A multi-cluster demand scheduling method based on anomaly perception

By obtaining and filtering exception data in a multi-cluster environment, calculating exception components and scheduling, the problem of insufficient real-time exception perception capabilities in the existing technology is solved, and efficient utilization of multi-cluster resources and improving task execution efficiency is achieved.

CN118656226BActive Publication Date: 2024-11-22HEFEI SHUJUAN INFORMATION TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411154264.2
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-08-22
Publication Date
2024-11-22
Estimated Expiration
2044-08-22

AI Technical Summary

Technical Problem

The existing multi-cluster scheduling methods are based on static rules and historical data, and lack the ability to perceive and handle real-time exceptions, resulting in problems such as waste of resources and task delays, making it difficult to achieve efficient and stable resource utilization in a multi-cluster environment.

Method used

By obtaining the number of transactions processed per second, filtering exception data, calculating exception components, and scheduling the cluster with the demand, dynamic balancing and load transfer across cluster resources can be achieved, and performance degradation caused by load inequality and resource bottlenecks or Service outage.

Benefits of technology

It effectively improves the average utilization rate of multi-cluster resources and task execution efficiency, ensures efficient and coordinated work of clusters in high-load environments, and avoids performance degradation or service interruption caused by load inequality or resource bottlenecks.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118656226B_ABST
    Figure CN118656226B_ABST
Patent Text Reader

Abstract

The present invention relates to the field of cluster scheduling, and provides a multi-cluster demand scheduling method based on abnormality perception, which obtains data on the number of transactions processed per second by multiple clusters during an active period, recorded as active data, screens out abnormal data from the active data, calculates the abnormal component of the cluster based on the abnormal data, and schedules the cluster based on the demand size of each cluster during the active period combined with the abnormal component. The method can effectively balance resource utilization across clusters, avoid performance degradation or service interruption caused by uneven load or resource bottlenecks, ensure the efficiency and stability of clusters in the process of processing tasks, and through predictive scheduling, can perform protective scheduling or load transfer before abnormalities occur in multiple clusters, achieve full scheduling of computing power between idle clusters and overloaded clusters, and ensure that multiple clusters can maintain efficient collaborative work in a complex computing environment with high load.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of cluster scheduling, and in particular to a multi-cluster demand scheduling method based on abnormality perception. Background Art

[0002] The development of cloud-native technology has provided a solid technical foundation for multi-clusters. By standardizing the underlying hardware, it has achieved load balancing and performance optimization of containers while improving resource utilization. Currently, multi-cluster architecture has become a key technology for large data centers to achieve efficient task execution.

[0003] Each cluster is an independent system that can be deployed in different places / environments. Generally speaking, determining the resource size of a cluster requires estimating the total demand for the target business to be processed. When the cluster faces insufficient resources, it can naturally be expanded by increasing the number of nodes. However, the cluster size cannot be increased indefinitely. One obvious reason is that the additional overhead of the control plane between clusters will increase accordingly, affecting the overall stability of the cluster.

[0004] In a multi-cluster environment, demand scheduling involves multiple key technical issues, including cluster resource allocation and load balancing based on the minimum and maximum resource amounts, and the consistent synchronization between the expansion and reduction of cluster containers by the control plane. At the same time, considering the performance bottlenecks and reliability issues of each cluster, when the cluster resource status changes, the monitoring and feedback mechanism based on resource rebalancing technology needs to be triggered to ensure the timeliness and correctness of resource scheduling.

[0005] However, most of the existing multi-cluster scheduling methods are based on static rules and historical data, lacking the ability to perceive and handle real-time abnormal situations, which easily leads to problems such as resource waste or task delays in practical applications. Therefore, a method that can dynamically perceive abnormalities and perform intelligent scheduling is needed to improve the average utilization of multi-cluster resources and the execution efficiency of processing tasks. Summary of the invention

[0006] The purpose of the present invention is to propose a multi-cluster demand scheduling method based on abnormal perception to solve one or more technical problems existing in the prior art and at least provide a beneficial option or create conditions.

[0007] The present invention provides a multi-cluster demand scheduling method based on abnormality perception, which obtains the data of the number of transactions processed per second by multiple clusters in the active period, recorded as active data, screens out abnormal data in the active data, calculates the abnormal component of the cluster according to the abnormal data, and schedules the cluster based on the demand size of each cluster in the active period combined with the abnormal component. The method can effectively balance the resource utilization across clusters, avoid performance degradation or service interruption caused by uneven load or resource bottleneck, ensure the efficiency and stability of the cluster in the process of processing tasks, and through predictive scheduling, can perform protective scheduling or load transfer before abnormalities occur in multiple clusters, realize full scheduling of computing power between idle clusters and overloaded clusters, and ensure that multiple clusters can maintain efficient collaborative work in a complex computing environment with high load.

[0008] In order to achieve the above object, according to one aspect of the present invention, a multi-cluster demand scheduling method based on abnormality perception is provided, and the method comprises the following steps:

[0009] Get the number of transactions processed per second by multiple clusters during the active period, and record it as active data;

[0010] Filter out abnormal data from active data and calculate the abnormal component of the cluster based on the abnormal data;

[0011] The clusters are scheduled based on the demand size of each cluster during the active period and the abnormal component.

[0012] Furthermore, data on the number of transactions processed per second by multiple clusters during an active period is obtained, wherein each cluster is composed of multiple nodes, and different clusters among the multiple clusters have different numbers of nodes, and the nodes are physical nodes or virtual nodes; the active period refers to any period of time when multiple clusters are processing tasks, and the number of transactions processed per second by a cluster refers to the number of transactions processed per second by the cluster during the active period.

[0013] Furthermore, the method of obtaining the data on the number of transactions processed per second by multiple clusters during the active period and recording it as the active data is as follows: let the number of clusters be N in total, let cla(i) represent the i-th cluster among the N clusters, let T represent the active period, let the length of T be M seconds, let tk(j) represent the j-th second in M ​​seconds (of the active period T), tk(j) is the time, Tps(i,j) represents the number of transactions processed by cla(i) at time tk(j), and Tps(i,j) is recorded as the active data.

[0014] Furthermore, the size of M is set to [3000, 5000].

[0015] Furthermore, the method of filtering out abnormal data in the activity data is:

[0016] S1. Let \(S = M\times N\). Store the \(S\) values (data) corresponding to \(Tps(i, j)\) in the array \(Tps\), and denote the \(k\)-th value among these \(S\) values as \(Tps(k)\). Randomly select a value from the array \(Tps\) and denote it as \(Tps(rand)\). Create an empty array \(Fd\), set the variable \(Curr = 1\) and the variable \(D = 0\).

[0017] From \(k = 1\) to \(k = S\), calculate the first difference \(fd(k)\) of \(Tps(k)\) in sequence. Then, for the \(S\) values \(Tps(1), Tps(2),\cdots, Tps(S)\) in \(Tps\), each corresponds to a first difference \(fd(1), fd(2),\cdots, fd(S)\). Add the value in \(Tps\) corresponding to the largest first difference to the array \(Fd\), and go to S2; where \(fd(k)=\vert1 - Tps(rand)\div Tps(k)\vert\).

[0018] S2. Execute the sub - statement, update the value of \(Curr\) to the non - zero minimum value among \(fd2(1), fd2(2),\cdots, fd2(S)\), and at the same time add the value in \(Tps\) corresponding to the non - zero minimum second difference to the array \(Fd\), and update the value of \(D\) to \(Curr\).

[0019] Execute the sub - statement again, and set the value of the variable \(x\) to \(x = 1\), then go to S3.

[0020] S3. If \(fd2(x)\) satisfies \(Curr\lt fd2(x)\lt SFd\), update the value of \(Curr\) to \(fd2(x)\), and at the same time add \(Tps(x)\) to the array \(Fd\), and then execute the sub - statement. After the sub - statement is executed, set the value of \(x\) to \(1\) and (re -) go to S3; if \(fd2(x)\) does not satisfy \(Curr\lt fd2(x)\lt SFd\), then go to S4; where \(SFd\) represents the sum of all values in the array \(Fd\).

[0021] S4. If the value of the variable \(x\) is less than \(N\), increase the value of \(x\) by \(1\) and go to S3; if the value of the variable \(x\) is equal to or greater than \(N\), output the array \(Fd\) and mark all values in the array \(Fd\) as abnormal data.

[0022] The beneficial effect of this step is that since there are many nodes in the cluster and data needs to be collected across clusters at the same time, and the cluster demand scheduling based on anomaly perception requires abnormal data as a support. The method in this step can quickly and effectively identify the abnormal points in the data by using the calculation methods of the first difference and the second difference, and by means of iterative and multiple calculations of the difference, continuously approach the limit of the abnormal value, so as to quickly screen out the points with large differences from the randomly selected value. Through the traversal from \(Tps(1)\) to \(Tps(S)\), while ensuring that all data points participate in the detection and screening, most of the non - abnormal data can be quickly filtered out, thus fully improving the screening efficiency.

[0023] Furthermore, the sub-statement is: If Curr>0.5×D, then the value at the end of the array Fd is recorded as Fd(end), and the second difference fd2(k) of Tps(k) is calculated from k=1 to k=S, then the S values ​​Tps(1), Tps(2),…, Tps(S) in Tps each correspond to a second difference fd2(1), fd2(2),…, fd2(S), where fd2(k)=|1-Fd(end)÷Tps(k)|; If Curr≤0.5×D, then the array Fd is output, and all values ​​in the array Fd are marked as abnormal data.

[0024] Furthermore, the method for calculating the abnormal component of the cluster based on the abnormal data is specifically as follows:

[0025] Create an all-zero array Cop of length N (i.e., array Cop contains N zero values), and use Cop(i) to represent the i-th value in array Cop;

[0026] The statistical algorithm is defined as follows: use the array TTps to store the M activity data Tps(i,1), Tps(i,2),…,Tps(i,M) corresponding to the cluster cla(i), accumulate the repeated values ​​between the array TTps and the array Fd, and record the accumulated value as Cop(i) (update Cop(i) from the original value of 0 to the accumulated value);

[0027] In the statistical algorithm, i is traversed from i=1 to i=N to obtain Cop(1), Cop(2),…, Cop(N), and Cop(1), Cop(2),…, Cop(N) are used as the abnormal components of clusters cla(1), cla(2),…, cla(N).

[0028] The beneficial effect of this step is: taking the abnormal component as the abnormal perception of each cluster in the current period, the abnormal component of each cluster is calculated by the intersection weight between the abnormal data and the cluster activity data, which reflects the frequency and warning level of abnormal behavior in the cluster. Through continuous monitoring and analysis, it can provide real-time warning and data basis for identifying problem clusters for the scheduling optimization of multiple clusters, effectively improving the accuracy and response speed of demand scheduling.

[0029] Furthermore, the method for scheduling the clusters based on the demand size of each cluster during the active period and the abnormal component is as follows: the demand size of the cluster during the active period and the abnormal component of the cluster are normalized, and the normalized value is used as the scheduling amount, then each cluster corresponds to a scheduling amount, and the P clusters with the highest scheduling amount and the P clusters with the lowest scheduling amount are selected for scheduling;

[0030] The scheduling method is to perform round-robin scheduling on all target nodes, where all target nodes refer to all nodes in the P clusters with the highest scheduling amounts and all nodes in the P clusters with the lowest scheduling amounts.

[0031] The beneficial effect of this step is: due to problems such as uneven load, resource bottlenecks or demand fluctuations, in order to ensure the efficient use and stability of resources in the multi-cluster system, it is necessary to balance the resources by demand scheduling, or when it is foreseen that the cluster will be maintained or may fail, the workload is scheduled to other clusters in advance, thereby avoiding service interruption and cluster performance degradation, and improving the resource utilization efficiency of the cluster.

[0032] Preferably, the round-robin scheduling can also be replaced by weighted round-robin scheduling or minimum load scheduling.

[0033] Furthermore, the demand size of the cluster during the active period refers to the sum of the network inbound and outbound traffic of the cluster during the active period.

[0034] The beneficial effects of the present invention are as follows: the method can effectively balance resource utilization across clusters, avoid performance degradation or service interruption due to uneven load or resource bottlenecks, ensure the efficiency and stability of clusters in processing tasks, and through predictive scheduling, perform protective scheduling or load transfer before abnormalities occur in multiple clusters, thereby achieving full scheduling of computing power between idle clusters and overloaded clusters, and ensuring that multiple clusters can maintain efficient collaborative work in a complex computing environment with high loads. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 Shown is a flow chart of a multi-cluster demand scheduling method based on anomaly perception. DETAILED DESCRIPTION

[0036] The following will be combined with the embodiments and drawings to clearly and completely describe the concept, specific structure and technical effects of the present invention, so as to fully understand the purpose, scheme and effect of the present invention. It should be noted that the embodiments and features in the embodiments of this application can be combined with each other without conflict.

[0037] In the description of the present invention, "several" means one or more, "more" means more than two, "greater than", "less than", "exceed" etc. are understood as not including the number itself, and "above", "below", "within" etc. are understood as including the number itself. If there is a description of "first" or "second", it is only used for the purpose of distinguishing the technical features, and cannot be understood as indicating or implying the relative importance or implicitly indicating the number of the indicated technical features or implicitly indicating the order of the indicated technical features.

[0038] like Figure 1 The flowchart of a multi-cluster demand scheduling method based on abnormal perception according to the present invention is shown below. Figure 1 A multi-cluster demand scheduling method based on abnormality perception according to an embodiment of the present invention is described below.

[0039] The present invention proposes a multi-cluster demand scheduling method based on abnormality perception, and the method comprises the following steps:

[0040] Get the number of transactions processed per second by multiple clusters during the active period, and record it as active data;

[0041] Filter out abnormal data from active data and calculate the abnormal component of the cluster based on the abnormal data;

[0042] The clusters are scheduled based on the demand size of each cluster during the active period and the abnormal component.

[0043] Furthermore, data on the number of transactions processed per second by multiple clusters during an active period is obtained, wherein each cluster is composed of multiple nodes, and different clusters among the multiple clusters have different numbers of nodes, and the nodes are physical nodes or virtual nodes; the active period refers to any period of time when multiple clusters are processing tasks, and the number of transactions processed per second by a cluster refers to the number of transactions processed per second by the cluster during the active period.

[0044] Furthermore, the method of obtaining the number of transactions processed per second by multiple clusters during the active period and recording it as the active data is as follows: let the number of clusters be N in total, let cla(i) represent the i-th cluster among the N clusters, i is the sequence number, i=1,2,…,N, let T represent the active period, let the length of T be M seconds, let tk(j) represent the j-th second in M ​​seconds (of the active period T), j is the sequence number, j=1,2,…,M, tk(j) is the time, tk(j)=tk(1),tk(2),…,tk(M), let Tps(i,j) represent the number of transactions processed by cla(i) at time tk(j), and record Tps(i,j) as the active data.

[0045] Specifically, the size of M is set to 4000.

[0046] Specifically, the value range of Tps(i,j) corresponds to the following two-dimensional matrix:

[0047]

[0048] Among them, the i-th row of the matrix of size M×N corresponds to the activity data Tps(i,1), Tps(i,2),…, Tps(i,M) of cluster cla(i) in the activity period T (a total of M seconds); for example, the first row of the matrix Tps(1,1), Tps(1,2),…, Tps(1,M) is the activity data of cla(1) in the activity period T (a total of M seconds), and the N-th row of the matrix Tps(N,1), Tps(N,2),…, Tps(N,M) is the activity data of cla(N) in the activity period T.

[0049] Furthermore, the method of filtering out abnormal data in the activity data is:

[0050] S1, from i=1,2,…,N and j=1,2,…,M, we know that Tps(i,j) corresponds to M×N data, let S=M×N, use array Tps to store the S values ​​(data) corresponding to Tps(i,j), use Tps(k) to represent the kth of these S values, k is the sequence number, k=1,2,…,S, randomly select a value from array Tps and record it as Tps(rand), create a blank array Fd, set variable Curr=1 and variable D=0;

[0051] From k=1 to k=S, calculate the first difference fd(k) of Tps(k) in sequence, then the S values ​​Tps(1), Tps(2), ..., Tps(S) in Tps each correspond to a first difference fd(1), fd(2), ..., fd(S), add the value corresponding to the largest first difference in Tps to the array Fd and transfer to S2; where fd(k)=|1-Tps(rand)÷Tps(k)|;

[0052] S2, execute the substatement, update the value of Curr to the non-zero minimum value among fd2(1), fd2(2), ..., fd2(S), add the value corresponding to the non-zero minimum second difference in Tps to the array Fd, and update the value of D to Curr;

[0053] Execute the substatement again and set the value of variable x to x=1. Use fd2(x) to represent the xth value of variable x in fd2(1), fd2(2), …, fd2(S). Go to S3.

[0054] S3, if fd2(x) satisfies Curr < fd2(x) < SFd, update the value of Curr to fd2(x), add Tps(x) to the array Fd, and execute the sub-statement. After the sub-statement is executed, set the value of x to 1 and (re-)go to S3; if fd2(x) does not satisfy Curr < fd2(x) < SFd, go to S4; where SFd represents the sum of all values in the array Fd.

[0055] S4, if the value of the variable x is less than N, increase the value of x by 1 and go to S3; if the value of the variable x is equal to or greater than N, output the array Fd and mark all values in the array Fd as abnormal data.

[0056] Further, the sub-statement is: if Curr > 0.5×D, record the value at the end of the array Fd as Fd(end), calculate the second difference fd2(k) of Tps(k) in sequence from k = 1 to k = S. Then, the S values Tps(1), Tps(2), …, Tps(S) in Tps each correspond to a second difference fd2(1), fd2(2), …, fd2(S), where fd2(k) = |1 - Fd(end) ÷ Tps(k)|; if Curr ≤ 0.5×D, output the array Fd and mark all values in the array Fd as abnormal data.

[0057] Executing the sub-statement multiple times is to continuously update the second differences corresponding to the values in Tps and use this as a screening criterion to continuously increase the value of Curr so as to achieve the effect of filtering most data points (saving search overhead).

[0058] Further, the method for calculating the abnormal component of the cluster based on the abnormal data is specifically as follows:

[0059] Create a zero-filled array Cop of length N (i.e., the array Cop contains N zero values), and represent the i-th value in the array Cop as Cop(i);

[0060] Define the statistical algorithm as: store the M active data Tps(i,1), Tps(i,2), …, Tps(i,M) corresponding to the cluster cla(i) in the array TTps, accumulate the duplicate values between the array TTps and the array Fd, and record this accumulated value as Cop(i) (update Cop(i) from the original 0 value to this accumulated value);

[0061] The duplicate values between the array TTps and the array Fd refer to those values that appear in both the array TTps and the array Fd;

[0062] In the statistical algorithm, i is traversed from i=1 to i=N to obtain Cop(1), Cop(2),…, Cop(N), and Cop(1), Cop(2),…, Cop(N) are used as the abnormal components of clusters cla(1), cla(2),…, cla(N).

[0063] Furthermore, the method for scheduling the clusters based on the demand size of each cluster during the active period and the abnormal component is as follows: the demand size of the cluster during the active period and the abnormal component of the cluster are normalized, and the normalized value is used as the scheduling amount, then each cluster corresponds to a scheduling amount, and the P clusters with the highest scheduling amount and the P clusters with the lowest scheduling amount are selected for scheduling;

[0064] The scheduling method is to perform round-robin scheduling on all target nodes, where all target nodes refer to all nodes in the P clusters with the highest scheduling amounts and all nodes in the P clusters with the lowest scheduling amounts.

[0065] Specifically, the method of normalizing the demand size of the cluster during the active period and the abnormal component of the cluster to obtain the normalized value is:

[0066] Sod(i) and Cop(i) represent the demand size of cluster cla(i) during the active period and the abnormal component of cluster cla(i) respectively. The calculation method of the normalized value norm(i) is norm(i)=Sod(i)÷Sum(Sod)+Cop(i)÷Sum(Cop) (sum normalization);

[0067] Or, norm(i)=Sod(i)÷Max(Sod)+Cop(i)÷Max(Cop) (maximum normalization);

[0068] Among them, Sum(Sod) represents the sum of the demand sizes of N clusters during the active period, and Sum(Cop) represents the sum of the abnormal components of N clusters; Max(Sod) represents the maximum value of the demand sizes of N clusters during the active period, and Max(Cop) represents the maximum value of the abnormal components of N clusters.

[0069] Furthermore, the demand size of the cluster during the active period refers to the sum of the network inbound and outbound traffic of the cluster during the active period.

[0070] Specifically, query the total network inflow and outflow traffic of the cluster in the historical period. In PromQL, use the following statement to query SnR=sum(rate(node_network_receive_bytes_total[T])), where SnR represents the total network inflow and outflow traffic of the cluster in the active period T.

[0071] The present invention provides a multi-cluster demand scheduling method based on abnormal perception, obtains the data of the number of transactions processed per second by multiple clusters in the active period, records it as active data, screens out abnormal data in the active data, calculates the abnormal component of the cluster according to the abnormal data, and schedules the cluster based on the demand size of each cluster in the active period combined with the abnormal component. The method can effectively balance the resource utilization across clusters, avoid performance degradation or service interruption caused by uneven load or resource bottleneck, ensure the efficiency and stability of the cluster in the process of processing tasks, and through predictive scheduling, can perform protective scheduling or load transfer before the abnormality occurs in multiple clusters, realize the full scheduling of computing power between idle clusters and overloaded clusters, and ensure that multiple clusters can maintain efficient collaborative work in a complex computing environment with high load. Although the description of the present invention has been quite detailed and several embodiments are described in particular, it is not intended to be limited to any of these details or embodiments or any special embodiments, so as to effectively cover the predetermined scope of the present invention. In addition, the above describes the present invention with the foreseeable embodiments of the inventor, the purpose of which is to provide a useful description, and those non-substantial changes to the present invention that have not yet been foreseen can still represent equivalent changes to the present invention.

Claims

1. A multi-cluster demand scheduling method based on abnormality perception, characterized in that: The method includes the following steps: Obtain the number of transactions processed per second data of multiple clusters during the active period, denoted as active data; Filter out abnormal data from the active data, and calculate the abnormal component of the cluster according to the abnormal data; Schedule the clusters by combining the demand magnitude of each cluster during the active period with the abnormal component; The method for obtaining the number of transactions processed per second data of multiple clusters during the active period, denoted as active data, is specifically as follows: Denote the total number of multiple clusters as N, represent the i-th cluster among these N clusters as cla(i), represent the active period as T, denote the length of T as M seconds, represent the j-th second among the M seconds as tk(j), where tk(j) is a moment, and represent the number of transactions processed by cla(i) at the moment tk(j) as Tps(i,j), and record Tps(i,j) as the active data; The method for filtering out abnormal data from the active data is as follows: S1, Denote S = M×N, store the S values corresponding to Tps(i,j) in the array Tps, represent the k-th value among these S values as Tps(k), randomly select a value from the array Tps and denote it as Tps(rand), create a blank array Fd, set the variable Curr = 1 and the variable D = 0; From k = 1 to k = S, calculate the first difference fd(k) of Tps(k) in sequence, then the S values Tps(1), Tps(2), …, Tps(S) in Tps each correspond to the first differences fd(1), fd(2), …, fd(S), add the value in Tps corresponding to the largest first difference to the array Fd, and go to S2; where, fd(k)=|1 - Tps(rand)÷Tps(k)|; S2, Execute the sub-statement, update the value of Curr to the non-zero minimum value among fd2(1), fd2(2), …, fd2(S), at the same time add the value in Tps corresponding to the non-zero minimum second difference to the array Fd, and update the value of D to Curr; Execute the sub-statement again, and set the value of the variable x to x = 1, and go to S3; S3, If fd2(x) satisfies Curr < fd2(x) < SFd, update the value of Curr to fd2(x), at the same time add Tps(x) to the array Fd, and execute the sub-statement, after the sub-statement is executed, set the value of x to 1, and go to S3; If fd2(x) does not satisfy Curr < fd2(x) < SFd, then go to S4; where, SFd represents the sum of all values in the array Fd; S4, If the value of the variable x is less than N, increase the value of x by 1, and go to S3; If the value of the variable x is equal to or greater than N, output the array Fd, and mark all values in the array Fd as abnormal data; The sub-statements are: If Curr>0.5×D, then the value at the end of the array Fd is recorded as Fd(end), and the second difference fd2(k) of Tps(k) is calculated from k=1 to k=S. Then the S values ​​Tps(1), Tps(2),…, Tps(S) in Tps each correspond to the second difference fd2(1), fd2(2),…, fd2(S), where fd2(k)=|1-Fd(end)÷Tps(k)|; If Curr≤0.5×D, then the array Fd is output, and all values ​​in the array Fd are marked as abnormal data; The specific method for calculating the abnormal component of a cluster based on abnormal data is: Create an all-zero array Cop of length N, and use Cop(i) to represent the i-th value in the array Cop; The statistical algorithm is defined as follows: use the array TTps to store the M activity data Tps(i,1), Tps(i,2),…,Tps(i,M) corresponding to the cluster cla(i), accumulate the repeated values ​​between the array TTps and the array Fd, and record the accumulated value as Cop(i); In the statistical algorithm, i is traversed from i=1 to i=N to obtain Cop(1), Cop(2),…, Cop(N), and Cop(1), Cop(2),…, Cop(N) are used as the abnormal components of clusters cla(1), cla(2),…, cla(N).

2. According to the method of multi-cluster demand scheduling based on abnormal perception according to claim 1, it is characterized in that: Obtain data on the number of transactions processed per second by multiple clusters during an active period, wherein each cluster consists of multiple nodes, and different clusters in the multiple clusters have different numbers of nodes, and the nodes are physical nodes or virtual nodes; an active period refers to any period in the process of multiple clusters processing tasks, and the number of transactions processed per second of a cluster refers to the number of transactions processed per second by the cluster during the active period.

3. The method for multi-cluster demand scheduling based on abnormality perception according to claim 1 is characterized in that: The method for scheduling clusters based on the demand size of each cluster during the active period and the abnormal component is as follows: the demand size of the cluster during the active period and the abnormal component of the cluster are normalized, and the normalized value is used as the scheduling amount. Then, each cluster corresponds to a scheduling amount, and the P clusters with the highest scheduling amount and the P clusters with the lowest scheduling amount are selected for scheduling; The scheduling method is to perform round-robin scheduling on all target nodes, where all target nodes refer to all nodes in the P clusters with the highest scheduling amounts and all nodes in the P clusters with the lowest scheduling amounts.

4. The method for multi-cluster demand scheduling based on abnormality perception according to claim 3 is characterized in that: The demand volume of a cluster during the active period refers to the total network inbound and outbound traffic of the cluster during the active period.