Energy-saving method and system based on multi-dimensional resource intelligent scheduling in cloud computing system

Abstract

The invention relates to an energy-saving method and system based on multi-dimensional resource intelligent scheduling in a cloud computing system, and belongs to the technical field of cloud computing. The data acquisition module acquires data of a host CPU, a memory, a network and storage resources in real time; calculating a load mean value based on a dynamic sliding window, wherein the window size is adaptively adjusted along with resource fluctuation; carrying out weighting calculation on a comprehensive load through an information entropy algorithm, and predicting a future load in combination with an ARIMA model; classifying the host as overload, normal or low load according to the current and predicted load; selecting the virtual machine with the maximum descending gradient of the resource utilization rate for the overloaded host to migrate, and closing the low-load host after migrating all the virtual machines; and matching resource complementary target hosts based on the Pearson's correlation coefficient, and verifying and triggering secondary migration after migration. Multi-dimensional resource collaborative optimization is realized through the monitoring and scheduling module, invalid migration is reduced by utilizing dynamic load prediction, the utilization rate is improved by utilizing resource complementary matching, and the service continuity is ensured while the energy consumption is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of cloud computing, and particularly relates to an energy-saving method and system based on multi-dimensional resource intelligent scheduling in a cloud computing system. Background Art

[0002] At present, a major problem faced by the data center of a cloud computing system is the problem of energy waste. Moreover, with the rapid growth of cloud computing applications, the scale of the data center is becoming larger and larger, and the power consumption of a large number of hosts is increasing day by day. Research shows that the average CPU utilization rate of physical hosts in a cloud data center is only 15% - 20%. A large number of servers are in an idle or no-load state for a long time, and the energy consumption of no-load servers is as high as 70% of that at full load, resulting in significant energy waste.

[0003] The prior art dynamically adjusts resource allocation through virtual machine migration, but its core defect lies in that it only takes the CPU utilization rate as a single optimization goal and ignores the collaborative influence of memory, network, and storage resources. For example, when the memory resources are exhausted, even if the CPU utilization rate is lower than the threshold, service degradation will still be triggered; network bandwidth congestion leads to a sharp increase in the time-consuming of virtual machine migration, further exacerbating resource fragmentation. This one-dimensional optimization strategy not only fails to achieve the global optimal energy efficiency but also causes frequent violations of the service level agreement (SLAv) due to the imbalance of multi-dimensional resources.

[0004] Furthermore, the prior art lacks the adaptability to dynamic loads and the ability of forward-looking scheduling. The sudden fluctuations in the load of virtual machines (such as the update of AI training task parameters) result in highly unstable resource requirements, and the load evaluation mechanism based on instantaneous sampling is difficult to respond in a timely manner, frequently triggering ineffective migrations. In addition, the migration strategy does not consider the multi-dimensional resource complementary characteristics of virtual machines and hosts (such as deploying virtual machines with high memory requirements to hosts with low memory margins), resulting in the resource utilization rate still being lower than expected after migration. For example, network-intensive virtual machines are allocated to hosts with limited bandwidth, leading to a sharp increase in the demand for secondary migrations, forming a vicious cycle of "migration - overload - re-migration". The blindness, radicalness, and lag of the prior art make the resource scheduling efficiency low, which can neither effectively reduce energy consumption nor ensure service stability. Summary of the Invention

[0005] The technical problems to be solved by the present invention are: to overcome the defects such as unbalanced single-dimensional optimization, poor dynamic load adaptability, and blindness in resource matching existing in the resource scheduling of existing cloud computing systems, and to provide an energy-saving method and system based on multi-dimensional resource intelligent scheduling, which realizes the accurate perception of multi-dimensional resource fluctuations and the prediction of future trends through dynamic sliding window mean calculation and ARIMA load prediction model; achieves the global resource optimization of virtual machine migration decision-making through information entropy algorithm dynamic weighting and Pearson correlation coefficient complementary matching; and ensures business continuity while significantly reducing the energy consumption of the data center through KVM virtualization non-interruptible migration and intelligent shutdown of low-load hosts.

[0006] The energy-saving method based on multi-dimensional resource intelligent scheduling in this cloud computing system includes the following steps:

[0007] S1 - Real-time collect the CPU, memory, network, and storage resource utilization data of host nodes;

[0008] S2 - Calculate the load mean of each dimension of resources based on a dynamic sliding window, and the size of the sliding window is dynamically adjusted according to the resource load fluctuation;

[0009] S3 - Determine the weights of each dimension of resources through the information entropy algorithm, and calculate the comprehensive load of the host by weighting;

[0010] S4 - Predict the comprehensive load value of the host at a future moment based on the ARIMA model;

[0011] S5 - Classify the host as overloaded, normal load, or low load type according to the current comprehensive load and the predicted value;

[0012] S6 - For overloaded hosts, determine the virtual machines to be migrated based on the virtual machine migration selection model;

[0013] S7 - For low-load hosts, migrate all their virtual machines to normal-load hosts and shut down the low-load hosts;

[0014] S8 - Select a target host for the virtual machine to be migrated based on the resource complementarity, complete the migration and verify the resource status; after migration, verify whether the comprehensive load and the predicted value of the target host exceed the upper limit threshold, and if so, trigger a secondary migration;

[0015] S9 - Shut down the low-load hosts after the migration is completed, and update the resource distribution status of the data center.

[0016] Further, the adjustment method of the dynamic sliding window size in step S2 is to dynamically expand or contract the window duration according to the absolute difference between the maximum fluctuation value of the resource utilization rate and the mean value at the previous moment.

[0017] Further, the information entropy algorithm in step S3 includes

[0018] - Construct a normalized matrix of the mean resource load for each dimension;

[0019] - Calculate the entropy value and contribution degree of each dimension of resources;

[0020] - Allocate weight coefficients according to the contribution degree.

[0021] Furthermore, the establishment of the ARIMA model described in step S4 includes,

[0022] - Conduct a stationarity test and differencing process on the comprehensive load time series;

[0023] - Determine the model order according to the autocorrelation function and partial autocorrelation function;

[0024] - Select the optimal prediction model through the Akaike information criterion.

[0025] Furthermore, the classification criteria for overloaded and underloaded types in the host described in step S5 are,

[0026] - The overloaded type is that both the current comprehensive load and the predicted value exceed the preset upper threshold;

[0027] - The underloaded type is that both the current comprehensive load and the predicted value are lower than the preset lower threshold.

[0028] Furthermore, the migration virtual machine selection model described in step S6 is to preferentially select the virtual machine that maximizes the gradient of the utilization rate reduction of each dimension of resources of the overloaded host.

[0029] Furthermore, the resource complementarity described in step S8 is calculated through the Pearson correlation coefficient, and the host that best matches the resource requirements of the virtual machine to be migrated is selected.

[0030] The energy-saving system based on multi-dimensional resource intelligent scheduling in this cloud computing system is used to implement the energy-saving method based on multi-dimensional resource intelligent scheduling in the above cloud computing system. The system includes the following functional modules,

[0031] - Local monitor: Set on each host node, collect CPU, memory, network, and storage resource utilization data through the agent inside the virtual machine;

[0032] - Global monitor: Set in the cloud management platform, receive and summarize the monitoring information from the local monitor;

[0033] - Load mean module: Set in the intelligent management scheduler, calculate the load mean of each dimension of resources based on a dynamic sliding window;

[0034] - Comprehensive load calculation module: Set in the intelligent management scheduler, calculate the comprehensive load of the host by weighted calculation through the information entropy algorithm;

[0035] - Comprehensive load prediction module: Set in the intelligent management scheduler, generating load prediction values for future moments based on the ARIMA model;

[0036] - Migrated virtual machine selection module: Set in the intelligent scheduler, determining the virtual machines to be migrated on overloaded hosts according to the migration model;

[0037] - Virtual machine placement module: Set in the intelligent scheduler, matching target hosts based on resource complementarity and performing migrations;

[0038] - Resource management module: Set in the cloud management platform, shutting down underloaded hosts and updating resource status.

[0039] Furthermore, the virtual machine placement module integrates the KVM virtualization interface and completes virtual machine migration through the live migration technology.

[0040] An energy-saving method and system for multi-dimensional resource intelligent scheduling in a cloud computing system according to the present invention overcome the defects such as unbalanced single-dimensional optimization, poor dynamic load adaptability, and blindness in resource matching existing in the resource scheduling of existing cloud computing systems. Through dynamic sliding window load mean calculation, ARIMA prediction model, and resource complementarity matching mechanism, the following beneficial effects are achieved:

[0041] (1) Multi-dimensional resource collaborative optimization: Synchronously monitoring CPU, memory, network, and storage resources, improving the accuracy of comprehensive load assessment and significantly reducing the incidence of service violations (SLAv);

[0042] (2) Dynamic load precise adaptation: Based on the sliding window mean and load prediction model, effectively coping with sudden load fluctuations and reducing the number of ineffective migrations;

[0043] (3) Resource complementary and efficient matching: Selecting the optimal target host through the Pearson correlation coefficient, reducing the need for secondary migrations and maximizing resource utilization;

[0044] (4) Dual optimization of energy efficiency and service: Shutting down underloaded hosts to reduce energy consumption, and predictive migration ensuring seamless scheduling of services, achieving the collaborative optimization of energy consumption reduction and service quality improvement in the data center. Description of the Drawings

[0045] The following further describes an energy-saving method and system for multi-dimensional resource intelligent scheduling in a cloud computing system according to the present invention with reference to the drawings:

[0046] Figure 1 is the overall flowchart block diagram of the energy-saving method for multi-dimensional resource intelligent scheduling in this cloud computing system;

[0047] Figure 2 is the system architecture diagram of the energy-saving system for multi-dimensional resource intelligent scheduling in this cloud computing system;

[0048] Figure 3 It is a schematic diagram of a sliding window in the time series in Embodiment 1 of the energy-saving system based on multi-dimensional resource intelligent scheduling in this cloud computing system. Specific implementation manner

[0049] The following uses specific embodiments to further describe the technical solution of the present invention, but the protection scope of the present invention is not limited to the following embodiments.

[0050] Embodiment 1: As Figure 1 shown, the energy-saving method based on multi-dimensional resource intelligent scheduling in this cloud computing system includes the following steps:

[0051] The specific steps of the energy-saving method based on multi-dimensional resource intelligent scheduling in this cloud computing system are as follows:

[0052] S1 Data collection - Through the local monitor (including the Agent inside the virtual machine) deployed on the host node, the utilization rate data of CPU, memory, network, and storage resources are collected in real time. Specifically, in the data center, the host is the carrier for hosting virtual machines. One host can place multiple virtual machines, but one virtual machine can only belong to one host. Assume that V(h i ) represents the set of virtual machines on host h i ; R = {cpu, mem, net, disk} represents the set of cpu, memory (mem), bandwidth (net), and storage (disk) resources, and r j ∈ R represents each dimension of resources; represents the actual utilization rate of virtual machine v i for resource r j at time t; is the capacity of the configured resource r i for host h j ; represents the actual utilization rate of physical host h i for resource r j at time t, which can be calculated from the relevant resources of the virtual machines deployed on this host, as shown in the following formula,

[0053]

[0054] S2 Dynamic sliding window calculation - Based on the dynamic sliding window, calculate the load average value of each dimension of resources, and the size of the sliding window is dynamically adjusted according to the resource load fluctuation. Specifically, in order to objectively and accurately calculate the load of the host and prevent the influence caused by occasional load fluctuations, the present invention uses the average value within the recent sliding time window T at time t as the actual load of the host at time t, and the size of the sliding window can be intelligently adjusted according to the size of the load fluctuation, as shown in the following formula,

[0055]

[0056] Among them, represents the resource r of host h at time t i within the sliding window T j load mean, as shown in Equation (2). This mean is calculated based on multiple discrete samplings within the sliding window T, with a sampling frequency of once per second, so the number of samplings is T. As Figure 2 shown, the sliding window is a time period on the time series at a certain moment point. The size of the sliding window T can be dynamically adjusted according to the maximum value of the resource r j sampled in the most recent T time and the load mean at the previous moment t - 1 absolute difference, as shown in Equation (3). The greater the difference between the two, the larger the time window will be, so the number of samples collected will be more, and then the ability to resist load fluctuations will be stronger. Among them, T min is the minimum value of the time window, and its size can be set artificially. As Figure 3 shown.

[0057] S3 Comprehensive load calculation - Determine the weights of each dimension of resources through the information entropy algorithm and calculate the comprehensive load of the host by weighting. Specifically, for host h i comprehensive load L i is obtained by weighting according to the load means of each dimension of resources as shown in the following formula,

[0058]

[0059] Among them, ω j is the weight coefficient of the load mean of each dimension of resources . Since the usage of various resources of different hosts is different, the higher the utilization rate of a certain resource, the greater the impact on the comprehensive load of the host, and its weight ω j should be higher. The weight coefficient ω j is solved based on the information entropy algorithm.

[0060] The information entropy algorithm determines the weight based on the variation degree of the index. The greater the variation degree, the greater the role played by the index, then the smaller the information entropy, and the greater the weight. Conversely, the smaller the variation degree of the index, the smaller the role played, then the greater the information entropy, and the smaller the weight. ω j The specific steps are as follows,

[0061] (1) Calculate the decision matrix U according to the following formula. Each column of the matrix records the mean load of the relevant resource r j where n is the number of samples.

[0062]

[0063] (2) Normalize the matrix U to obtain the matrix R, as shown in the following formula:

[0064]

[0065] Wherein,

[0066] (3) Calculate the entropy value E j is within the range of 0 and 1.

[0067] (4) Calculate the contribution degree d j = 1 - E j , then the weight ω j can be obtained by dividing the current contribution degree d j by the sum of all contribution degrees, as shown in the following formula:

[0068]

[0069] That is, the comprehensive load L of the host is calculated through formula (4) i , and the weight ω j in formula (4) is determined by the information entropy algorithm of formulas (5)-(7).

[0070] S4 Load prediction - Predict the comprehensive load value of the host at the future moment based on the ARIMA model. Specifically, due to the instability and high variability of cloud resource usage and workload, it is necessary to fully predict this change trend during resource allocation management to avoid excessive resource scheduling. This method predicts the comprehensive load based on the Autoregressive Integrated Moving Average (ARIMA) model. ARIMA combines the Auto Regressive Model (AR), Moving Average Model (MA), Auto Regressive and Moving Average Model (ARMA), and the differencing method. The basic idea is to regard the data sequence formed by the prediction object over time as a random sequence, make the non-stationary random sequence stationary through multiple differencings, and then approximately describe this sequence with a certain mathematical model. Once this model is identified, it can predict future values from the past values of the time series. When applied to the present invention, the future value of L i can be predicted through the historical data of the comprehensive load L i at the next moment.

[0071] Let \(p\) represent the order of the autoregressive AR, \(d\) represent the order of differencing required for the data, and \(q\) represent the order of the moving average MA. The prediction equation of the ARIMA(p,d,q) model with \(p\), \(d\), and \(q\) as parameters can be expressed as the following formula.

[0072]

[0073] Among them, \(y\) t is the sample value of the time series, that is, the time series of the comprehensive load \(L\) i . and \(\theta\) i (\(i = 1,2,\cdots,p\)) are model parameters, and \(\varepsilon\) t is white noise with a normal distribution.

[0074] Then the predicted value \(L_P\) of the comprehensive load at the next moment i can be calculated through Equation (8). The specific process of solving the ARIMA model is as follows:

[0075] (1) Stationarity processing: Determine the stationarity of the time series \(y\) t according to the autocorrelation function and the partial autocorrelation function. If it is a stationary sequence, the autocorrelation function will quickly decay to zero. If it is a non-stationary sequence, the \(d\)-order differencing method is used for stationarity processing. The autocorrelation function \(\rho\) k and the partial autocorrelation function \(\alpha\) k+1,k+1 are shown as the following formula

[0076]

[0077] (2) Model identification: Determine the model and its order by analyzing the truncation or tailing of the autocorrelation function and the partial autocorrelation function of the time series. Truncation means that the autocorrelation function or the partial autocorrelation function of the time series is 0 after a certain order, and tailing means that the autocorrelation function or the partial autocorrelation function is not 0 after a certain order. If the autocorrelation function tails and the partial autocorrelation function truncates, use the AR model. If the autocorrelation function truncates and the partial autocorrelation function tails, use the MA model. If both the autocorrelation function and the partial autocorrelation function tail, use the ARMA model. \(p\) and \(q\) are determined by the order of the ending point or the tailing point, as shown in the following table.

[0078] Selection model Autocorrelation function Partial autocorrelation function AR(p) Trailing Truncated at order p MA(q) Truncated at order q Trailing ARMA(p, q) Trailing at order p Trailing at order q

[0079] (3) Parameter solution: When the following formula (11) is minimized, the values of and \(\theta\) i (\(i = 1,2,\cdots,p\)) are estimated by the least squares method.

[0080]

[0081] (4) Parameter Estimation: There can be multiple model selections for ARIMA prediction. However, to select the best model, the optimal model is obtained through the Akaike Information Criterion, which is defined as shown in the following formula.

[0082] A = 2n - ln(L) (12)

[0083] Where n is the number of parameters of the model, and L is the maximum likelihood function of the model. Select the model with the smallest A.

[0084] The host comprehensive load L is solved through the above process. i and the predicted load LP i . Assume that the upper limit value of the host comprehensive load L i is Thr max , and the lower limit value is Thr min . If L i > Thr max and LP i > Thr max , then add this host to the list of overloaded types; conversely, if L i < Thr min and LP i < Thr min , then add this host to the list of underloaded types; the remaining hosts are added to the list of normal load types.

[0085] That is, use the ARIMA(p, d, q) model (Equation 8) to predict the future comprehensive load LP i , and the model parameters are determined by Equations (9)-(12).

[0086] S5 Host Classification - According to the current comprehensive load and the predicted value, classify the host as overloaded, normal load, or underloaded type. Specifically, the host comprehensive load L i and the predicted load LP i are solved through the above process. Assume that the upper limit value of the host comprehensive load L i is Thr max , and the lower limit value is Thr min . If L i > Thr max and LP i > Thr max , then add this host to the list of overloaded types; conversely, if L i < Thr min and LP i < Thr min , then add this host to the list of underloaded types; the remaining hosts are added to the list of normal load types. That is, if L i > Thr max and LPi >Thr max ,, marked as overloaded; if L i <Thr min and LP i <Thr min , marked as underloaded.

[0087] S6 - S8 Virtual Machine Migration Decision - For overloaded hosts, determine the virtual machines to be migrated based on the virtual machine migration selection model; for underloaded hosts, migrate all their virtual machines to hosts with normal load and shut down the underloaded hosts; select target hosts for the virtual machines to be migrated based on the resource complementarity degree, complete the migration and verify the resource status; after migration, verify whether the comprehensive load and predicted value of the target host exceed the upper threshold, and if so, trigger a secondary migration. Specifically, for underloaded hosts, migrate all the virtual machines on them to hosts with normal load and shut them down to save energy. For overloaded hosts, migrate some virtual machines to hosts with normal load to restore them to normal load to ensure their performance. To quickly and accurately eliminate the overloaded operation of overloaded hosts and restore them to normal load, the present invention designs a model for selecting virtual machines to be migrated, as shown in the following formula,

[0088]

[0089] where, is the utilization rate of resource r i for overloaded host h j , is the utilization rate of resource r i after migrating virtual machine v i from overloaded host h j , is the descending gradient of the comprehensive load of host h(v i ) after migrating virtual machine v i , and the greater the descending gradient, the greater the probability that v i is selected as the virtual machine to be migrated.

[0090] From the above analysis, it can be seen that the list of virtual machines to be migrated is all the virtual machines on underloaded hosts and some virtual machines on overloaded hosts. Next, re - select new host machines for the virtual machines to be migrated to achieve dynamic allocation of resources. Realizing the balanced and full utilization of the overall resources of the data center to obtain the minimum number of host machines in operation and achieving the purpose of energy conservation is the main goal to be achieved by the present invention. For this reason, when the present invention allocates resources, it takes virtual machines as the resource scheduling unit and selects the host machine with the greatest degree of resource complementarity in each dimension with the virtual machines to be migrated as the new host machine, so as to ensure the full utilization of resources. The complementarity degree evaluation model calculates the correlation between each dimension of resources of virtual machines and the remaining resources of each dimension of host machines based on the Pearson correlation coefficient, as shown in the following formula,

[0091]

[0092] wherein represents the remaining amount of the host resource r i remaining amount, represents the correlation between the resources of each dimension of the virtual machine v i and the remaining resources of each dimension of the host h i The greater the value, the greater the degree of complementarity, and the more fully the resources can be utilized. The full utilization of resources can thus ensure the minimum number of operating hosts and the minimum energy consumption.

[0093] In summary, that is, the overloaded host selects the virtual machine v with the largest migration gradient according to Equation (13) i ; calculate the degree of complementarity based on the Pearson correlation coefficient in Equation (14), and select the target host h(v i ); after migration, check the load of the target host. If it is overloaded, trigger a secondary migration.

[0094] S9 Resource management - Shut down the underloaded host that has completed migration and update the resource distribution status of the data center. Shut down the underloaded host and update the resource status.

[0095] Through the above S1 - S9 methods, intelligent allocation is performed according to the proposed relevant models in the entire resource scheduling process of the data center, and finally the minimum energy consumption is obtained on the premise of ensuring the service quality of the host.

[0096] Example 2: The adjustment method of the dynamic sliding window size described in step S2 of the energy - saving method based on multi - dimensional resource intelligent scheduling in this cloud computing system is to dynamically expand or contract the window duration according to the absolute difference between the maximum fluctuation value of the resource utilization rate and the mean value at the previous moment. As described in detail for step S2 in Example 1.

[0097] Example 3: The information entropy algorithm described in step S3 of the energy - saving method based on multi - dimensional resource intelligent scheduling in this cloud computing system includes

[0098] - Construct a normalized matrix of the mean load of each dimension of resources;

[0099] - Calculate the entropy value and contribution degree of each dimension of resources;

[0100] - Allocate weight coefficients according to the contribution degree.

[0101] As described in detail for step S3 in Example 1.

[0102] Example 4: The establishment of the ARIMA model described in step S4 of the energy - saving method based on multi - dimensional resource intelligent scheduling in this cloud computing system includes

[0103] - Conduct a stationarity test and differencing process on the comprehensive load time series;

[0104] - Determine the model order according to the autocorrelation function and partial autocorrelation function;

[0105] - Select the optimal prediction model through the Akaike information criterion.

[0106] As described in detail for step S4 in Embodiment 1.

[0107] Embodiment 5: In the energy-saving method based on multi-dimensional resource intelligent scheduling in this cloud computing system, the classification criteria for overloaded and under-loaded types in the host described in step S5 are as follows

[0108] - The overloaded type is that both the current comprehensive load and the predicted value exceed the preset upper threshold;

[0109] - The under-loaded type is that both the current comprehensive load and the predicted value are lower than the preset lower threshold.

[0110] As described in detail for step S5 in Embodiment 1.

[0111] Embodiment 6: In the energy-saving method based on multi-dimensional resource intelligent scheduling in this cloud computing system, the virtual machine migration selection model described in step S6 is to preferentially select the virtual machine that maximizes the gradient of the utilization rate reduction of each dimension of resources of the overloaded host. As described in detail for step S6 in Embodiment 1.

[0112] Embodiment 7: In the energy-saving method based on multi-dimensional resource intelligent scheduling in this cloud computing system, the resource complementarity described in step S8 is calculated by the Pearson correlation coefficient, and the host that best matches the resource requirements of the virtual machine to be migrated is selected. As described in detail for step S8 in Embodiment 1.

[0113] Embodiment: The energy-saving system based on multi-dimensional resource intelligent scheduling in this cloud computing system is characterized by including the following functional modules

[0114] - Local monitor: Set on each host node, collect CPU, memory, network, and storage resource utilization data through an agent inside the virtual machine;

[0115] - Global monitor: Set in the cloud management platform, receive and aggregate the monitoring information from the local monitor;

[0116] - Load mean module: Set in the intelligent management scheduler, calculate the load mean of each dimension of resources based on a dynamic sliding window;

[0117] - Comprehensive load calculation module: Set in the intelligent management scheduler, calculate the comprehensive load of the host by weighted calculation through the information entropy algorithm;

[0118] - Comprehensive load prediction module: Set in the intelligent management scheduler, generate the load prediction value at a future moment based on the ARIMA model;

[0119] - Migration virtual machine selection module: Set in the intelligent scheduler to determine the virtual machines to be migrated on overloaded hosts according to the migration model;

[0120] - Virtual machine placement module: Set in the intelligent scheduler to match target hosts based on resource complementarity and perform migrations;

[0121] - Resource management module: Set in the cloud management platform to shut down underloaded hosts and update resource status.

[0122] Among them, the virtual machine placement module integrates the KVM virtualization interface and completes virtual machine migration through live migration technology.

[0123] The energy-saving method and system based on multi-dimensional resource intelligent scheduling in this cloud computing system overcome the defects such as unbalanced single-dimensional optimization, poor dynamic load adaptability, and blindness in resource matching existing in the resource scheduling of existing cloud computing systems. Through dynamic sliding window mean calculation and ARIMA load prediction model, it realizes the accurate perception of multi-dimensional resource fluctuations and the prediction of future trends; through information entropy algorithm dynamic weighting and Pearson correlation coefficient complementary matching, it achieves global resource optimization for virtual machine migration decisions; through KVM virtualization live migration and intelligent shutdown of underloaded hosts, it ensures business continuity while significantly reducing the energy consumption of the data center.

[0124] The above description shows the main features, basic principles, and advantages of the present invention. For those skilled in the art, it is obvious that the present invention is not limited to the details of the above exemplary embodiments or examples, and can be implemented in other specific forms without departing from the spirit or basic characteristics of the present invention. Therefore, the above embodiments or examples should be regarded as exemplary and non-limiting. The scope of the present invention is defined by the appended claims rather than the above description. Therefore, all changes falling within the meaning and scope of the equivalent elements of the claims are intended to be included in the present invention. Any reference signs in the claims should not be regarded as limiting the claimed claims.

[0125] In addition, it should be understood that although this specification is described according to embodiments, not every embodiment only contains an independent technical solution. This narrative way of the specification is only for clarity. Those skilled in the art should regard the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. An energy-saving method based on intelligent scheduling of multi-dimensional resources in a cloud computing system, characterized in that: It includes the following steps: S1 - Collect CPU, memory, network, and storage resource utilization data of the host node in real time; S2 - Calculate the load mean of each dimension of resources based on a dynamic sliding window, and the size of the sliding window is dynamically adjusted according to the resource load fluctuation; S3 - Determine the weight of each dimension of resources through the information entropy algorithm, and calculate the comprehensive load of the host by weighting; S4 - Predict the comprehensive load value of the host at a future time based on the ARIMA model; S5 - Classify the host as overloaded, normal load, or underloaded type according to the current comprehensive load and the predicted value; S6 - For overloaded hosts, determine the virtual machines to be migrated based on the virtual machine migration selection model; S7 - For underloaded hosts, migrate all their virtual machines to normal load hosts and shut down the underloaded hosts; S8 - Select a target host for the virtual machine to be migrated based on the resource complementarity, complete the migration and verify the resource status; after migration, verify whether the comprehensive load and the predicted value of the target host exceed the upper limit threshold, and if so, trigger a secondary migration; S9 - Shut down the underloaded host after the migration is completed, and update the resource distribution status of the data center.

2. The energy-saving method based on intelligent scheduling of multi-dimensional resources in the cloud computing system according to claim 1, characterized in that: The adjustment method of the dynamic sliding window size in step S2 is to dynamically expand or contract the window duration according to the absolute difference between the maximum fluctuation value of the resource utilization rate and the mean value at the previous moment.

3. The energy-saving method based on intelligent scheduling of multi-dimensional resources in the cloud computing system according to claim 1, characterized in that: The information entropy algorithm described in step S3 includes - Construct a normalized matrix of the load mean of each dimension of resources; - Calculate the entropy value and contribution degree of each dimension of resources; - Allocate weight coefficients according to the contribution degree.

4. The energy-saving method based on intelligent scheduling of multi-dimensional resources in the cloud computing system according to claim 1, characterized in that: The establishment of the ARIMA model described in step S4 includes - Conduct a stationarity test and differencing processing on the comprehensive load time series; - Determine the model order according to the autocorrelation function and partial autocorrelation function; - Select the optimal prediction model through the Akaike information criterion.

5. The energy-saving method based on intelligent scheduling of multi-dimensional resources in the cloud computing system according to claim 1, characterized in that: The classification criteria for overloaded and underloaded types of hosts in step S5 are - The overloaded type is that both the current comprehensive load and the predicted value exceed the preset upper limit threshold; - The underloaded type is that both the current comprehensive load and the predicted value are lower than the preset lower limit threshold.

6. The energy-saving method based on intelligent scheduling of multi-dimensional resources in the cloud computing system according to claim 1, characterized in that: The virtual machine migration selection model described in step S6 is to preferentially select the virtual machine that can make the utilization rate of each dimension of resources of the overloaded host decrease with the largest gradient.

7. The energy-saving method based on intelligent scheduling of multi-dimensional resources in the cloud computing system according to claim 1, characterized in that: The resource complementarity described in step S8 is calculated through the Pearson correlation coefficient, and the host that best matches the resource requirements of the virtual machine to be migrated is selected.

8. An energy-saving system based on intelligent scheduling of multi-dimensional resources in a cloud computing system, characterized in that: The energy-saving system based on multi-dimensional resource intelligent scheduling in the cloud computing system is used to implement the energy-saving method based on multi-dimensional resource intelligent scheduling in the cloud computing system described in claims 1-7; the system includes the following functional modules - Local monitor: Set on each host node, and collect CPU, memory, network, and storage resource utilization data through the agent inside the virtual machine; - Global monitor: Set in the cloud management platform, and receive and summarize the monitoring information from the local monitor; - Load mean module: Set in the intelligent management scheduler, and calculate the load mean of each dimension of resources based on a dynamic sliding window; - Comprehensive load calculation module: Set in the intelligent management scheduler, and calculate the comprehensive load of the host by weighting through the information entropy algorithm; - Comprehensive load prediction module: It is set in the intelligent management scheduler and generates load prediction values for future moments based on the ARIMA model; - Migrated virtual machine selection module: It is set in the intelligent scheduler and determines the virtual machines to be migrated on overloaded hosts according to the migration model; - Virtual machine placement module: It is set in the intelligent scheduler, matches the target host based on the resource complementarity degree and performs migration; - Resource management module: It is set in the cloud management platform, shuts down under-loaded hosts and updates the resource status.

9. The energy-saving system based on intelligent scheduling of multi-dimensional resources in the cloud computing system according to claim 8, characterized in that: The virtual machine placement module integrates the KVM virtualization interface and completes virtual machine migration through the live migration technology.

Images