A data center computing power scheduling management method and system

Abstract

The application discloses a data center computing power scheduling management method and system, and relates to the technical field of computing resource management. Through obtaining virtual machine computing power use data and demand prediction, prediction data is obtained, and the load state of a physical machine is judged; a to-be-migrated virtual machine set is determined according to the load state of each physical machine; in combination with the physical machine specifications, a meta-heuristic algorithm is used to optimize to-be-migrated virtual machine distribution, and a virtual machine placement scheme is generated. Compared with a traditional passive response strategy, the scheme can plan migration according to future load states, reduce unnecessary migration operations and migration overhead, improve physical machine resource utilization at the same time, and realize efficient scheduling and resource allocation optimization of virtual machine migration.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computing resource management, and in particular to a data center computing power scheduling management method and system. BACKGROUND

[0002] With the rapid development of cloud computing and big data, data centers are expanding in size. The core goal is to meet the needs of users for high-performance computing and quality of service, while reducing energy consumption and improving resource utilization. Virtual machines, as a virtualized instance running on a physical server, are used to carry applications and services, and to achieve resource isolation and elastic allocation.

[0003] However, existing resource scheduling and virtual machine management techniques still face many challenges. First, traditional virtual machine integration and migration strategies are mostly passive response types, i.e. triggering migration when detecting that a physical machine is overloaded or lightly loaded. This approach has significant shortcomings: on the one hand, due to the sensitivity to transient load fluctuations, frequent and even unnecessary virtual machine migration often occurs, increasing network and migration overhead and causing virtual machine performance to decline; on the other hand, passive response lags behind actual load changes, causing virtual machines to not be able to obtain sufficient resources in a timely manner, affecting service quality and increasing the risk of service level agreement violations. SUMMARY

[0004] The present application aims to solve the problems of migration lag and frequent migration mentioned in the background art, and proposes a data center computing power scheduling management method and system.

[0005] In the first aspect of the present application, a data center computing power scheduling management method is provided. The servers of the data center include a plurality of physical machines, and each running physical machine has at least one virtual machine deployed thereon. The method comprises:

[0006] Obtaining usage data of any virtual machine on computing power resources and making demand prediction to obtain demand prediction data;

[0007] Judging the load state of a target physical machine according to the demand prediction data of the virtual machines on the target physical machine; the target physical machine is any one of the plurality of physical machines;

[0008] Determining a set of virtual machines to be migrated according to the load state of each physical machine;

[0009] According to the specifications of each physical machine, a meta-heuristic algorithm is used to optimize the allocation of the set of virtual machines to be migrated to obtain a virtual machine placement scheme;

[0010] According to the virtual machine placement scheme, the virtual machines to be migrated are migrated to the corresponding physical machines, and the resource allocation of the physical machines is adjusted according to the resource requirements of each virtual machine.

[0011] Optionally, the judging the load state of the target physical machine according to the demand prediction data of the virtual machines on the target physical machine comprises:

[0012] obtaining current load data and load prediction data of the target physical machine according to current resource usage data and demand prediction data of all virtual machines on the target physical machine;

[0013] if both the current load data and the load prediction data exceed a first preset threshold, determining that the target physical machine is in a high load state;

[0014] if both the current load data and the load prediction data are lower than a second preset threshold, determining that the target physical machine is in a light load state;

[0015] if both the current load data and the load prediction data do not exceed the first preset threshold, but at least one exceeds the second preset threshold, determining that the first target physical machine is in a normal load state.

[0016] Optionally, the determining the set of virtual machines to be migrated according to the load state of each physical machine comprises:

[0017] including all virtual machines on a physical machine in a light load state into the set of virtual machines to be migrated;

[0018] including all virtual machines on a physical machine in a normal state as a whole virtual machine instance into the set of virtual machines to be migrated;

[0019] for any physical machine in a high load state, sequentially including a virtual machine with the smallest memory into the set of virtual machines to be migrated until the load of the physical machine is lower than the first preset threshold.

[0020] Optionally, each physical machine has the same specification; the set of virtual machines to be migrated is distributed and optimized by using an improved genetic algorithm; the specific process comprises:

[0021] Step 1: generating a distribution scheme based on the Best Fit strategy, denoted as a first feasible solution; and determining an initial maximum gene value according to the number of physical machines used by the first feasible solution;

[0022] Step 2: randomly generating a plurality of different chromosomes and repairing them by using a preset repair strategy to form an initial population;

[0023] Step 3: calculating the fitness of each chromosome to determine a current optimal solution; if the number of physical machines used by the current optimal solution is less than the current maximum gene value, updating the maximum gene value to the number of physical machines used by the current optimal solution, and scaling the gene values of the chromosomes;

[0024] Step four, a termination condition judgment is performed, if the termination condition is met, the iteration is stopped, and the chromosome with the highest fitness is output; otherwise, step five is entered;

[0025] Step five, genetic operation is performed, and the chromosome is updated; return to step three.

[0026] Optionally, the repair strategy of the chromosome comprises:

[0027] For any chromosome, denoted as a target chromosome, the load conditions of each physical machine in the virtual machine allocation scheme corresponding to the target chromosome are detected, and the physical machine whose load exceeds an upper threshold is recorded as an overloaded physical machine and is included in an overloaded physical machine set;

[0028] The virtual machines on the target overloaded physical machine are arranged in descending order of resource demand, and are sequentially included in a to-be-reallocated set until the load of the target overloaded physical machine is less than the upper threshold; the target overloaded physical machine is any one of the overloaded physical machines;

[0029] The virtual machines in the to-be-reallocated set are arranged in descending order of resource demand, and are allocated based on a Best Fit strategy.

[0030] If the reallocation is successful, the gene value of the target chromosome is updated according to the obtained allocation scheme.

[0031] The second aspect of the embodiment of the application provides a data center computing power scheduling management system, the server of the data center comprises a plurality of physical machines, and at least one virtual machine is deployed on each running physical machine; the system comprises:

[0032] A demand prediction module is configured to obtain usage data of computing power resources of any virtual machine, and perform demand prediction to obtain demand prediction data;

[0033] A state classification module is configured to determine the load state of a target physical machine according to the demand prediction data of the virtual machines on the target physical machine; the target physical machine is any one of the plurality of physical machines;

[0034] A to-be-migrated determination module is configured to determine a to-be-migrated virtual machine set according to the load state of each physical machine;

[0035] A scheme generation module is configured to perform allocation optimization on the to-be-migrated virtual machine set by using a meta-heuristic algorithm according to the specifications of the physical machines, to obtain a virtual machine placement scheme;

[0036] A scheme execution module is configured to migrate the to-be-migrated virtual machines to corresponding physical machines according to the virtual machine placement scheme, and adjust the resource allocation of the physical machines according to the resource demands of the virtual machines.

[0037] Optionally, the state classification module comprises:

[0038] a load determination module configured to obtain current load data and load prediction data of the target physical machine according to current resource usage data and demand prediction data of all virtual machines on the target physical machine;

[0039] a first state determination module configured to determine that the target physical machine is in a high load state if the current load data and the load prediction data both exceed a first preset threshold value;

[0040] a second state determination module configured to determine that the target physical machine is in a light load state if the current load data and the load prediction data both are lower than a second preset threshold value;

[0041] a third state determination module configured to determine that the target physical machine is in a normal load state if the current load data and the load prediction data both do not exceed the first preset threshold value, but at least one of them exceeds the second preset threshold value.

[0042] 8. The data center computing power scheduling management system of claim 7, wherein the to-be-migrated determination module comprises:

[0043] a light load emptying module configured to include all virtual machines on a physical machine in a light load state into a to-be-migrated virtual machine set;

[0044] a normal state integrating module configured to include all virtual machines on a physical machine in a normal state as a whole virtual machine instance into the to-be-migrated virtual machine set;

[0045] a high load migration-out module configured to sequentially include a virtual machine with the smallest memory on a physical machine in a high load state into the to-be-migrated virtual machine set until the load of the physical machine is lower than the first preset threshold value.

[0046] Optionally, the specifications of each physical machine are consistent; the scheme generation module performs distribution optimization on the to-be-migrated virtual machine set by using an improved genetic algorithm; specifically, the scheme generation module comprises:

[0047] a feasible solution generation module configured to generate an allocation scheme based on a Best Fit strategy, denoted as a first feasible solution; and determine an initial gene maximum value according to the number of physical machines used by the first feasible solution;

[0048] a population initialization module configured to randomly generate a plurality of different chromosomes;

[0049] a chromosome repair module configured to repair the chromosomes by using a preset repair strategy to form an initial population;

[0050] An adaptability calculation module is configured to calculate the adaptability of each chromosome and determine a current optimal solution;

[0051] A range adjustment module is configured to update the gene maximum value to the number of physical machines used by the current optimal solution and scale the gene values of the chromosomes if the number of physical machines used by the current optimal solution is less than the current gene maximum value;

[0052] A termination determination module is configured to determine a termination condition, stop iteration, and output the chromosome with the highest adaptability if the termination condition is met, or enter the genetic operation module if the termination condition is not met.

[0053] The genetic operation module is configured to perform genetic operations and update the chromosomes, and return to the adaptability calculation module.

[0054] Optionally, the chromosome repair module comprises:

[0055] An over-limit detection module is configured to detect the load of each physical machine in the virtual machine allocation scheme corresponding to a target chromosome, which is any chromosome, and record the physical machine with a load exceeding an upper limit threshold as an overloaded physical machine and include the overloaded physical machine in an overloaded physical machine set.

[0056] A distribution target determination module is configured to arrange the virtual machines on the target overloaded physical machine in ascending order of resource demand, and sequentially include the virtual machines in a to-be-reallocated set until the load of the target overloaded physical machine is less than the upper limit threshold; the target overloaded physical machine is any overloaded physical machine.

[0057] A reallocation module is configured to arrange the virtual machines in the to-be-reallocated set in descending order of resource demand and allocate the virtual machines based on a Best Fit strategy.

[0058] A conversion module is configured to update the gene values of the target chromosome according to the obtained allocation scheme if the reallocation is successful.

[0059] The present application has the following advantages:

[0060] By predicting the resource demand of the virtual machines, determining the change trend of the load state of the physical machines, and optimizing the allocation of the to-be-migrated virtual machines using a meta-heuristic algorithm, a reasonable migration scheme is generated. Compared with the traditional passive response strategy, it can plan migration according to the future load state, reduce unnecessary migration operations and migration overhead, and improve the utilization rate of physical machine resources, thereby achieving efficient scheduling and resource allocation optimization of virtual machine migration. BRIEF DESCRIPTION OF DRAWINGS

[0061] Figure 1 A flowchart of a data center computing power scheduling management method is provided for the embodiments of the present application.

[0062] Figure 2An architecture diagram of a data center computing power scheduling management system is provided for the embodiment of the present application. DETAILED DESCRIPTION

[0063] In order to further illustrate the technical means and effects adopted by the present application to achieve the predetermined application purposes, the specific embodiments, structures, features and effects according to the present application are described in detail below in combination with the drawings and preferred embodiments.

[0064] The embodiment of the present application provides a data center computing power scheduling management method. Figure 1 , Figure 1 A flowchart of a data center computing power scheduling management method is provided for the embodiment of the present application. The method comprises the following steps:

[0065] S101, obtaining the use data of computing power resources of each virtual machine and performing demand prediction to obtain demand prediction data.

[0066] S102, judging the load state of the target physical machine according to the demand prediction data of the virtual machine on the target physical machine.

[0067] S103, determining the set of virtual machines to be migrated according to the load state of each physical machine.

[0068] S104, performing distribution optimization on the set of virtual machines to be migrated by using a meta-heuristic algorithm according to the specifications of each physical machine to obtain a virtual machine placement scheme.

[0069] S105, migrating the virtual machines to be migrated to the corresponding physical machines according to the virtual machine placement scheme, and adjusting the resource allocation of the physical machines according to the resource requirements of each virtual machine.

[0070] The resource requirement data includes CPU resources and memory resources. The target physical machine is any one of the plurality of physical machines.

[0071] Based on the data center computing power scheduling management method provided by the embodiment of the present application, the change trend of the load state of the physical machine is judged by predicting the resource requirements of the virtual machine, and the meta-heuristic algorithm is used to optimize the distribution of the virtual machines to be migrated, so as to generate a reasonable migration scheme. Compared with the traditional passive response strategy, it can plan migration according to the future load state, reduce unnecessary migration operations and migration overhead, and at the same time improve the resource utilization rate of the physical machine, so as to realize efficient scheduling and resource allocation optimization of virtual machine migration.

[0072] In one embodiment, the specifications of each physical machine are consistent. Each type of resource is normalized in the form of a proportion. If a physical machine provides CPU and memory resources in the amounts CM and MM, respectively, the resource amounts of the physical machine are normalized to 1, and the resource requirements (CPU i , MEM i ) of the virtual machine i are normalized to (CPU i / CM, MEM i / MM) after normalization. By normalizing the CPU and memory resources of the physical machine to 1, different types of resources are mapped to the same scale, so that the dimensions of CPU and memory are consistent, facilitating comparison and calculation in the scheduling algorithm.

[0073] In one embodiment, a bidirectional long short-term memory network model (BiLSTM) can be used for prediction. For example, data is collected at a period of 5 minutes to obtain time series data of virtual machine resource usage, which is preprocessed and input into the prediction model to obtain subsequent resource requirement data. BiLSTM can simultaneously learn past dependencies and future trends of time series, and has high accuracy in time series prediction.

[0074] In one embodiment, step S102, according to the demand prediction data of the virtual machines on the target physical machine, the load state of the target physical machine is determined, including:

[0075] Step one, according to the current resource usage data and demand prediction data of all virtual machines on the target physical machine, the current load data and load prediction data of the target physical machine are obtained. The load of the physical machine is the demand of the virtual machines thereon, which is an equivalent relationship, including CPU load and memory load.

[0076] Step two, if the current load data and the load prediction data both exceed the first preset threshold, it is determined that the target physical machine is in a high load state. For example, the first preset threshold can be set to 0.9.

[0077] Step three, if the current load data and the load prediction data are both lower than the second preset threshold, it is determined that the target physical machine is in a light load state. For example, the second preset threshold can be set to 0.5.

[0078] Step four, if the current load data and the load prediction data do not exceed the first preset threshold, but at least one of them exceeds the second preset threshold, it is determined that the first target physical machine is in a normal load state.

[0079] Among them, the maximum value in the load prediction data is compared with the threshold.

[0080] In one implementation, if the current load data exceeds the first preset threshold, but the load prediction data falls back and does not exceed the first preset threshold, it is considered as a higher load state and does not participate in scheduling.

[0081] This embodiment considers both current usage and future demand, can more accurately determine the load state of the physical machine, and avoid misjudgment caused by instantaneous data, such as the physical machine with short-term fluctuation of current load not participating in migration immediately, avoiding performance loss and network overhead caused by frequent migration.

[0082] In one embodiment, step S103, determining the set of virtual machines to be migrated according to the load state of each physical machine comprises:

[0083] Step one, all virtual machines on the physical machine in the light load state are included in the set of virtual machines to be migrated.

[0084] Step two, all virtual machines on the physical machine in the normal state are included in the set of virtual machines to be migrated as a whole virtual machine instance. For example, a physical machine has two virtual machines VM1 and VM2, and the resource requirements are (CPU1, MEM1) and (CPU2, MEM2) respectively, then the whole is included as a virtual machine to be migrated, and the resource requirement is (CPU1+CPU2, MEM1+MEM2).

[0085] Step three, for any physical machine in the high load state, the virtual machine with the smallest memory is sequentially included in the set of virtual machines to be migrated until the load of the physical machine is lower than the first preset threshold.

[0086] In one implementation, the virtual machines of the normal load physical machine are migrated as a whole, reducing the complexity and unnecessary scheduling caused by scattered migration. The combination of light load, normal load, and high load strategies realizes targeted processing of different load conditions, and balances resource release and system stability.

[0087] In one embodiment, in step S104, an improved genetic algorithm is used to optimize the allocation of the set of virtual machines to be migrated. The specific process includes:

[0088] Step one, based on the Best Fit strategy, generate an allocation scheme, denoted as the first feasible solution; and determine the number of physical machines used by the first feasible solution as the initial maximum gene value.

[0089] Step two, randomly generate multiple different chromosomes and use a preset repair strategy to repair them to form an initial population.

[0090] Step three, calculate the fitness of each chromosome to determine the current optimal solution; if the number of physical machines used by the current optimal solution is less than the current maximum gene value, update the maximum gene value to the number of physical machines used by the current optimal solution, and scale the gene value of the chromosome, specifically: ; wherein, , It represents the maximum value of the gene before and after the update. , These are the gene values ​​before and after scaling. `round` is the integer function, which rounds the values ​​to the nearest integer.

[0091] Step four: Determine the termination condition. If the termination condition is met, stop the iteration and output the chromosome with the highest fitness; otherwise, proceed to step five. Termination conditions include reaching the maximum number of iterations or the fitness of the optimal solution remaining unchanged for several consecutive generations.

[0092] Step 5: Perform genetic manipulation to update the chromosomes; return to step 3.

[0093] This embodiment uses a Best Fit strategy to generate initial feasible solutions, allowing the algorithm to start the search from a better starting point and improving search efficiency. A pre-defined repair strategy is employed to repair the randomly generated chromosomes, ensuring that each allocation scheme meets resource constraints (CPU, memory, etc.) and preventing infeasible solutions from interfering with the optimization process.

[0094] In one implementation, chromosomes are encoded using integers. For example, a chromosome X... i The vector form is [a i1 a i2 , ..., a in ];a ij This indicates that the j-th virtual machine is included in the a-th virtual machine. ij Group. Among them, size is the population size; , where n is the number of virtual machines to be migrated; , where m is the maximum gene value, representing the number of physical machines selected for migration.

[0095] In one implementation, the search space is reduced by dynamically updating the maximum value of the gene, thus avoiding the algorithm from wasting computational resources in the redundant range and improving the convergence speed.

[0096] In one implementation, each physical machine j has a certain amount of CPU remaining. and remaining memory Each virtual machine has CPU requirements. and memory requirements (The quantity demanded for each category equals the maximum of its current demand and its forecast), total demand L i = + .

[0097] Based on the Best Fit strategy, the generated allocation scheme includes:

[0098] Step 1: Sort the virtual machines to be migrated from largest to smallest according to the total demand, forming a queue.

[0099] Step two, take out the head virtual machine, and screen the candidate physical machine set meeting the following conditions: And If there is no physical machine meeting the conditions, a failure is generated.

[0100] Step three, for each candidate physical machine, calculate its remaining capacity and the matching distance of the virtual machine , select the candidate physical machine with the minimum distance, put the virtual machine into it, and update the resource remaining amount of the physical machine.

[0101] Step four, if the queue is not empty, return to step two. Otherwise, end the allocation process.

[0102] Wherein, the remaining amount is the difference between the upper threshold and the existing load, rather than the difference between the resource rated value and the existing load. For example, the upper threshold can be set to 0.9. In this way, the physical machine will not run at a high load after the virtual machine is allocated.

[0103] The implementation mode sorts the total demand of virtual machines from large to small, and allocates large demand virtual machines first, which can avoid large virtual machines from being unable to be placed due to insufficient space, thereby reducing resource fragmentation. By calculating the matching distance and selecting the physical machine with the minimum distance, the virtual machine allocation fills the remaining resources of the physical machine as much as possible, reducing resource waste.

[0104] In an implementation mode, the repair strategy of the chromosome includes:

[0105] Step one, for any chromosome, denoted as a target chromosome, detect the load of each physical machine in the virtual machine allocation scheme corresponding to the target chromosome. If there is any physical machine with a load exceeding the upper threshold, record the chromosome as an illegal solution, and record the physical machine with a load exceeding the upper threshold as an overloaded physical machine, which is included in the overloaded physical machine set.

[0106] Step two, arrange the virtual machines on the target overloaded physical machine in ascending order of resource demand, and sequentially include them in the to-be-reallocated set until the load of the target overloaded physical machine is less than the upper threshold. The target overloaded physical machine is any overloaded physical machine.

[0107] Step three, arrange the virtual machines in the to-be-reallocated set in descending order of resource demand, and allocate them based on the BestFit strategy.

[0108] Step four, if the reallocation is successful, update the gene value of the target chromosome according to the obtained allocation scheme. If the reallocation fails, keep the original illegal solution unchanged.

[0109] In this implementation mode, the repair strategy can convert part of the illegal solution into a feasible solution, reduce the iterative calculation wasted by the algorithm due to invalid solutions, and improve the convergence speed.

[0110] In one implementation, the fitness function is:

[0111] ; wherein F is a fitness value; apm is the number of physical machines used by the chromosome; total_overload is the total overload amount; overloadj is the load overload amount of the jth physical machine, which is 0 when not overloaded; and a is a penalty coefficient, which can be set to 1000 to ensure that the fitness of illegal solutions is low.

[0112] The optimization goal of this implementation is to minimize the number of physical machines used, and the fitness function reduces the fitness of solutions that do not meet the resource constraints through a penalty term, thereby reducing the spread of illegal solutions in the population and guiding the algorithm to gradually converge to a feasible solution that uses the fewest physical machines and meets the resource constraints.

[0113] In one embodiment, the virtual machine placement scheme includes grouping results, i.e., which virtual machines should be placed on the same physical machine.

[0114] Step S105, according to the virtual machine placement scheme, migrating the to-be-migrated virtual machines to the corresponding physical machines includes:

[0115] Step one, calculate the total resource demand of each group of virtual machines, and form a queue in descending order.

[0116] Step two, take out the head virtual machine group in turn, calculate the memory load caused by the virtual machines in the group on each physical machine according to the distribution of the virtual machines in the group, and select the physical machine with the largest memory load as the anchor physical machine. Migrate the other virtual machines in the group to the anchor physical machine. That is, do not migrate the virtual machines that have formed a large memory load.

[0117] This embodiment selects the physical machine where the virtual machine that has formed a large memory load in the group as the anchor point, and migrates the other virtual machines in the group to this physical machine, thereby avoiding migrating high-load virtual machines, effectively reducing the amount of memory data transmission generated during migration, reducing migration overhead, and improving overall migration efficiency.

[0118] The embodiment of the application provides a data center computing power scheduling management system. Referring to Figure 2 , Figure 2 An architectural diagram of a data center computing power scheduling management system is provided for the embodiment of the application. The system includes:

[0119] The demand prediction module is configured to obtain usage data of any virtual machine on computing power resources and perform demand prediction to obtain demand prediction data.

[0120] The state classification module is configured to determine the load state of the target physical machine according to the demand prediction data of the virtual machines on the target physical machine.

[0121] The to-migrate determining module is configured to determine a set of to-migrate virtual machines according to the load state of each physical machine.

[0122] The scheme generating module is configured to optimize the allocation of the set of to-migrate virtual machines by using a meta-heuristic algorithm according to the specifications of the physical machines, to obtain a virtual machine placement scheme.

[0123] The scheme executing module is configured to migrate the to-migrate virtual machines to the corresponding physical machines according to the virtual machine placement scheme, and adjust the resource allocation of the physical machines according to the resource requirements of the virtual machines.

[0124] The resource requirement data includes CPU resources and memory resources. The target physical machine is any one of the plurality of physical machines.

[0125] The data center computing power scheduling management system provided by the embodiment of the present application can predict the resource requirements of virtual machines, determine the change trend of the load state of physical machines, and optimize the allocation of to-migrate virtual machines by using a meta-heuristic algorithm, so as to generate a reasonable migration scheme. Compared with the traditional passive response strategy, it can plan migration according to the future load state, reduce unnecessary migration operations and migration overhead, improve the resource utilization rate of physical machines, and realize efficient scheduling and resource allocation optimization of virtual machine migration.

[0126] In one embodiment, the state classification module includes:

[0127] The load determining module is configured to obtain current load data and load prediction data of the target physical machine according to the current resource usage data and requirement prediction data of all virtual machines on the target physical machine.

[0128] The first state determining module is configured to determine that the target physical machine is in a high load state if the current load data and the load prediction data both exceed a first preset threshold.

[0129] The second state determining module is configured to determine that the target physical machine is in a light load state if the current load data and the load prediction data are both lower than a second preset threshold.

[0130] The third state determining module is configured to determine that the first target physical machine is in a normal load state if the current load data and the load prediction data do not exceed the first preset threshold, but at least one of them exceeds the second preset threshold.

[0131] In one embodiment, the to-migrate determining module includes:

[0132] The light load emptying module is configured to include all virtual machines on the physical machine in the light load state in the set of to-migrate virtual machines.

[0133] The normal state integration module is configured to integrate all virtual machines on a physical machine in a normal state as a whole virtual machine instance into the set of virtual machines to be migrated.

[0134] The high-load migration-out module is configured to, for any physical machine in a high-load state, sequentially integrate the virtual machine with the least memory into the set of virtual machines to be migrated until the load of the physical machine is lower than the first preset threshold.

[0135] In an embodiment, the specifications of each physical machine are consistent. The scheme generation module optimizes allocation of the set of virtual machines to be migrated using an improved genetic algorithm. Specifically, the scheme generation module includes:

[0136] The feasible solution generation module is configured to generate an allocation scheme based on a Best Fit strategy, denoted as a first feasible solution; and determine an initial maximum gene value based on the number of physical machines used by the first feasible solution.

[0137] The population initialization module is configured to randomly generate a plurality of different chromosomes.

[0138] The chromosome repair module is configured to repair the chromosomes using a preset repair strategy to form an initial population.

[0139] The fitness calculation module is configured to calculate the fitness of each chromosome to determine a current optimal solution.

[0140] The range adjustment module is configured to, if the number of physical machines used by the current optimal solution is less than the current maximum gene value, update the maximum gene value to the number of physical machines used by the current optimal solution, and scale the gene values of the chromosomes.

[0141] The termination determination module is configured to determine a termination condition. If the termination condition is met, stop iteration and output the chromosome with the highest fitness. Otherwise, enter the genetic operation module.

[0142] The genetic operation module is configured to perform genetic operations and update the chromosomes. The genetic operation module returns to the fitness calculation module.

[0143] In an embodiment, the chromosome repair module includes:

[0144] The overrun detection module is configured to, for any chromosome, denoted as a target chromosome, detect the load of each physical machine in the virtual machine allocation scheme corresponding to the target chromosome, and integrate any physical machine with a load exceeding an upper threshold value into a set of overloaded physical machines.

[0145] The allocation target determination module is configured to arrange the virtual machines on a target overloaded physical machine in ascending order of resource demand, and sequentially integrate the virtual machines into a set of virtual machines to be reallocated until the load of the target overloaded physical machine is less than the upper threshold value. The target overloaded physical machine is any overloaded physical machine.

[0146] The re-distribution module is arranged to arrange the virtual machines in the set to be re-distributed according to the resource demand from large to small, and to allocate based on the Best Fit strategy.

[0147] The conversion module is arranged to update the gene value of the target chromosome according to the obtained allocation scheme if the re-distribution is successful.

[0148] The above is only the preferred embodiment of the present application, and does not limit the present application in any form. Although the present application has been disclosed as above with the preferred embodiment, it is not intended to limit the present application. Any person skilled in the art can make some changes or modifications to the above disclosed technical content to obtain equivalent embodiments with equivalent changes, without departing from the technical solution of the present application. Any simple modification, equivalent change and modification made to the above embodiments according to the technical essence of the present application, without departing from the technical solution of the present application, shall fall within the scope of the technical solution of the present application.

Claims

1. A data center computing power scheduling management method, characterized in that, The server of the data center comprises a plurality of physical machines, and at least one virtual machine is deployed on each running physical machine; the method comprises: obtaining the use data of computing power resources of any virtual machine and performing demand prediction to obtain demand prediction data; judging the load state of a target physical machine according to the demand prediction data of the virtual machines on the target physical machine; the target physical machine is any of the plurality of physical machines; the load state comprises a high load state, a light load state and a normal state; determining a set of virtual machines to be migrated according to the load state of each physical machine; performing allocation optimization on the set of virtual machines to be migrated according to the specifications of each physical machine by using a meta-heuristic algorithm to obtain a virtual machine placement scheme; migrating the virtual machines to be migrated to the corresponding physical machines according to the virtual machine placement scheme, and adjusting the resource allocation of the physical machines according to the resource requirements of each virtual machine; the determination of the set of virtual machines to be migrated according to the load state of each physical machine comprises: including all the virtual machines on the physical machine in the light load state into the set of virtual machines to be migrated; including all the virtual machines on the physical machine in the normal state as a whole virtual machine instance into the set of virtual machines to be migrated; for any physical machine in the high load state, sequentially including the virtual machine with the smallest memory into the set of virtual machines to be migrated until the load of the physical machine is lower than a first preset threshold.

2. The data center computing power scheduling management method of claim 1, wherein, the judgment of the load state of the target physical machine according to the demand prediction data of the virtual machines on the target physical machine comprises: obtaining the current load data and the load prediction data of the target physical machine according to the current resource use data and the demand prediction data of all the virtual machines on the target physical machine; if both the current load data and the load prediction data exceed a first preset threshold, determining that the target physical machine is in a high load state; if both the current load data and the load prediction data are lower than a second preset threshold, determining that the target physical machine is in a light load state; if both the current load data and the load prediction data do not exceed the first preset threshold, but at least one of them exceeds the second preset threshold, determining that the target physical machine is in a normal load state.

3. The data center computing power scheduling management method of claim 1, wherein, the specifications of each physical machine are consistent; an improved genetic algorithm is used to perform allocation optimization on the set of virtual machines to be migrated; the specific process comprises: Step 1: based on the Best Fit strategy, generate an allocation scheme, denoted as the first feasible solution; and determine the initial maximum gene value according to the number of physical machines used by the first feasible solution; Step 2: randomly generate a plurality of different chromosomes and repair them using a preset repair strategy to form an initial population; Step 3: calculate the fitness of each chromosome to determine the current optimal solution; if the number of physical machines used by the current optimal solution is less than the current maximum gene value, update the maximum gene value to the number of physical machines used by the current optimal solution, and scale the gene value of the chromosome; Step 4: perform termination condition judgment; if the termination condition is met, stop iteration and output the chromosome with the highest fitness; otherwise, go to Step 5; Step 5: perform genetic operation and update the chromosome; return to Step 3.

4. The data center computing power scheduling management method of claim 3, wherein, the repair strategy of the chromosome comprises: For any chromosome, denoted as a target chromosome, the load conditions of each physical machine in the virtual machine allocation scheme corresponding to the target chromosome are detected, and a physical machine whose load exceeds an upper threshold is denoted as an overloaded physical machine and is included in an overloaded physical machine set; The virtual machines on the target overloaded physical machine are arranged in ascending order of resource demand, and are sequentially included in a to-be-reallocated set until the load of the target overloaded physical machine is less than the upper threshold; the target overloaded physical machine is any overloaded physical machine; The virtual machines in the to-be-reallocated set are arranged in descending order of resource demand, and are allocated based on a Best Fit strategy; If the reallocation is successful, the gene value of the target chromosome is updated according to the obtained allocation scheme.

5. A data center computing power scheduling management system, characterized in that, The server of the data center includes a plurality of physical machines, and at least one virtual machine is deployed on each running physical machine; the system includes: A demand prediction module is configured to obtain usage data of computing power resources of any virtual machine and perform demand prediction to obtain demand prediction data; A state classification module is configured to determine a load state of a target physical machine according to demand prediction data of virtual machines on the target physical machine; the target physical machine is any of the plurality of physical machines; the load state includes a high load state, a light load state, and a normal state; A to-be-migrated determination module is configured to determine a to-be-migrated virtual machine set according to the load state of each physical machine; A scheme generation module is configured to perform allocation optimization on the to-be-migrated virtual machine set by using a meta-heuristic algorithm according to the specifications of each physical machine to obtain a virtual machine placement scheme; A scheme execution module is configured to migrate the to-be-migrated virtual machines to corresponding physical machines according to the virtual machine placement scheme, and adjust resource allocation of the physical machines according to resource demands of each virtual machine; The to-be-migrated determination module includes: A light load emptying module is configured to include all virtual machines on a physical machine in a light load state in the to-be-migrated virtual machine set; A normal state integration module is configured to include all virtual machines on a physical machine in a normal state as a whole virtual machine instance in the to-be-migrated virtual machine set; A high load migration-out module is configured to sequentially include a virtual machine with the smallest memory in the to-be-migrated virtual machine set for any physical machine in a high load state until the load of the physical machine is lower than a first preset threshold.

6. The data center computing power scheduling management system of claim 5, wherein, The state classification module includes: A load determination module is configured to obtain current load data and load prediction data of a target physical machine according to current resource usage data and demand prediction data of all virtual machines on the target physical machine; A first state discrimination module is configured to determine that the target physical machine is in a high load state if both the current load data and the load prediction data exceed a first preset threshold; A second state discrimination module is configured to determine that the target physical machine is in a light load state if both the current load data and the load prediction data are lower than a second preset threshold; A third state discrimination module is configured to determine that a first target physical machine is in a normal load state if neither the current load data nor the load prediction data exceeds the first preset threshold, but at least one of them exceeds the second preset threshold.

7. The data center computing power scheduling management system of claim 5, wherein, The specifications of each physical machine are consistent. The scheme generation module adopts an improved genetic algorithm to optimize allocation of the set of virtual machines to be migrated. Specifically, the scheme generation module comprises: a feasible solution generation module configured to generate an allocation scheme based on a Best Fit strategy, denoted as a first feasible solution, and determine an initial gene maximum value based on a number of physical machines used by the first feasible solution; a population initialization module configured to randomly generate a plurality of different chromosomes; a chromosome repair module configured to repair the chromosomes using a preset repair strategy to form an initial population; a fitness calculation module configured to calculate fitness of each chromosome and determine a current optimal solution; a range adjustment module configured to, if a number of physical machines used by the current optimal solution is less than the current gene maximum value, update the gene maximum value to the number of physical machines used by the current optimal solution and scale gene values of the chromosomes; a termination determination module configured to determine a termination condition, if the termination condition is met, stop iteration and output a chromosome with the highest fitness; otherwise, enter a genetic operation module; the genetic operation module configured to perform genetic operation and update the chromosomes; and return to the fitness calculation module. 8.The data center computing power scheduling management system of claim 7, wherein, The chromosome repair module comprises: an overrun detection module configured to, for any chromosome, denoted as a target chromosome, detect load conditions of each physical machine in a virtual machine allocation scheme corresponding to the target chromosome, and record a physical machine with load exceeding an upper threshold value as an overloaded physical machine and include the overloaded physical machine in an overloaded physical machine set; an allocation target determination module configured to arrange virtual machines on a target overloaded physical machine in ascending order of resource demand, and sequentially include the virtual machines in a set to be reallocated until load of the target overloaded physical machine is less than the upper threshold value; the target overloaded physical machine is any overloaded physical machine; a reallocation module configured to arrange virtual machines in the set to be reallocated in descending order of resource demand and allocate the virtual machines based on a Best Fit strategy; a conversion module configured to, if reallocation is successful, update gene values of the target chromosome based on the obtained allocation scheme.

Images