factor β, respectively, worst 、f best and f(x) represent the worst value, optimal value and current value of the population respectively, maxiter is the maximum number of iterations, and iter is the current number of iterations.
[0027] Among them, the load vector quantization standard for establishing the load balancing model is as follows:
[0028]
[0029] Where load is the node N i The load rate at time t, L(N i ) t is the load of node N at time t, ML(N i ) is node N i Maximum load capacity, load L (N i ) by node N i The CPU load, memory load and disk IO load are jointly determined.
[0030] The real-time load matrix status of each node is shown as follows:
[0031]
[0032] When there are s tasks waiting to be scheduled, node N i The estimated load rate required is:
[0033]
[0034] Among them, cost is the load rate consumed by processing q tasks, x′ i is the result of normalizing the node weight.
[0035] Then node N i The expected load state when the assigned tasks are processed can be expressed as:
[0036]
[0037] The load balance of each resource of the node in process b is calculated by the mean square error formula:
[0038]
[0039] The fixed period in process c is 2 to 10 seconds. The master control node in process c calculates the load balance degree according to the current load status of each node and draws a load balance curve from the start of scheduling to the completion of all task processing with time as the horizontal axis.
[0040] Due to the adoption of the above scheme, the present invention has the following beneficial effects:
[0041] 1. This method sets up a feedback mechanism to feed back the load status of each node to the master node at a fixed period, and the algorithm adjusts the weights according to the current status of each node. The master node calculates the load balance based on the current load status of each node. When it exceeds the threshold, it can timely optimize the weights according to the current load status of the cluster, find a high-quality weight solution in a short time, and adjust the weights of the current load balancing algorithm to avoid the load balancing algorithm continuing to distribute according to the previous weights, resulting in further imbalance of the load. This process is repeated, and the cluster status is continuously monitored and adjusted to ensure that the entire cluster works in a stable and balanced environment.
[0042] 2. The present invention adopts an adaptive discrete quantum particle swarm algorithm for position update. The algorithm is an improvement and upgrade based on the quantum particle swarm algorithm (QPSO). The algorithm combines the selection, crossover, and mutation operators in the genetic algorithm to help QPSO be applied to discrete problems. At the same time, it introduces a parameter adaptive mechanism to help the algorithm enhance its search capability and improve convergence.
[0043] 3. The expansion and contraction factors of the present invention are adaptively adjusted according to the formula. When the individual particles are relatively advantageous, the β value can be appropriately reduced to improve the local search ability of the particles and focus on exploring the area around the particles. Otherwise, the β value should be appropriately increased to improve the global optimization ability.
[0044] 4. The load balancing degree using the algorithm of the present invention has a narrow fluctuation range, and has no obvious upward trend over time, and has high stability. BRIEF DESCRIPTION OF THE DRAWINGS
[0045] Figure 1 It is a flow chart of the dynamic weight optimization load balancing algorithm of the present invention.
[0046] Figure 2 It is a flow chart of the adaptive discrete quantum particle swarm algorithm (ADQPSO) of the present invention.
[0047] Figure 3 This is a comparison curve of the load balancing degree of the algorithm of the present invention, the polling algorithm and the smooth weighted polling algorithm. (In the figure, RR is the polling algorithm, SWRR is the smooth weighted polling algorithm, and ADSWRR is the dynamic weight optimization algorithm of the present invention) DETAILED DESCRIPTION
[0048] In order to enable those skilled in the art to better understand the technical solution of the present invention, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.
[0049] like Figure 1 As shown, the present invention discloses a dynamic weight optimization load balancing algorithm for wafer surface defect detection, including the following process:
[0050] a. Establish a load balancing model for the computer cluster used for wafer surface defect detection.
[0051] The load vector quantization standard for establishing the load balancing model is as follows:
[0052]
[0053] Where load is the node N i The load rate at time t, L(N i ) t is the load of node N at time t, ML(N i ) is node N i Maximum load capacity, load L (N i ) by node N i The CPU load, memory load and disk IO load are jointly determined.
[0054] The real-time load matrix status of each node is shown as follows:
[0055]
[0056] When there are s tasks waiting to be scheduled, node N i The estimated load rate required is:
[0057]
[0058] Among them, cost is the load rate consumed by processing q tasks, x′ i is the result of normalizing the node weight.
[0059] Then node N i The estimated load state when the assigned tasks are processed can be expressed as:
[0060]
[0061] The load balance of each resource of the node is calculated by the mean square error formula:
[0062]
[0063] b. Construct the objective function, which is designed as a linear function of the load balance of each indicator after introducing the influencing factors:
[0064] f(x)=aσ cpu +bσ mem +cσ io
[0065] Among them, a, b, and c are the influencing factors of CPU utilization, memory utilization, and disk IO utilization, respectively, which are determined according to the degree of influence of each indicator on cluster load balancing, and the sum of a, b, and c is 1; σ cpu , σ mem , σ io They are the current load balancing degree of the three indicators: CPU utilization, memory utilization, and disk IO utilization.
[0066] c. Periodically obtain real-time node load information
[0067] The load status of each node is fed back to the master node at a fixed period. The master node calculates the load balance based on the current load status of each node and determines whether the cluster maintains load balance. The judgment method is: pre-set the threshold and determine whether the load balance of each node exceeds the threshold. If so, proceed to process c. If not, continue to monitor the real-time load status.
[0068] The fixed period can be 2 to 10 seconds. The master node calculates the load balance degree according to the current load status of each node and draws a load balance curve from the start of scheduling to the completion of all tasks with time as the horizontal axis. Drawing the load balance curve makes it easy to observe the fluctuation of the load balance degree over a period of time.
[0069] d. Use dynamic weight optimization algorithm to solve new weights and optimize system resource allocation.
[0070] In order to quickly solve the high-quality weight solution, the present invention designs an adaptive discrete quantum particle swarm algorithm (ADQPSO), which combines the selection, crossover and mutation operators in the genetic algorithm to help QPSO be applied to discrete problems, and introduces a parameter adaptation mechanism to help the algorithm enhance its search capability and improve convergence.
[0071] like Figure 2 As shown, the adaptive discrete quantum particle swarm algorithm (ADQPSO) includes the following steps:
[0072] d1. Randomly initialize each particle in the population.
[0073] d2. Calculate the optimal, worst, optimal, worst and average optimal positions of the initial individual, the group, based on the initial population.
[0074] The particle swarm algorithm first randomly initializes a particle swarm in a given solution space. The number of variables in the problem to be optimized determines the dimension of the solution space. Each particle has an initial position and initial velocity, and then iterates to find the best solution. In each iteration, each particle updates its position and velocity in the solution space by tracking two extreme values. One is the optimal particle found by a single particle itself in the iteration, that is, the individual optimal and worst value (individual extreme value), and the other is the optimal solution particle of all particles in the iteration process, that is, the group optimal and worst value (global extreme value).
[0075] d3. Each particle in the population updates its current position according to the position update formula.
[0076] The position update formula of the adaptive discrete quantum particle swarm algorithm is:
[0077]
[0078] Among them, x(t) is the current particle position, x(t+1) is the updated individual particle position, α and u are random numbers uniformly distributed on (0, 1), and pbest i is the optimal position of the individual particle, gbest is the optimal position of the particle group, mbest is the average optimal position of the particle group, p1 is the crossover probability of the advantage guidance term, p2 is the crossover probability of the individual exploration term, and p3 is the crossover probability of the comprehensive term.
[0079] ⊙ is the individual mutation operator. The individual mutation operation randomly selects a mutation point in the multidimensional particle position space and generates a random probability r in the range of (0,1). If r is less than the mutation probability, the value of the mutation point is replaced by any random integer in the range of this dimension, and the remaining elements are retained, thereby generating a new individual. If r is greater than or equal to the mutation probability, the original individual is maintained unchanged.
[0080] is the crossover selection operator. The crossover selection operation uses uniform crossover and roulette wheel selection to generate new individuals. The operator Each dimension of the individuals at both ends generates a random probability r in the range of (0,1) i , i is the dimension index, if r i If the crossover probability is less than , each element is exchanged with a certain probability. After the two individuals cross, two new individuals are generated. At this time, the roulette wheel is used to select one of the two individuals to become the final individual obtained by the crossover selection operation. If r i If the crossover probability is greater than or equal to , no crossover selection is performed, and the operator is directly set to The previous individual is taken as the final individual obtained by this operation.
[0081] β is the expansion-contraction factor, and the value range of β is (0,1). The calculation formula of the expansion-contraction factor β is as follows:
[0082]
[0083] Among them, β min With β max are the minimum and maximum values of the expansion-contraction factor β, respectively, worst 、f best and f(x) represent the worst value, optimal value and current value of the population respectively, maxiter is the maximum number of iterations, and iter is the current number of iterations.
[0084] When individual particles are more advantageous, the β value should be appropriately reduced to improve the local search ability of particles and focus on exploring the area around the particles. Conversely, the β value should be appropriately increased to improve the global optimization ability. Generally speaking, the algorithm will focus more on global optimization ability in the early stage of optimization, so the β value is generally larger, while in the later stage of optimization, it will focus more on local optimization ability, and the β value will gradually decrease.
[0085] d4. Calculate the particle objective function value and update the individual optimal, worst, group optimal, worst and average optimal positions;
[0086] d5. Determine whether the algorithm has reached the termination condition, that is, whether the maximum number of iterations has been reached or whether the solution quality has met the requirements. If so, output the optimal solution as the weighted solution for this time; if not, return to step d3 to continue optimizing.
[0087] e: Replace the original weight with the obtained weight solution.
[0088] f: The master node continues to distribute tasks using the new weights. When all tasks are distributed, the algorithm ends. If not, it returns to process c to continue obtaining real-time node load information.
[0089] The algorithm of the present invention (ADSWRR), the existing polling algorithm (RR), and the smooth weighted polling algorithm (SWRR) are experimentally tested for load balancing in the same environment, and the process is as follows:
[0090] The load balancing situation was tested under the same task number of 10,000 wafer images. The real-time load of each service node was recorded every 2 seconds. After calculating the load balancing degree, a load balancing curve was drawn with time as the horizontal axis from the start of scheduling to the completion of all tasks. Figure 3 shown.
[0091] Depend on Figure 3It can be clearly seen that the load balancing degree of the polling algorithm and the smooth weighted polling algorithm fluctuates greatly throughout the process and increases slightly over time, while the load balancing degree of the present method is always stable at around 0.08, with a narrow fluctuation range and no obvious upward trend. According to calculations, during the entire experimental process, the average load balancing degree of the polling algorithm is 0.1021, the average load balancing degree of the weighted polling algorithm is 0.0948, and the average load balancing degree of the algorithm of the present invention is 0.0776, which are 24.0% and 18.1% higher than the other two algorithms, respectively.
[0092] In summary, this method first establishes a load balancing model and designs the objective function for the wafer defect detection cluster environment, and then proposes a dynamic weight optimization algorithm based on adaptive discrete quantum particle swarm (ADQPSO). The quantum particle swarm algorithm is used as the main basis. While maintaining the good global optimization ability of the quantum particle swarm algorithm, the quantum particle swarm algorithm is not suitable for discrete optimization problems. With the help of operators such as selection, mutation, and crossover in the genetic algorithm, the algorithm can be effectively applied to discrete problems. An adaptive mechanism is introduced for the expansion-contraction factor to help the algorithm improve population diversity in the early stage and improve local search capabilities in the later stage, so that it can dynamically adjust the weights in the load balancing algorithm according to the current status of the cluster to achieve the purpose of maintaining cluster load balance for a long time.
[0093] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by any technician familiar with the technical field within the technical scope disclosed by the present invention should be covered within the protection scope of the present invention.
Claims
1. A dynamic weight optimization load balancing algorithm for wafer surface defect detection, characterized in that: The process includes: a. Establish a load balancing model for the computer cluster used for wafer surface defect detection; the load vector quantization standard for establishing the load balancing model is as follows: Where load is the node N i The load rate at time t, L(N i ) t is the load of node N at time t, ML(N i ) is node N i Maximum load capacity, load L (N i ) by node N i The CPU load, memory load and disk IO load are jointly determined; b. Construct the objective function, which is designed as a linear function of the load balance of each indicator after introducing the influencing factors: f(x)=aσ cpu +bσ mem +cσ io Among them, a, b, and c are the influencing factors of CPU utilization, memory utilization, and disk IO utilization, respectively, which are determined according to the degree of influence of each indicator on cluster load balancing, and the sum of a, b, and c is 1; σ cpu , σ mem , σ io The three indicators are the current load balancing degree of CPU utilization, memory utilization and disk IO utilization; c. Feedback the load status of each node to the master node at a fixed period. The master node calculates the load balance according to the current load status of each node and determines whether it exceeds the threshold. If so, proceed to process d. If not, continue to monitor the real-time load status; d. Using a dynamic weight optimization algorithm to solve the new weight; including the following steps: d1. Randomly initialize each particle in the population; d2. Calculate the optimal, worst, optimal, worst and average optimal positions of the initial individual, the group, based on the initial population; d3. Each particle in the population updates its current position according to the position update formula; d4. Calculate the particle objective function value and update the individual optimal, worst, group optimal, worst and average optimal positions; d5. Determine whether the algorithm has reached the termination condition, that is, whether the maximum number of iterations has been reached or whether the solution quality has met the requirements. If so, output the optimal solution as the weighted solution for this time; if not, return to step d3 to continue optimizing; e: Replace the original weight with the obtained weight solution; f: The master node continues to distribute tasks according to the new weights; In step d2, the position update formula of the adaptive discrete quantum particle swarm algorithm is: Among them, ⊙ is the individual mutation operator. The individual mutation operation randomly selects a mutation point in the multidimensional particle position space and generates a random probability r in the range of (0,1). If r is less than the mutation probability, the value of the mutation point is replaced by any random integer in the range of the dimension, and the remaining elements are retained, thereby generating a new individual; if r is greater than or equal to the mutation probability, the original individual is maintained unchanged; is the crossover selection operator. The crossover selection operation uses uniform crossover and roulette wheel selection to generate new individuals. The operator Each dimension of the individuals at both ends generates a random probability r in the range of (0,1) i , i is the dimension index, if r i If the crossover probability is less than , each element is exchanged with a certain probability. After the two individuals cross, two new individuals are generated. At this time, the roulette wheel is used to select one of the two individuals to become the final individual obtained by the crossover selection operation. If r i If the crossover probability is greater than or equal to , no crossover selection is performed, and the operator is directly set to The previous individual is taken as the final individual obtained by the operation; β is the expansion and contraction factor, and the value range of β is (0,1); x(t) is the current particle position, x(t+1) is the updated individual particle position, α and u are both random numbers uniformly distributed on (0,1), pbest i is the optimal position of the individual particle, gbest is the optimal position of the particle group, mbest is the average optimal position of the particle group, p1 is the crossover probability of the advantage guidance term, p2 is the crossover probability of the individual exploration term, and p3 is the crossover probability of the comprehensive term; The expansion-contraction factor β is calculated as follows: Among them, β min With β max are the minimum and maximum values of the expansion-contraction factor β, respectively, worst 、f best and f(x) represent the worst value, optimal value and current value of the population respectively, maxiter is the maximum number of iterations, and iter is the current number of iterations.
2. The dynamic weight optimization load balancing algorithm for wafer surface defect detection according to claim 1, characterized in that: The real-time load matrix status of each node is shown as follows: When there are s tasks waiting to be scheduled, node N i The estimated load rate required is: Among them, cost is the load rate consumed by processing q tasks, x′ i is the result of normalizing the node weight; Then node N i The expected load state when the assigned tasks are processed can be expressed as: The load balance of each resource of the node in process b is calculated by the mean square error formula:
3. The dynamic weight optimization load balancing algorithm for wafer surface defect detection according to claim 1, characterized in that: The fixed period in process c is 2 to 10 seconds.
4. The dynamic weight optimization load balancing algorithm for wafer surface defect detection according to claim 1, characterized in that: The master control node in process c calculates the load balance degree according to the current load status of each node and draws a load balance curve from the start of scheduling to the completion of all tasks with time as the horizontal axis.
