A cloud computing resource scheduling method and system based on multi-dimensional resource fingerprint

Abstract

The present invention provides a cloud computing resource scheduling method and system based on multi-dimensional resource fingerprints in the field of cloud computing resource management technology. The method includes: step S1, collecting cloud computing tasks and cloud computing resource usage data; step S2, extracting statistical features from the cloud computing resource usage data to construct a high-dimensional resource fingerprint matrix; step S3, performing dimensionality reduction operations on the high-dimensional resource fingerprint matrix to obtain a compressed resource fingerprint matrix; step S4, calculating the task affinity between each cloud computing task based on the compressed resource fingerprint matrix to construct a task affinity matrix; step S5, constructing system affinity based on the task affinity matrix, setting resource constraints, and optimizing system affinity through resource constraints; step S6, scheduling cloud computing resources based on the compressed resource fingerprint matrix and system affinity. The advantages of the present invention are: greatly improving the resource utilization, timeliness, stability and energy efficiency of cloud computing resource scheduling.

Description

Technical Field

[0001] The present invention relates to the technical field of cloud computing resource management, and in particular to a cloud computing resource scheduling method and system based on multi-dimensional resource fingerprints. Background Art

[0002] With the widespread application and rapid development of cloud computing, how to efficiently manage increasingly complex cloud computing resources (cloud infrastructure) and meet the diverse needs of users and applications has become a key issue that cloud service providers need to solve. For the management of cloud computing resources, the traditional management method is to combine static resource configuration with coarse-grained dynamic adjustment strategy. However, the traditional method faces many challenges and is difficult to adapt to the high dynamics, high heterogeneity and high uncertainty of the cloud environment, mainly including the following problems:

[0003] 1. Cloud workloads present complex and dynamic characteristics, while static resource configuration lacks the ability to adapt to load changes, resulting in a prominent imbalance in the supply and demand of cloud computing resources. 2. There are a large number of heterogeneous tasks in the cloud environment, and coarse-grained dynamic adjustment strategies cannot match personalized task requirements and performance goals, resulting in problems such as waste of cloud computing resources and task delays. 3. Traditional management methods have difficulty capturing the fine-grained resource usage patterns and mutual influences between tasks, resulting in frequent cloud computing resource conflicts and performance interference problems. 4. The trade-off between the utilization efficiency of cloud computing resources and energy costs is becoming increasingly prominent.

[0004] Therefore, how to provide a cloud computing resource scheduling method and system based on multi-dimensional resource fingerprints to improve the resource utilization, timeliness, stability and energy efficiency of cloud computing resource scheduling has become a technical problem that needs to be solved urgently. Summary of the invention

[0005] The technical problem to be solved by the present invention is to provide a cloud computing resource scheduling method and system based on multi-dimensional resource fingerprints, so as to improve the resource utilization, timeliness, stability and energy efficiency of cloud computing resource scheduling.

[0006] In a first aspect, the present invention provides a cloud computing resource scheduling method based on multi-dimensional resource fingerprints, comprising the following steps:

[0007] Step S1, collecting cloud computing tasks of each physical machine running in the cloud environment, and collecting cloud computing resource usage data of each cloud computing task through a sliding window;

[0008] Step S2: extracting statistical features including at least a mean, a standard deviation, a maximum value, and a minimum value from the cloud computing resource usage data, and constructing a high-dimensional resource fingerprint matrix for each cloud computing task based on the statistical features;

[0009] Step S3, performing a dimensionality reduction operation on the high-dimensional resource fingerprint matrix by a principal component analysis method to obtain a compressed resource fingerprint matrix;

[0010] Step S4: calculating the task affinity between each cloud computing task based on the compressed resource fingerprint matrix, and constructing a task affinity matrix based on each task affinity;

[0011] Step S5: constructing system affinity based on the task affinity matrix, setting a greedy task placement strategy, setting a resource constraint condition based on the greedy task placement strategy, and optimizing the system affinity through the resource constraint condition;

[0012] Step S6: scheduling the cloud computing resources of each physical machine based on the compressed resource fingerprint matrix and the system affinity, and ensuring that the cloud computing resources allocated to each physical machine do not exceed the rated resources based on the proportional allocation method during the scheduling process;

[0013] Step S7: monitor the operating status of each physical machine, and migrate the cloud computing task based on the operating status and task affinity;

[0014] Step S8: During the cloud computing resource scheduling process, the compressed resource fingerprint matrix and task affinity are dynamically updated.

[0015] Furthermore, in step S1, the cloud computing resource usage data includes at least CPU usage, memory usage, disk I / O speed, and network bandwidth usage;

[0016] In step S4, the calculation formula of the task affinity is: ;

[0017] in, represents the i-th compressed resource fingerprint matrix; represents the jth compressed resource fingerprint matrix; express The corresponding cloud computing tasks and The task affinity of the corresponding cloud computing task; α represents the trade-off factor; Sim() represents the similarity measurement function; Comp() represents the complementarity measurement function.

[0018] Furthermore, the step S5 is specifically as follows:

[0019] Based on the task affinity matrix, the task affinities between each cloud computing task running in each physical machine are obtained, the task affinities of each physical machine are summed up to obtain the physical machine affinity, and the task affinities of each physical machine are summed up to obtain the system affinity;

[0020] Set a greedy task placement strategy , specifically:

[0021] For each cloud computing task to be scheduled, k , calculate t k The accumulated affinity with the scheduled cloud computing tasks is used to select a physical machine that maximizes the accumulated affinity and satisfies the resource constraint conditions for allocation;

[0022] A resource constraint condition is set based on the greedy task placement strategy: ;

[0023] Among them, t k represents the kth cloud computing task; r k Indicates t k The resource demand vector of i represents the i-th physical machine; Indicates allocation to m i A collection of cloud computing tasks; c i Indicates m i The resource capacity vector; M represents the total number of physical machines; represents the greedy task placement strategy; Indicates system affinity;

[0024] The system affinity is optimized through the resource constraint conditions.

[0025] Furthermore, the step S6 is specifically as follows:

[0026] Obtaining statistical features based on the compressed resource fingerprint matrix, scheduling the cloud computing resources of each physical machine based on the statistical features, constraining the scheduling process through the system affinity, and ensuring that the cloud computing resources allocated to each physical machine do not exceed the rated resources based on the proportional allocation method;

[0027] The proportional allocation method is specifically as follows: when the cloud computing resources required for all cloud computing tasks running on the physical machine exceed the rated resources, a scaling factor is calculated based on the ratio of the required cloud computing resources to the rated resources, and the allocated cloud computing resources are adjusted based on the scaling factor so that the cloud computing resources allocated to the physical machine do not exceed the rated resources.

[0028] Furthermore, the step S7 is specifically as follows:

[0029] Monitoring the operating status of each physical machine, wherein the operating status is overload and underload;

[0030] The cloud computing task of the physical machine whose running status is overloaded is migrated to the physical machine whose running status is underloaded and whose task affinity gain is maximized for execution.

[0031] In a second aspect, the present invention provides a cloud computing resource scheduling system based on multi-dimensional resource fingerprints, comprising the following modules:

[0032] A cloud computing task and resource collection module, used to collect cloud computing tasks of each physical machine running in the cloud environment, and collect cloud computing resource usage data of each cloud computing task through a sliding window;

[0033] A high-dimensional resource fingerprint matrix construction module is used to extract statistical features including at least a mean, a standard deviation, a maximum value, and a minimum value from the cloud computing resource usage data, and to construct a high-dimensional resource fingerprint matrix for each cloud computing task based on the statistical features;

[0034] A compressed resource fingerprint matrix construction module is used to perform a dimensionality reduction operation on the high-dimensional resource fingerprint matrix through a principal component analysis method to obtain a compressed resource fingerprint matrix;

[0035] A task affinity matrix construction module, used for calculating the task affinity between each cloud computing task based on the compressed resource fingerprint matrix, and constructing a task affinity matrix based on each task affinity;

[0036] A system affinity construction module, used to construct system affinity based on the task affinity matrix, set a greedy task placement strategy, set a resource constraint condition based on the greedy task placement strategy, and optimize the system affinity through the resource constraint condition;

[0037] A cloud computing resource scheduling module, used to schedule the cloud computing resources of each physical machine based on the compressed resource fingerprint matrix and the system affinity, and to ensure that the cloud computing resources allocated to each physical machine do not exceed the rated resources based on the proportional allocation method during the scheduling process;

[0038] A cloud computing task migration module, used to monitor the operating status of each physical machine and migrate the cloud computing tasks based on the operating status and task affinity;

[0039] The dynamic update module is used to dynamically update the compressed resource fingerprint matrix and task affinity during the cloud computing resource scheduling process.

[0040] Further, in the cloud computing task and resource acquisition module, the cloud computing resource usage data includes at least CPU usage, memory usage, disk I / O speed, and network bandwidth usage;

[0041] In the task affinity matrix construction module, the calculation formula of the task affinity is: ;

[0042] in, represents the i-th compressed resource fingerprint matrix; represents the jth compressed resource fingerprint matrix; express The corresponding cloud computing tasks and The task affinity of the corresponding cloud computing task; α represents the trade-off factor; Sim() represents the similarity measurement function; Comp() represents the complementarity measurement function.

[0043] Furthermore, the system affinity building module is specifically used for:

[0044] Based on the task affinity matrix, the task affinities between each cloud computing task running in each physical machine are obtained, the task affinities of each physical machine are summed up to obtain the physical machine affinity, and the task affinities of each physical machine are summed up to obtain the system affinity;

[0045] Set a greedy task placement strategy , specifically:

[0046] For each cloud computing task to be scheduled, k , calculate t k The accumulated affinity with the scheduled cloud computing tasks is used to select a physical machine that maximizes the accumulated affinity and satisfies the resource constraint conditions for allocation;

[0047] A resource constraint condition is set based on the greedy task placement strategy: ;

[0048] Among them, t k represents the kth cloud computing task; r k Indicates t k The resource demand vector of i represents the i-th physical machine; Indicates allocation to m i A collection of cloud computing tasks; c i Indicates m i The resource capacity vector; M represents the total number of physical machines; represents the greedy task placement strategy; Indicates system affinity;

[0049] The system affinity is optimized through the resource constraint conditions.

[0050] Furthermore, the cloud computing resource scheduling module is specifically used to:

[0051] Obtaining statistical features based on the compressed resource fingerprint matrix, scheduling the cloud computing resources of each physical machine based on the statistical features, constraining the scheduling process through the system affinity, and ensuring that the cloud computing resources allocated to each physical machine do not exceed the rated resources based on the proportional allocation method;

[0052] The proportional allocation method is specifically as follows: when the cloud computing resources required for all cloud computing tasks running on the physical machine exceed the rated resources, a scaling factor is calculated based on the ratio of the required cloud computing resources to the rated resources, and the allocated cloud computing resources are adjusted based on the scaling factor so that the cloud computing resources allocated to the physical machine do not exceed the rated resources.

[0053] Furthermore, the cloud computing task migration module is specifically used to:

[0054] Monitoring the operating status of each physical machine, wherein the operating status is overload and underload;

[0055] The cloud computing task of the physical machine whose running status is overloaded is migrated to the physical machine whose running status is underloaded and whose task affinity gain is maximized for execution.

[0056] The advantages of the present invention are:

[0057] By collecting cloud computing tasks of each physical machine running in the cloud environment, collecting cloud computing resource usage data of each cloud computing task through a sliding window, and extracting statistical features including at least mean, standard deviation, maximum value and minimum value from the cloud computing resource usage data, a high-dimensional resource fingerprint matrix of each cloud computing task is constructed based on each statistical feature, and then the high-dimensional resource fingerprint matrix is ​​reduced in dimension by the principal component analysis method to obtain a compressed resource fingerprint matrix, and the task affinity between each cloud computing task is calculated based on the compressed resource fingerprint matrix, and a task affinity matrix is ​​constructed based on each task affinity; then the system affinity is constructed based on the task affinity matrix, and a greedy task placement strategy is set, and resource constraints are set based on the greedy task placement strategy, and the system affinity is optimized through the resource constraints; then the cloud computing resources of each physical machine are scheduled based on the compressed resource fingerprint matrix and the system affinity, and the cloud computing resources allocated to each physical machine are ensured not to exceed the rated resources based on the proportional allocation method during the scheduling process; then the operating status of each physical machine is monitored, and based on the operating status and task affinity The cloud computing tasks are migrated, and the compressed resource fingerprint matrix and task affinity are dynamically updated; that is, a high-dimensional resource fingerprint matrix is ​​constructed based on statistical features to capture the fine-grained features of cloud computing resource usage, which can accurately characterize the resource usage patterns of cloud computing tasks at different execution stages; the task affinity between each cloud computing task is calculated based on the compressed resource fingerprint matrix to evaluate the compatibility and complementarity between each cloud computing task, which is used to guide the configuration optimization of cloud computing resources and facilitate the subsequent migration of cloud computing tasks; the cloud computing resources of each physical machine are scheduled based on the compressed resource fingerprint matrix and system affinity, while ensuring the quality of cloud computing task execution, the utilization efficiency of cloud computing resources is improved; combined with the running status monitoring of the physical machine, the compressed resource fingerprint matrix and the dynamic update of task affinity, each cloud computing task can be timely allocated to the most suitable physical machine for execution, avoiding cloud computing resource conflicts and performance interference problems, and can effectively balance the utilization efficiency of cloud computing resources and energy costs, and ultimately greatly improve the resource utilization, timeliness, stability and energy efficiency of cloud computing resource scheduling. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] The present invention will be further described below in conjunction with embodiments with reference to the accompanying drawings.

[0059] Figure 1 It is a flow chart of a cloud computing resource scheduling method based on multi-dimensional resource fingerprints of the present invention.

[0060] Figure 2 It is a structural schematic diagram of a cloud computing resource scheduling system based on multi-dimensional resource fingerprints of the present invention. DETAILED DESCRIPTION

[0061] The technical solution in the embodiments of the present application has the following overall ideas: constructing a high-dimensional resource fingerprint matrix based on statistical features to capture the fine-grained features of cloud computing resource usage, which can accurately characterize the resource usage patterns of cloud computing tasks at different execution stages; calculating the task affinity between each cloud computing task based on the compressed resource fingerprint matrix to evaluate the compatibility and complementarity between each cloud computing task, which is used to guide the configuration optimization of cloud computing resources and facilitate the subsequent migration of cloud computing tasks; scheduling the cloud computing resources of each physical machine based on the compressed resource fingerprint matrix and system affinity, while ensuring the quality of cloud computing task execution and improving the utilization efficiency of cloud computing resources; combining the running status monitoring of the physical machine, the compressed resource fingerprint matrix and the dynamic update of the task affinity, so that each cloud computing task can be timely allocated to the most suitable physical machine for execution, avoiding cloud computing resource conflicts and performance interference problems, and can effectively balance the utilization efficiency of cloud computing resources and energy costs, thereby improving the resource utilization, timeliness, stability and energy efficiency of cloud computing resource scheduling.

[0062] Please refer to Figure 1 to Figure 2 As shown, a preferred embodiment of a cloud computing resource scheduling method based on multi-dimensional resource fingerprints of the present invention comprises the following steps:

[0063] Step S1, collecting cloud computing tasks of each physical machine running in the cloud environment, and collecting cloud computing resource usage data of each cloud computing task through a sliding window;

[0064] Step S2: extracting statistical features including at least a mean, a standard deviation, a maximum value, and a minimum value from the cloud computing resource usage data, and constructing a high-dimensional resource fingerprint matrix for each cloud computing task based on the statistical features;

[0065] The high-dimensional resource fingerprint matrix is ​​expressed as: ;in, represents a real number; D represents the dimension of cloud computing resources. For example, cloud computing resources include CPU, memory, disk I / O, and network bandwidth, so the value of D is 4; T represents the number of time steps;

[0066] Step S3, reducing the dimension of the high-dimensional resource fingerprint matrix by principal component analysis (PCA) to obtain a compressed resource fingerprint matrix; reducing the dimension of the high-dimensional resource fingerprint matrix to a compressed resource fingerprint matrix to reduce computational complexity and improve cloud computing resource scheduling efficiency; the compressed resource fingerprint matrix accurately describes the usage characteristics and time dynamics of cloud computing tasks on multi-dimensional cloud computing resources;

[0067] The compressed resource fingerprint matrix is ​​expressed as: ; Wherein, d represents the dimension of cloud computing resources after dimensionality reduction, d<D;

[0068] In the specific implementation, the principal component of 95% variance is retained by the principal component analysis method; that is, the minimum number of principal components d that satisfies the cumulative variance contribution rate of not less than 95% is selected, that is, the minimum d is selected so that the sum of the first d eigenvalues ​​accounts for not less than 95% of the total eigenvalue sum;

[0069] Assume that the compressed resource fingerprint matrix of a cloud computing task has 12 dimensions (3 resource types × 4 statistical features), and the eigenvalues ​​are: [100, 50, 20, 10, 5, 3, 2, 1, 0.5, 0.3, 0.1, 0.1], and the total is 192;

[0070] When taking the first four eigenvalues: (100+50+20+10) / 192≈0.938<0.95;

[0071] When taking the first 5 eigenvalues: (100+50+20+10+5) / 192≈0.964>0.95;

[0072] Therefore, d=5 is selected, which means retaining the first 5 principal components;

[0073] Step S4: calculating the task affinity between each cloud computing task based on the compressed resource fingerprint matrix, and constructing a task affinity matrix based on each task affinity;

[0074] Step S5: constructing system affinity based on the task affinity matrix, setting a greedy task placement strategy, setting a resource constraint condition based on the greedy task placement strategy, and optimizing the system affinity through the resource constraint condition;

[0075] Step S6: scheduling the cloud computing resources of each physical machine based on the compressed resource fingerprint matrix and the system affinity, and ensuring that the cloud computing resources allocated to each physical machine do not exceed the rated resources based on the proportional allocation method during the scheduling process;

[0076] Step S7: monitor the operating status of each physical machine, and migrate the cloud computing task based on the operating status and task affinity;

[0077] Step S8: During the cloud computing resource scheduling process, the compressed resource fingerprint matrix and task affinity are dynamically updated to continuously optimize performance; in specific implementation, the task affinity is also smoothed using the exponential moving average method to balance historical information and new observations:

[0078] ;

[0079] ;

[0080] Among them, β represents the smoothing factor, which is used to control the weight of historical information and new observations; Represents the cloud computing resource usage vector at time t+1.

[0081] The present invention dynamically generates a compressed resource fingerprint matrix, which can accurately characterize the resource usage mode of cloud computing tasks in different execution stages, laying the foundation for fine-grained resource management; reveals the similarity and complementarity between cloud computing tasks through task affinity, and guides the optimal configuration of cloud computing resources; optimizes system affinity through resource constraints, maximizes system affinity while ensuring service quality, and improves resource utilization efficiency; combines the compressed resource fingerprint matrix and the dynamic update mechanism of task affinity, so that the scheduling strategy of cloud computing resources can quickly adapt to changes in workload and maintain long-term optimization effects.

[0082] In step S1, the cloud computing resource usage data includes at least CPU usage, memory usage, disk I / O speed, and network bandwidth usage;

[0083] In step S4, the calculation formula of the task affinity is: ;

[0084] in, represents the i-th compressed resource fingerprint matrix; represents the j-th compressed resource fingerprint matrix; express The corresponding cloud computing tasks and The task affinity of the corresponding cloud computing tasks; α represents the trade-off factor, which can be dynamically adjusted to balance similarity and complementarity; Sim() represents the similarity measurement function; Comp() represents the complementarity measurement function.

[0085] ;

[0086] ;

[0087] in, represents the Pearson correlation coefficient; represents the time series of the k-th dimension of cloud computing resources for cloud computing task i; represents the time series of the k-th dimension of cloud computing resources for cloud computing task j; represents the normalized usage of the k-th dimension of cloud computing resources by cloud computing task i at time t; represents the normalized usage of the k-th dimension cloud computing resources of cloud computing task j at time t.

[0088] The element in the i-th row and j-th column of the task affinity matrix is ​​expressed as: .

[0089] The step S5 is specifically as follows:

[0090] Based on the task affinity matrix, the task affinities between each cloud computing task running in each physical machine are obtained, the task affinities of each physical machine are summed up to obtain the physical machine affinity, and the task affinities of each physical machine are summed up to obtain the system affinity;

[0091] Assume that there are two physical machines m1 and m2. Tasks t1, t2, and t3 are running on m1, and tasks t4 and t5 are running on m2. Then the calculation process of system affinity is:

[0092] S(π)=[A(t1,t2)+A(t1,t3)+A(t2,t3)] / / task pair affinity on m1+[A(t4,t5)] / / task pair affinity on m2;

[0093] where A(ti,tj) comes from the corresponding element in the task affinity matrix A.

[0094] Set a greedy task placement strategy , specifically:

[0095] For each cloud computing task to be scheduled, k , calculate t k The accumulated affinity with the scheduled cloud computing tasks is used to select a physical machine that maximizes the accumulated affinity and satisfies the resource constraint conditions for allocation;

[0096] The greedy task placement strategy is used to make placement decisions for cloud computing tasks to balance exploration and utilization; in specific implementation, experience replay technology can be used to improve learning efficiency and stability;

[0097] A resource constraint condition is set based on the greedy task placement strategy: ;

[0098] Among them, t k represents the kth cloud computing task; r k Indicates t k The resource demand vector of i represents the i-th physical machine; Indicates allocation to m i A collection of cloud computing tasks; c i Indicates m i The resource capacity vector; M represents the total number of physical machines; represents the greedy task placement strategy; Indicates system affinity;

[0099] The system affinity is optimized through the resource constraint conditions.

[0100] In specific implementation, multi-objective optimization problems can be defined, such as considering resource utilization, energy consumption and service quality at the same time.

[0101] The step S6 is specifically as follows:

[0102] Obtaining statistical features based on the compressed resource fingerprint matrix, scheduling the cloud computing resources of each physical machine based on the statistical features, constraining the scheduling process through the system affinity, and ensuring that the cloud computing resources allocated to each physical machine do not exceed the rated resources based on the proportional allocation method;

[0103] In the cloud computing resource scheduling process, the statistical characteristics reflect the resource usage pattern of the cloud computing task; when calculating new cloud computing resource requirements, multiple factors are taken into consideration, such as using the mean of resource usage as a benchmark value, considering the standard deviation to reserve fluctuation space, and referring to historical peak values ​​(maximum values) to avoid resource shortages; if the current usage rate is close to the maximum value, the allocation is appropriately increased; if it is in a low usage state for a long time, consider reducing the allocation; and the importance and priority of the task also need to be considered.

[0104] The proportional allocation method is specifically as follows: when the cloud computing resources required for all cloud computing tasks running on the physical machine exceed the rated resources, a scaling factor is calculated based on the ratio of the required cloud computing resources to the rated resources, and the allocated cloud computing resources are adjusted based on the scaling factor so that the cloud computing resources allocated to the physical machine do not exceed the rated resources.

[0105] Assume that a physical machine has a 100-core CPU, 256GB memory, and 10Gbps network bandwidth. When the original requirements of the three tasks are: Task 1 requires a 60-core CPU, 100GB memory, and 4Gbps bandwidth; Task 2 requires a 50-core CPU, 150GB memory, and 3Gbps bandwidth; Task 3 requires a 40-core CPU, 80GB memory, and 5Gbps bandwidth. Since the total CPU requirement of 150 cores exceeds the machine capacity of 100 cores, a scaling factor of 2 / 3 is calculated, and then the CPU allocation of the three tasks is adjusted to 40 cores, 33.3 cores, and 26.7 cores, respectively. The allocation of memory and network bandwidth also adopts a similar proportional scaling method.

[0106] The advantages of the proportional allocation method are: it can maintain the relative proportion of resource allocation between tasks to ensure fairness of allocation; it supports the collaborative management of multi-dimensional resources; and it can dynamically adapt to changes in resource demand. In actual operation, the system will consider the minimum resource requirements of the task to ensure that the scaled allocation will not be lower than the basic operating requirements of the task, and will also reserve a certain amount of resource margin to cope with load fluctuations. In addition, in order to avoid system shocks, small changes in demand will not immediately trigger reallocation, but will only be adjusted after accumulating to a certain extent.

[0107] The step S7 is specifically as follows:

[0108] The operating status of each physical machine is monitored, and the operating status is overload and underload; the overload and underload are judged based on statistical characteristics, combined with preset thresholds and multi-dimensional resource usage.

[0109] The cloud computing task of the physical machine whose running status is overloaded is migrated to the physical machine whose running status is underloaded and whose task affinity gain is maximized for execution.

[0110] During the specific implementation, system indicators such as resource utilization, task completion time, and SLA breach rate are also monitored, and the trade-off factor α in the task affinity calculation formula and the exploration rate of the greedy task placement strategy are dynamically adjusted, and the corresponding processing mechanism is triggered when an abnormal situation is detected.

[0111] For the migration of cloud computing tasks, the migration strategy is divided into three main steps: task screening, target machine selection and migration execution.

[0112] The first is the task screening stage, in which all cloud computing tasks running on the overloaded physical machine are evaluated, mainly considering two aspects: 1. The task affinity between the cloud computing task and other cloud computing tasks on the current physical machine, and selecting the cloud computing task with the lowest task affinity as the migration candidate; 2. The migration cost of the cloud computing task, including factors such as task scale, running time, and data dependency. Specifically, the "migration value score" of each cloud computing task is calculated: Score(tk) = w1 *(1 - LocalAffinity(tk)) + w2 * (1 - MigrationCost(tk)); where LocalAffinity is the average task affinity between the cloud computing task and other cloud computing tasks on the same physical machine, MigrationCost is the normalized migration cost, and w1 and w2 are weight factors.

[0113] Then comes the selection of the target machine (target physical machine). Select the most suitable target machine from the set of physical machines with lower load. The selection criteria are: 1. Evaluate the resource sufficiency of the target machine to ensure that there will be no new overload after migration; 2. Calculate the affinity gain between the cloud computing task and the existing cloud computing task on the target machine, that is, the improvement in system affinity that may be brought about after migration; 3. Network distance and data locality. Calculate the comprehensive score for each candidate target machine: Score(mi)=α* ResourceAvailability(mi) +β* AffinityGain(tk, mi)+γ*LocalityBenefit(tk,mi).

[0114] During the migration execution phase, a progressive migration strategy is adopted. First, the cloud computing tasks with the highest scores are migrated to the most suitable target machines, and then the system status changes are observed. If the overload situation still exists, the suboptimal cloud computing tasks are selected for migration until the overload situation is alleviated or the preset migration limit is reached. The impact of the migration is continuously monitored throughout the process. If it is found that the migration causes a significant performance degradation, the rollback mechanism will be triggered.

[0115] In order to avoid system instability caused by frequent migration, a protection mechanism is also set up: setting a minimum migration interval, controlling the number of concurrent migrations, recording the migration history of tasks to avoid repeated migrations, etc. In addition, the effectiveness of migration decisions is regularly evaluated, and relevant parameters are optimized through machine learning methods to improve the accuracy and efficiency of migration strategies.

[0116] A preferred embodiment of a cloud computing resource scheduling system based on multi-dimensional resource fingerprints of the present invention includes the following modules:

[0117] A cloud computing task and resource collection module, used to collect cloud computing tasks of each physical machine running in the cloud environment, and collect cloud computing resource usage data of each cloud computing task through a sliding window;

[0118] A high-dimensional resource fingerprint matrix construction module is used to extract statistical features including at least a mean, a standard deviation, a maximum value, and a minimum value from the cloud computing resource usage data, and to construct a high-dimensional resource fingerprint matrix for each cloud computing task based on the statistical features;

[0119] The high-dimensional resource fingerprint matrix is ​​expressed as: ;in, represents a real number; D represents the dimension of cloud computing resources. For example, cloud computing resources include CPU, memory, disk I / O, and network bandwidth, so the value of D is 4; T represents the number of time steps;

[0120] A compressed resource fingerprint matrix construction module is used to reduce the dimension of the high-dimensional resource fingerprint matrix through the principal component analysis method (PCA) to obtain a compressed resource fingerprint matrix; the high-dimensional resource fingerprint matrix is ​​reduced to a compressed resource fingerprint matrix in order to reduce the computational complexity and improve the efficiency of cloud computing resource scheduling; the compressed resource fingerprint matrix accurately describes the usage characteristics and time dynamics of cloud computing tasks on multi-dimensional cloud computing resources;

[0121] The compressed resource fingerprint matrix is ​​expressed as: ; Wherein, d represents the dimension of cloud computing resources after dimensionality reduction, d<D;

[0122] In the specific implementation, the principal component of 95% variance is retained by the principal component analysis method; that is, the minimum number of principal components d that satisfies the cumulative variance contribution rate of not less than 95% is selected, that is, the minimum d is selected so that the sum of the first d eigenvalues ​​accounts for not less than 95% of the total eigenvalue sum;

[0123] Assume that the compressed resource fingerprint matrix of a cloud computing task has 12 dimensions (3 resource types × 4 statistical features), and the eigenvalues ​​are: [100, 50, 20, 10, 5, 3, 2, 1, 0.5, 0.3, 0.1, 0.1], and the total is 192;

[0124] When taking the first four eigenvalues: (100+50+20+10) / 192≈0.938<0.95;

[0125] When taking the first 5 eigenvalues: (100+50+20+10+5) / 192≈0.964>0.95;

[0126] Therefore, d=5 is selected, which means retaining the first 5 principal components;

[0127] A task affinity matrix construction module, used for calculating the task affinity between each cloud computing task based on the compressed resource fingerprint matrix, and constructing a task affinity matrix based on each task affinity;

[0128] A system affinity construction module, used to construct system affinity based on the task affinity matrix, set a greedy task placement strategy, set a resource constraint condition based on the greedy task placement strategy, and optimize the system affinity through the resource constraint condition;

[0129] A cloud computing resource scheduling module, used to schedule the cloud computing resources of each physical machine based on the compressed resource fingerprint matrix and the system affinity, and to ensure that the cloud computing resources allocated to each physical machine do not exceed the rated resources based on the proportional allocation method during the scheduling process;

[0130] A cloud computing task migration module, used to monitor the operating status of each physical machine and migrate the cloud computing tasks based on the operating status and task affinity;

[0131] The dynamic update module is used to dynamically update the compressed resource fingerprint matrix and task affinity during the cloud computing resource scheduling process to continuously optimize performance; in specific implementation, the exponential moving average method is also used to smooth the task affinity to balance historical information and new observations:

[0132] ;

[0133] ;

[0134] in, represents the smoothing factor, which is used to control the weight of historical information and new observations; Represents the cloud computing resource usage vector at time t+1.

[0135] The present invention dynamically generates a compressed resource fingerprint matrix, which can accurately characterize the resource usage mode of cloud computing tasks in different execution stages, laying the foundation for fine-grained resource management; reveals the similarity and complementarity between cloud computing tasks through task affinity, and guides the optimal configuration of cloud computing resources; optimizes system affinity through resource constraints, maximizes system affinity while ensuring service quality, and improves resource utilization efficiency; combines the compressed resource fingerprint matrix and the dynamic update mechanism of task affinity, so that the scheduling strategy of cloud computing resources can quickly adapt to changes in workload and maintain long-term optimization effects.

[0136] In the cloud computing task and resource acquisition module, the cloud computing resource usage data includes at least CPU usage, memory usage, disk I / O speed, and network bandwidth usage;

[0137] In the task affinity matrix construction module, the calculation formula of the task affinity is: ;

[0138] in, represents the i-th compressed resource fingerprint matrix; represents the jth compressed resource fingerprint matrix; express The corresponding cloud computing tasks and The task affinity of the corresponding cloud computing tasks; α represents the trade-off factor, which can be dynamically adjusted to balance similarity and complementarity; Sim() represents the similarity measurement function; Comp() represents the complementarity measurement function.

[0139] ;

[0140] ;

[0141] in, represents the Pearson correlation coefficient; represents the time series of the k-th dimension of cloud computing resources for cloud computing task i; represents the time series of the k-th dimension of cloud computing resources for cloud computing task j; represents the normalized usage of the k-th dimension of cloud computing resources by cloud computing task i at time t; represents the normalized usage of the k-th dimension cloud computing resources of cloud computing task j at time t.

[0142] The element in the i-th row and j-th column of the task affinity matrix is ​​expressed as: .

[0143] The system affinity building module is specifically used for:

[0144] Based on the task affinity matrix, the task affinities between each cloud computing task running in each physical machine are obtained, the task affinities of each physical machine are summed up to obtain the physical machine affinity, and the task affinities of each physical machine are summed up to obtain the system affinity;

[0145] Assume that there are two physical machines m1 and m2. Tasks t1, t2, and t3 are running on m1, and tasks t4 and t5 are running on m2. Then the calculation process of system affinity is:

[0146] S(π)=[A(t1,t2)+A(t1,t3)+A(t2,t3)] / / task pair affinity on m1+[A(t4,t5)] / / task pair affinity on m2;

[0147] where A(ti,tj) comes from the corresponding element in the task affinity matrix A.

[0148] Set a greedy task placement strategy , specifically:

[0149] For each cloud computing task to be scheduled, k , calculate t k The accumulated affinity with the scheduled cloud computing tasks is used to select a physical machine that maximizes the accumulated affinity and satisfies the resource constraint conditions for allocation;

[0150] The greedy task placement strategy is used to make placement decisions for cloud computing tasks to balance exploration and utilization; in specific implementation, experience replay technology can be used to improve learning efficiency and stability;

[0151] A resource constraint condition is set based on the greedy task placement strategy: ;

[0152] Among them, t k represents the kth cloud computing task; r k Indicates t k The resource demand vector of i represents the i-th physical machine; Indicates allocation to m i A collection of cloud computing tasks; c i Indicates m i The resource capacity vector; M represents the total number of physical machines; represents the greedy task placement strategy; Indicates system affinity;

[0153] The system affinity is optimized through the resource constraint conditions.

[0154] In specific implementation, multi-objective optimization problems can be defined, such as considering resource utilization, energy consumption and service quality at the same time.

[0155] The cloud computing resource scheduling module is specifically used for:

[0156] Obtaining statistical features based on the compressed resource fingerprint matrix, scheduling the cloud computing resources of each physical machine based on the statistical features, constraining the scheduling process through the system affinity, and ensuring that the cloud computing resources allocated to each physical machine do not exceed the rated resources based on the proportional allocation method;

[0157] In the cloud computing resource scheduling process, the statistical characteristics reflect the resource usage pattern of the cloud computing task; when calculating new cloud computing resource requirements, multiple factors are taken into consideration, such as using the mean of resource usage as a benchmark value, considering the standard deviation to reserve fluctuation space, and referring to historical peak values ​​(maximum values) to avoid resource shortages; if the current usage rate is close to the maximum value, the allocation is appropriately increased; if it is in a low usage state for a long time, consider reducing the allocation; and the importance and priority of the task also need to be considered.

[0158] The proportional allocation method is specifically as follows: when the cloud computing resources required for all cloud computing tasks running on the physical machine exceed the rated resources, a scaling factor is calculated based on the ratio of the required cloud computing resources to the rated resources, and the allocated cloud computing resources are adjusted based on the scaling factor so that the cloud computing resources allocated to the physical machine do not exceed the rated resources.

[0159] Assume that a physical machine has a 100-core CPU, 256GB memory, and 10Gbps network bandwidth. When the original requirements of the three tasks are: Task 1 requires a 60-core CPU, 100GB memory, and 4Gbps bandwidth; Task 2 requires a 50-core CPU, 150GB memory, and 3Gbps bandwidth; Task 3 requires a 40-core CPU, 80GB memory, and 5Gbps bandwidth. Since the total CPU requirement of 150 cores exceeds the machine capacity of 100 cores, a scaling factor of 2 / 3 is calculated, and then the CPU allocation of the three tasks is adjusted to 40 cores, 33.3 cores, and 26.7 cores, respectively. The allocation of memory and network bandwidth also adopts a similar proportional scaling method.

[0160] The advantages of the proportional allocation method are: it can maintain the relative proportion of resource allocation between tasks to ensure fairness of allocation; it supports the collaborative management of multi-dimensional resources; and it can dynamically adapt to changes in resource demand. In actual operation, the system will consider the minimum resource requirements of the task to ensure that the scaled allocation will not be lower than the basic operating requirements of the task, and will also reserve a certain amount of resource margin to cope with load fluctuations. In addition, in order to avoid system shocks, small changes in demand will not immediately trigger reallocation, but will only be adjusted after accumulating to a certain extent.

[0161] The cloud computing task migration module is specifically used for:

[0162] The operating status of each physical machine is monitored, and the operating status is overload and underload; the overload and underload are judged based on statistical characteristics, combined with preset thresholds and multi-dimensional resource usage.

[0163] The cloud computing task of the physical machine whose running status is overloaded is migrated to the physical machine whose running status is underloaded and whose task affinity gain is maximized for execution.

[0164] During the specific implementation, system indicators such as resource utilization, task completion time, and SLA breach rate are also monitored, and the trade-off factor α in the task affinity calculation formula and the exploration rate of the greedy task placement strategy are dynamically adjusted, and the corresponding processing mechanism is triggered when an abnormal situation is detected.

[0165] For the migration of cloud computing tasks, the migration strategy is divided into three main steps: task screening, target machine selection and migration execution.

[0166] The first is the task screening stage, in which all cloud computing tasks running on the overloaded physical machine are evaluated, mainly considering two aspects: 1. The task affinity between the cloud computing task and other cloud computing tasks on the current physical machine, and selecting the cloud computing task with the lowest task affinity as the migration candidate; 2. The migration cost of the cloud computing task, including factors such as task scale, running time, and data dependency. Specifically, the "migration value score" of each cloud computing task is calculated: Score(tk) = w1 *(1 - LocalAffinity(tk)) + w2 * (1 - MigrationCost(tk)); where LocalAffinity is the average task affinity between the cloud computing task and other cloud computing tasks on the same physical machine, MigrationCost is the normalized migration cost, and w1 and w2 are weight factors.

[0167] Then comes the selection of the target machine (target physical machine). Select the most suitable target machine from the set of physical machines with lower load. The selection criteria are: 1. Evaluate the resource sufficiency of the target machine to ensure that there will be no new overload after migration; 2. Calculate the affinity gain between the cloud computing task and the existing cloud computing task on the target machine, that is, the improvement in system affinity that may be brought about after migration; 3. Network distance and data locality. Calculate the comprehensive score for each candidate target machine: Score(mi)=α* ResourceAvailability(mi) +β* AffinityGain(tk, mi)+γ*LocalityBenefit(tk,mi).

[0168] During the migration execution phase, a progressive migration strategy is adopted. First, the cloud computing tasks with the highest scores are migrated to the most suitable target machines, and then the system status changes are observed. If the overload situation still exists, the suboptimal cloud computing tasks are selected for migration until the overload situation is alleviated or the preset migration limit is reached. The impact of the migration is continuously monitored throughout the process. If it is found that the migration causes a significant performance degradation, the rollback mechanism will be triggered.

[0169] In order to avoid system instability caused by frequent migration, a protection mechanism is also set up: setting a minimum migration interval, controlling the number of concurrent migrations, recording the migration history of tasks to avoid repeated migrations, etc. In addition, the effectiveness of migration decisions is regularly evaluated, and relevant parameters are optimized through machine learning methods to improve the accuracy and efficiency of migration strategies.

[0170] In summary, the advantages of the present invention are:

[0171] By collecting cloud computing tasks of each physical machine running in the cloud environment, collecting cloud computing resource usage data of each cloud computing task through a sliding window, and extracting statistical features including at least mean, standard deviation, maximum value and minimum value from the cloud computing resource usage data, a high-dimensional resource fingerprint matrix of each cloud computing task is constructed based on each statistical feature, and then the high-dimensional resource fingerprint matrix is ​​reduced in dimension by the principal component analysis method to obtain a compressed resource fingerprint matrix, and the task affinity between each cloud computing task is calculated based on the compressed resource fingerprint matrix, and a task affinity matrix is ​​constructed based on each task affinity; then the system affinity is constructed based on the task affinity matrix, and a greedy task placement strategy is set, and resource constraints are set based on the greedy task placement strategy, and the system affinity is optimized through the resource constraints; then the cloud computing resources of each physical machine are scheduled based on the compressed resource fingerprint matrix and the system affinity, and the cloud computing resources allocated to each physical machine are ensured not to exceed the rated resources based on the proportional allocation method during the scheduling process; then the operating status of each physical machine is monitored, and based on the operating status and task affinity The cloud computing tasks are migrated, and the compressed resource fingerprint matrix and task affinity are dynamically updated; that is, a high-dimensional resource fingerprint matrix is ​​constructed based on statistical features to capture the fine-grained features of cloud computing resource usage, which can accurately characterize the resource usage patterns of cloud computing tasks at different execution stages; the task affinity between each cloud computing task is calculated based on the compressed resource fingerprint matrix to evaluate the compatibility and complementarity between each cloud computing task, which is used to guide the configuration optimization of cloud computing resources and facilitate the subsequent migration of cloud computing tasks; the cloud computing resources of each physical machine are scheduled based on the compressed resource fingerprint matrix and system affinity, while ensuring the quality of cloud computing task execution, the utilization efficiency of cloud computing resources is improved; combined with the running status monitoring of the physical machine, the compressed resource fingerprint matrix and the dynamic update of task affinity, each cloud computing task can be timely allocated to the most suitable physical machine for execution, avoiding cloud computing resource conflicts and performance interference problems, and can effectively balance the utilization efficiency of cloud computing resources and energy costs, and ultimately greatly improve the resource utilization, timeliness, stability and energy efficiency of cloud computing resource scheduling.

[0172] Although the specific implementation modes of the present invention are described above, those skilled in the art should understand that the specific implementation modes described are only illustrative and are not intended to limit the scope of the present invention. Equivalent modifications and changes made by those skilled in the art in accordance with the spirit of the present invention should be included in the scope of protection of the claims of the present invention.

Claims

1. A cloud computing resource scheduling method based on multi-dimensional resource fingerprints, characterized in that: The steps include: Step S1, collecting cloud computing tasks of each physical machine running in the cloud environment, and collecting cloud computing resource usage data of each cloud computing task through a sliding window; Step S2: extracting statistical features including at least a mean, a standard deviation, a maximum value, and a minimum value from the cloud computing resource usage data, and constructing a high-dimensional resource fingerprint matrix for each cloud computing task based on the statistical features; Step S3, performing a dimensionality reduction operation on the high-dimensional resource fingerprint matrix by a principal component analysis method to obtain a compressed resource fingerprint matrix; Step S4: calculating the task affinity between each cloud computing task based on the compressed resource fingerprint matrix, and constructing a task affinity matrix based on each task affinity; Step S5: constructing system affinity based on the task affinity matrix, setting a greedy task placement strategy, setting a resource constraint condition based on the greedy task placement strategy, and optimizing the system affinity through the resource constraint condition; Step S6: scheduling the cloud computing resources of each physical machine based on the compressed resource fingerprint matrix and the system affinity, and ensuring that the cloud computing resources allocated to each physical machine do not exceed the rated resources based on the proportional allocation method during the scheduling process; Step S7: monitor the operating status of each physical machine, and migrate the cloud computing task based on the operating status and task affinity; Step S8: During the cloud computing resource scheduling process, the compressed resource fingerprint matrix and task affinity are dynamically updated; In step S1, the cloud computing resource usage data includes at least CPU usage, memory usage, disk I / O speed, and network bandwidth usage; In step S4, the calculation formula of the task affinity is: ; in, represents the i-th compressed resource fingerprint matrix; represents the j-th compressed resource fingerprint matrix; express The corresponding cloud computing tasks and The task affinity of the corresponding cloud computing task; represents the trade-off factor; represents the similarity measure function; represents the complementarity metric function; The step S5 is specifically as follows: Based on the task affinity matrix, task affinities between each cloud computing task running in each physical machine are obtained, the task affinities of each physical machine are summed up to obtain the physical machine affinity, and the affinities of each physical machine are summed up to obtain the system affinity; Set a greedy task placement strategy , specifically: For each cloud computing task to be scheduled ,calculate The accumulated affinity with the scheduled cloud computing tasks is used to select a physical machine that maximizes the accumulated affinity and satisfies the resource constraint conditions for allocation; A resource constraint condition is set based on the greedy task placement strategy: ; in, represents the kth cloud computing task; express The resource demand vector of represents the i-th physical machine; Indicates allocation to A collection of cloud computing tasks; express The resource capacity vector; M represents the total number of physical machines; represents the greedy task placement strategy; Indicates system affinity; The system affinity is optimized through the resource constraint conditions.

2. A cloud computing resource scheduling method based on multi-dimensional resource fingerprints as claimed in claim 1, characterized in that: The step S6 is specifically as follows: Obtaining statistical features based on the compressed resource fingerprint matrix, scheduling the cloud computing resources of each physical machine based on the statistical features, constraining the scheduling process through the system affinity, and ensuring that the cloud computing resources allocated to each physical machine do not exceed the rated resources based on the proportional allocation method; The proportional allocation method is specifically as follows: when the cloud computing resources required for all cloud computing tasks running on the physical machine exceed the rated resources, a scaling factor is calculated based on the ratio of the required cloud computing resources to the rated resources, and the allocated cloud computing resources are adjusted based on the scaling factor so that the cloud computing resources allocated to the physical machine do not exceed the rated resources.

3. The cloud computing resource scheduling method based on multi-dimensional resource fingerprint according to claim 1, characterized in that: The step S7 is specifically as follows: Monitoring the operating status of each physical machine, wherein the operating status is overload and underload; The cloud computing task of the physical machine whose running status is overloaded is migrated to the physical machine whose running status is underloaded and whose task affinity gain is maximized for execution.

4. A cloud computing resource scheduling system based on multi-dimensional resource fingerprints, characterized in that: Includes the following modules: A cloud computing task and resource collection module, used to collect cloud computing tasks of each physical machine running in the cloud environment, and collect cloud computing resource usage data of each cloud computing task through a sliding window; A high-dimensional resource fingerprint matrix construction module is used to extract statistical features including at least a mean, a standard deviation, a maximum value, and a minimum value from the cloud computing resource usage data, and to construct a high-dimensional resource fingerprint matrix for each cloud computing task based on the statistical features; A compressed resource fingerprint matrix construction module is used to perform a dimensionality reduction operation on the high-dimensional resource fingerprint matrix through a principal component analysis method to obtain a compressed resource fingerprint matrix; A task affinity matrix construction module, used for calculating the task affinity between each cloud computing task based on the compressed resource fingerprint matrix, and constructing a task affinity matrix based on each task affinity; A system affinity construction module, used to construct system affinity based on the task affinity matrix, set a greedy task placement strategy, set a resource constraint condition based on the greedy task placement strategy, and optimize the system affinity through the resource constraint condition; A cloud computing resource scheduling module, used to schedule the cloud computing resources of each physical machine based on the compressed resource fingerprint matrix and the system affinity, and to ensure that the cloud computing resources allocated to each physical machine do not exceed the rated resources based on the proportional allocation method during the scheduling process; A cloud computing task migration module, used to monitor the operating status of each physical machine and migrate the cloud computing tasks based on the operating status and task affinity; A dynamic update module, used for dynamically updating the compressed resource fingerprint matrix and task affinity during cloud computing resource scheduling; In the cloud computing task and resource acquisition module, the cloud computing resource usage data includes at least CPU usage, memory usage, disk I / O speed, and network bandwidth usage; In the task affinity matrix construction module, the calculation formula of the task affinity is: ; in, represents the i-th compressed resource fingerprint matrix; represents the j-th compressed resource fingerprint matrix; express The corresponding cloud computing tasks and The task affinity of the corresponding cloud computing task; represents the trade-off factor; represents the similarity measure function; represents the complementarity metric function; The system affinity building module is specifically used for: Based on the task affinity matrix, task affinities between each cloud computing task running in each physical machine are obtained, the task affinities of each physical machine are summed up to obtain the physical machine affinity, and the affinities of each physical machine are summed up to obtain the system affinity; Set a greedy task placement strategy , specifically: For each cloud computing task to be scheduled ,calculate The accumulated affinity with the scheduled cloud computing tasks is used to select a physical machine that maximizes the accumulated affinity and satisfies the resource constraint conditions for allocation; A resource constraint condition is set based on the greedy task placement strategy: ; in, represents the kth cloud computing task; express The resource demand vector of represents the i-th physical machine; Indicates allocation to A collection of cloud computing tasks; express The resource capacity vector of ; M represents the total number of physical machines; represents the greedy task placement strategy; Indicates system affinity; The system affinity is optimized through the resource constraint conditions.

5. A cloud computing resource scheduling system based on multi-dimensional resource fingerprints as claimed in claim 4, characterized in that: The cloud computing resource scheduling module is specifically used for: Obtaining statistical features based on the compressed resource fingerprint matrix, scheduling the cloud computing resources of each physical machine based on the statistical features, constraining the scheduling process through the system affinity, and ensuring that the cloud computing resources allocated to each physical machine do not exceed the rated resources based on the proportional allocation method; The proportional allocation method is specifically as follows: when the cloud computing resources required for all cloud computing tasks running on the physical machine exceed the rated resources, a scaling factor is calculated based on the ratio of the required cloud computing resources to the rated resources, and the allocated cloud computing resources are adjusted based on the scaling factor so that the cloud computing resources allocated to the physical machine do not exceed the rated resources.

6. A cloud computing resource scheduling system based on multi-dimensional resource fingerprints as claimed in claim 4, characterized in that: The cloud computing task migration module is specifically used for: Monitoring the operating status of each physical machine, wherein the operating status is overload and underload; The cloud computing task of the physical machine whose running status is overloaded is migrated to the physical machine whose running status is underloaded and whose task affinity gain is maximized for execution.

Images