CN110308965B

CN110308965B - Rule-based heuristic virtual machine distribution method and system for cloud data center

Info

Publication number: CN110308965B
Application number: CN201910470400.1A
Authority: CN
Inventors: 余显; 张广兴; 张春阳; 黄昆; 谢高岗
Original assignee: Institute of Computing Technology of CAS
Current assignee: Institute of Computing Technology of CAS
Priority date: 2019-05-31
Filing date: 2019-05-31
Publication date: 2021-09-24
Anticipated expiration: 2039-05-31
Also published as: CN110308965A

Abstract

The invention provides a heuristic virtual machine distribution method and system based on rules for a cloud data center, which comprises the following steps: all possible host states are scored according to the core rule that the score of each state depends on the score of the state to which the state can be transferred and the probability distribution of the corresponding state of the virtual machine; then, when the virtual machine migration occurs, the virtual machine selects a host with the highest state score after being placed for migration. Where a state transition refers to the process of one host changing to another resource utilization state by placing a virtual machine in a particular resource utilization state. Therefore, the invention can control the energy consumption and improve the service quality of the user.

Description

Rule-based heuristic virtual machine distribution method and system for cloud data center

Technical Field

The invention relates to the field of cloud computing, which is mainly applied to a cloud data center and used for guiding the migration of a virtual machine so as to reduce the energy consumption of the cloud data center and improve the computing power of the cloud data center.

Background

Unlike the traditional computing mode, cloud computing provides on-demand elastic resource services for users by means of virtualization technologies. With the expansion of the scale of the cloud computing data center, the problem of energy consumption is increasingly prominent. The existing virtual machine allocation technology guarantees the resource requirements of users by reasonably allocating the host positions of the virtual machines, reduces the number of active hosts as far as possible and improves the energy consumption efficiency.

The existing virtual machine allocation technology can be mainly divided into two modes of static allocation and dynamic allocation. The static allocation of the virtual machine refers to that the host is selected for placement at one time according to the resource request at the initialization moment of the virtual machine, and no additional processing is performed subsequently. The typical work of the method is that Oktopus guarantees the service quality of tenants by allocating enough network bandwidth for the tenants; or a Sync method is used for coordinating the relationship between the network policy and the virtual machine, so that the resource management of the virtual machine can also achieve sufficient service performance under different network policy environments; besides the scheduling of network resources, the Euclidean distance can be used for measuring the residual resource capacity of the host under various resource types, and the host with the minimum distance is statically selected for the virtual machine to be placed by taking the measured resource capacity as an index. The static allocation mode is simple and intuitive, and the greatest advantage is that the service requirement of the user can be fully met.

The dynamic allocation of the virtual machines is realized by using a virtual machine migration technology. When the resource requested by the virtual machine carried by the host exceeds the resource capacity limit, selecting part of the virtual machines from the host according to a certain virtual machine selection algorithm (such as random selection), and migrating the virtual machines to other hosts with enough resources, so as to ensure that all the virtual machines can obtain the required resource; in addition, when the amount of resources requested by the virtual machine carried by the host is very small, all the virtual machines of the host are migrated for the purpose of energy saving, so that the host can be shut down to reduce static power consumption. The most typical examples of the virtual machine allocation method using virtual machine migration as a means are First (FF), First critical Sum (FFDSum), and Round Robin (RR). These methods are often widely applied to various cloud platforms such as OpenStack because of their simplicity. Wherein the FF algorithm selects a first host with sufficient resources for the virtual machine to be migrated each time; the FFDSum considers the influence of different types of resources, and judges the resource occupation size of each host after the virtual machine is placed by setting the weight of the resource of each dimension and calculating the weight, namely, the host with the largest weight resource occupation amount is selected for placing each virtual machine; the RR algorithm is a very popular load balancing algorithm, which balances the load status of each host by selecting different hosts each time, thereby reducing the possibility of overloading the hosts due to the load fluctuation of the virtual machines. Different from the way that the RR algorithm guarantees the user service level, some researches implement refined virtual machine migration management and control by means of a single-step or multi-step long load prediction technology, and the main motivation for doing so is to prevent the host where the virtual machine is located from being overloaded twice after the virtual machine is migrated. These methods are referred to herein as PredBasedCoVM. On the other hand, in order to reduce the system energy consumption, the comp vm method researches the characteristics of load cycle line changes, and also comprehensively considers the complementarity of resources of different virtual machines in the space dimension (i.e., different resource dimensions) and the time dimension by a prediction means, thereby achieving the purposes of fully utilizing the host resources and reducing the number of the actually used hosts. The latest research work, PageRank vm, indicates that the key to reducing the number of active hosts is to accurately determine the probability of the current state of the host reaching a full resource utilization state (the best state), and proposes to use a PageRank web-ordering algorithm to define the transition relationships between host states and to compute the priority ranking of each other's states.

FIG. 1 depicts the basic workings of ordering host state in the PageRankVM method. Where each circle represents a host resource occupancy state and the arrows indicate that the host can transition from one state to the next by placing a virtual machine. In order to rank the host states, the PageRankVM consults the process of ranking web pages by the PageRank algorithm: the web page in the PageRank algorithm corresponds to the host state, and the link in the web page corresponds to the arrow in FIG. 1; at this time, the state rank at the back end of the arrow is affected by the previous hop state rank. A comparison of the PageRankVM method and the CompVM method shows that the PageRankVM has important significance in reducing energy consumption. The PageRankVM approach still has deficiencies.

Although the method for statically allocating virtual machines can effectively guarantee the service quality of users, the dynamic change of the load of the virtual machines causes that the methods cannot fully and effectively utilize the resources of the hosts, thereby causing resource waste, which also means that a large number of hosts are required to meet the resource requests of the virtual machines, thereby generating great energy consumption overhead; in addition, the methods such as Oktopus and Sync only consider the allocation of network resources, but in an actual environment, not only the network resources but also resources such as CPU, memory, and storage need to be considered, so that the practicability of these methods is also a great problem.

The dynamic allocation method is to improve the resource utilization rate of the host by using a virtual machine migration technology, which brings about many virtual machine migration overheads. In addition, the load states of the current virtual machine and the host machine are only considered in the classic FF and FFDSum, and the influence possibly generated by future load change is ignored, so that the defects in the aspects of energy conservation and user service quality guarantee are large; the RR algorithm sacrifices host resources for higher and better service experience, but it also ignores the dynamic factor of virtual machine load; although the PredBasedCoVM considers the load state of the virtual machine with a single step length or multiple steps for a long time, thereby better avoiding unnecessary migration and host overload, the PredBasedCoVM does not consider the energy consumption problem in the process of reallocating the virtual machine; the PageRankVM provides a very heuristic viewpoint on how to reduce the number of active hosts, but the PageRank algorithm is used for neglecting the influence of the host state when ordering the host state, so that a lot of unnecessary host overload and migration cost are generated. Moreover, the introduction of the PageRank algorithm can greatly increase the time complexity of sorting the host states, and is not suitable for processing the situation of dynamic load change of the virtual machine.

In summary, the existing methods cannot achieve a good balance among the overall energy consumption of the system, the user service quality and the virtual machine migration overhead; or the time complexity is high, and the method is difficult to be applied to the actual cloud computing system environment.

Disclosure of Invention

The invention provides a virtual machine allocation method (RHMM) based on rule scoring, which mainly aims at solving the problem that the existing virtual machine allocation method cannot effectively consider the relationship among user service quality, virtual machine migration overhead and energy consumption performance under the cloud computing environment of heterogeneous host capacity configuration and multi-type resources.

Aiming at the defects of the prior art, the invention provides a heuristic virtual machine allocation method based on rules for a cloud data center, which comprises the following steps:

step 1, a cloud data center comprises a plurality of hosts, and each host at least creates a virtual machine to execute the resource requirements of users;

step 2, collecting the utilization rate of each one-dimensional resource of the host at the current moment to obtain a host state vector, and collecting the utilization rate of each one-dimensional resource of the virtual machine at the current moment to obtain a virtual machine state vector;

step 3, judging whether the host is in an idle state, if so, setting the score of the host to zero, otherwise, obtaining the host state size of the host by weighting and summing the host state vector;

step 4, judging whether the residual resources of the host can accommodate additional virtual machines, if not, the host is in a termination state and the score of the host is set as the state size of the host, if so, the host is in an intermediate state and the probability distribution of the state size of the host multiplied by the resource request state of the virtual machine is set as the score of the host;

step 5, collecting the state and the score of each host to obtain a host state-score mapping table, acquiring the virtual machine to be migrated, and initializing an alternative active host list, an idle host list and a maximum host state score variable;

step 6, sequentially traversing each host in the cloud data center, and for any host, if the host is in an idle state, adding the host into the idle host list, otherwise, executing step 7;

step 7, judging whether the host is overloaded when the virtual machine to be migrated is migrated to the host or not, if not, calculating the score of the host after the virtual machine to be migrated is migrated to the host, and if the score is greater than the maximum host state score variable, emptying the standby active host list, adding the current host to the standby active host list as a standby, and updating the maximum host state score variable; otherwise, executing step 8;

step 8, if the score of the host is equal to the maximum host state score variable after the virtual machine to be migrated is migrated to the host, directly adding the current host to the candidate active host list;

and 9, if the size of the candidate active host list is not equal to 0, selecting a host with the largest state size from the candidate active host list as a virtual machine allocation result, and otherwise, selecting a host from the idle host list as a virtual machine allocation result.

In the rule-based heuristic virtual machine allocation method for the cloud data center, the host state vector in step 2 includes: CPU utilization, memory utilization, disk utilization, and network bandwidth.

The heuristic virtual machine distribution method based on the rules of the cloud data center is characterized in that the virtual machine to be migrated is obtained by randomly selecting one virtual machine from an overloaded host until the host is not overloaded.

The heuristic virtual machine allocation method based on the rules of the cloud data center is characterized in that the step 3 comprises the following steps:

the host state Size (P) is obtained by the above formula_i(t)), wherein α_dWeight, P, of resource in d-th dimension_i ^d(t) to represent the host P_iAnd D is the total resource type number.

The heuristic virtual machine allocation method based on the rules of any cloud data center is characterized in that the step 3 further comprises a host isomorphism processing step:

traversing all types of hosts, for each dimension of resources, finding the maximum resource capacity Max^d)：

{Max.Cap(P^d)|d∈[1,D]}

Max.Cap(P^d) The virtual capacity of the host on the resource of the D-th dimension is shown, and D is the total resource type number;

under the condition of considering the resource capacity of the host of the virtual machine, the virtual resource occupation condition of each host is obtained by recalculation, { Max^d)-Cap(P_i ^d)|d∈[1,D]And determining the state of each host according to the resource occupation condition.

The invention also provides a heuristic virtual machine distribution system based on rules for the cloud data center, which comprises the following steps:

the module 1, the cloud data center includes a plurality of host computers, and each host computer has at least one virtual machine created, in order to carry out the resource demand of the user;

the module 2 is used for collecting the utilization rate of each one-dimensional resource of the host at the current moment to obtain a host state vector, and collecting the utilization rate of each one-dimensional resource of the virtual machine at the current moment to obtain a virtual machine state vector;

module 3, judge whether the host computer is in idle state, if yes, set the score of the host computer to zero, otherwise through weighting and summing the host computer state vector, get the host computer state size of the host computer;

module 4, judge whether the surplus resource of the host can hold the extra virtual machine again, if not, the host is the termination state and set the score of the host as the size of the host state, if, the host is the intermediate state and the probability distribution of the size of the host state multiplied by the resource request state of the virtual machine is set as the score of the host;

the module 5 is used for collecting the state and the score of each host to obtain a host state-score mapping table, acquiring a virtual machine to be migrated, and initializing an alternative active host list, an idle host list and a maximum host state score variable;

module 6, sequentially traversing each host in the cloud data center, and for any host, if the host is in an idle state, adding the host into the idle host list, otherwise, executing module 7;

module 7, judge if migrating the virtual machine to be migrated to the host will cause the host overload, if not, calculate the score of the host after migrating the virtual machine to be migrated to the host, if greater than the maximum host state score variable, empty the candidate active host list, and add the current host to the candidate active host list as the candidate, and update the maximum host state score variable; otherwise, the module 8 is executed;

module 8, if the score of the host is equal to the maximum host state score variable after the virtual machine to be migrated is migrated to the host, directly adding the current host to the candidate active host list;

and 9, if the size of the candidate active host list is not equal to 0, selecting a host with the largest state size from the candidate active host list as a virtual machine allocation result, otherwise, selecting a host from the idle host list as a virtual machine allocation result.

The heuristic virtual machine distribution system based on rules of the cloud data center is characterized in that the host state vector in the module 2 comprises: CPU utilization, memory utilization, disk utilization, and network bandwidth.

The heuristic virtual machine distribution system based on the rules of the cloud data center is characterized in that the virtual machine to be migrated is obtained by randomly selecting one virtual machine from an overloaded host until the host is not overloaded.

The heuristic virtual machine distribution system based on rules of the cloud data center is characterized in that the module 3 comprises:

The heuristic virtual machine distribution system based on the rules of any cloud data center is characterized in that the module 3 further comprises a host isomorphism processing module:

{Max.Cap(P^d)|d∈[1,D]}

According to the scheme, the invention has the advantages that:

when the host computer is overloaded or underloaded and migrated in the cloud computing system, the virtual machine selects a proper target host computer according to the virtual machine placement algorithm provided by the invention. After the system runs for a period of time, compared with a classical method and a latest method, the system power consumption under the RHMM virtual machine allocation method can keep the same level basically, and the performance is greatly improved on two indexes of virtual machine migration times and service level grade violation degrees.

Drawings

FIG. 1 is a diagram of the ordering relationship of different host states in the PageRankVM method;

FIG. 2 is a diagram of a host state scoring algorithm;

FIG. 3 is a diagram of a virtual machine placement algorithm;

FIG. 4 is a diagram of a host state transition diagram.

Detailed Description

If the usage quantity of the active hosts is effectively controlled and the user service quality is guaranteed at the same time, the core of the method lies in how to process the relationship between the current and future load states of the virtual machines and the host states. The invention aims to avoid violating the service level of the user as much as possible from the viewpoint of energy saving. In addition, these methods also need to be able to handle the problem caused by heterogeneous host resource capacity and multiple types of resources, and the PageRank vm allocation mechanism proposes a concept of measuring the probability that the host state reaches the optimal state, but the proposed method of ranking different host states by measuring the probability through the PageRank algorithm still causes many unnecessary migration overheads and causes the degradation of user service quality. The main reason for this problem is that the core idea of the PageRank algorithm is as follows: the more authoritative sites that link to the website, the higher the trustworthiness or value, the higher the ranking of the website. By mapping it to a virtual machine allocation problem, the more recent (i.e., more resources occupied) host states tend to be ranked higher, so that the host is more likely to be the target of virtual machine migration. Although this can effectively reduce the number of used hosts, it also causes unnecessary host overload and virtual machine migration problems. Moreover, to accurately determine the probability of the host state reaching the optimal state, the probability distribution of the virtual machine resource request state is also important, which directly determines how likely a host is to transition to the next state. The calculation process of the probability distribution is, for example: assuming that there are 10 virtual machines, of which 4 are a and 6 are b, the distribution probability of the virtual machine state a is (4/10 ═ 0.4), and the distribution probability of the virtual machine state 6 is (6/10 ═ 0.6)

Therefore, the inventor further analyzes the relationship between each state of the host and the state of the virtual machine on the basis of the PageRankVM method, and focuses on how to accurately rank each state of the host. Moreover, the inventor elaborates the processing method of the invention under the environment of heterogeneous hosts with various types of resources.

The invention provides a rule-based heuristic virtual machine allocation method (RHMM), which is mainly applied to a cloud data center and used for guiding a virtual machine to be migrated to a host, and the definition of a heuristic algorithm is as follows: an algorithm based on intuitive or empirical construction. The heuristics below refer to virtual machine placement algorithms. That is, a virtual machine selects one host at a time for placement, and the score for this host after placement of this virtual machine is highest relative to the other hosts. Firstly, all possible host states are scored, and the scoring is based on the core rule that the score of each state depends on the score of the state to which the state can be transferred and the probability distribution of the corresponding virtual machine state; then, when the virtual machine migration occurs, the virtual machine selects a host with the highest state score after being placed for migration. Where a state transition refers to the process of one host changing to another resource utilization state by placing a virtual machine in a particular resource utilization state.

The key points of the application include:

1. in order to enable the virtual machine allocation algorithm to be also suitable for the heterogeneous host resource capacity scene, the invention firstly proposes to perform isomorphic processing on the host capacity; after the isomorphic processing, each host can be regarded as having the same resource capacity configuration, and the processed host is regarded as having partial initial resource occupation.

2. The grading rules of the three host states and the corresponding grading algorithm are provided by combining the transfer relation among the host states; each host resource occupation state corresponds to a score, and the key is that the score of the current state is equal to the weighted sum of the host state score which can be reached by the current state and the arrival probability of the current state.

3. A virtual machine placement method; when the virtual machine selects the target host, all the hosts preferentially carry out isomorphism processing according to the current resource occupation state of the hosts, and then the virtual machine selects a host with the highest host placement state score to carry out migration or placement.

In order to make the aforementioned features and effects of the present invention more comprehensible, embodiments accompanied with figures are described in detail below.

1. Defining host states and virtual machine states

First, the host state and the virtual machine state need to be defined. P for the present invention_i ^d(t) to represent the host P_iAt time t, the utilization rate on the d-th dimension resource, then the host state can be represented as the vector { P }_i ^d(t)|d∈[1,D]D is the total number of resource types, and the resource types include: CPU utilization, memory utilization, disk utilization, network bandwidth, and the like. Considering P_i ^d(t) is a group of [0,1]Real number between, the invention pairsIt is discretized. The present invention divides the resource utilization rate of each dimension of the host at 5% intervals (the size of the interval can be determined according to specific performance requirements), that is, each dimension of the resource can be divided into { 0%],(5％,10％],…,(95％,100％]Where 0 is a special interval indicating that the current host is in an idle (off) state. For the convenience of subsequent calculation, each resource utilization interval is represented by {0, 5%, 10%, …, 100% }, then the host state space size of a D-dimensional resource is 21^D. Similar to the host state, the invention uses

Virtual machine representation V_jThe value range of the request size of the d-dimension resource at the time t is also [0,1 ]]Then the virtual machine state can be represented as a vector

For convenience, 1% is used as

Is the smallest unit of measure of (1), that is to say if

It will be in accordance with the formula

Converted to 2%. Doing so also implies proper over-allocation of the virtual machines, helping to improve their quality of service. Used in the invention

To indicate the virtual machine initialization state, which means the amount of resource utilization requested at the time of virtual machine initialization. The meaning of the virtual machine state vector is the amount of resources that the virtual machine is currently actually requesting. When one state is { P_i ^d(t)|d∈[1,D]The host computer of the station is put in a state of

The state of the host machine is changed to

Thus, that is, the subsequent involved host state changes require the participation of the virtual machine state vector.

2. Host isomorphism processing

Considering that in a practical application scenario, there may be many hosts with different capacity configurations, such as two heterogeneous server hosts existing on amazon cloud platform: the CPU and memory configurations of M3 and C3 are (2.6 GHz, 8 cores, 64GB memory) and (2.8 GHz, 8 cores, 75GB memory). In this case, the host resource capacities need to be uniformized. The invention is processed according to the following steps:

step 1: cap (P) is found out for the maximum resource capacity of each dimension of resources by traversing all types of hosts^d). Cap (P) is called Max in the invention^d) Obtaining a maximum value vector (virtual host resource capacity) of each dimension resource for the virtual capacity of the host on the resource of the d dimension: cap (P)^d)|d∈[1,D]}；

Step 2: under the condition of considering the resource capacity of the host of the virtual machine, recalculating the virtual resource occupation condition of each host, namely { Max^d)-Cap(P_i ^d)|d∈[1,D]}。

For example, the following steps are carried out: assuming that there are two types of idle hosts and their respective resource capacities are [4,3] and [3,5] considering only two dimensions of resources, the virtual host resource capacity obtained after processing is [4,5 ]. At this time, the virtual resource occupation situations of the two types of hosts are called as [0,2] and [1,0], respectively.

3. Host state scoring algorithm

On the basis of the above-defined host states, the present invention further defines four different types of host states: zero state, intermediate state, end state, optimal state. The zero state indicates an idle host state, and it should be noted that although some zero state hosts are going through isomorphismThe virtual resource occupation condition may exist after the chemical processing, but the virtual resource occupation condition is still called as a zero host state; the intermediate state can be changed into another state by placing a virtual machine, and it should be noted that when a host cannot accommodate a virtual machine, the host is said to be changed into the state of the host itself under the current virtual machine, for example, the maximum capacity of the host is 6, and the host state with the size of 3 is considered, and then the host can be changed into the state 6 (the allowed resource capacity 6 is not exceeded) by placing a virtual machine with the size of 3; however, if a virtual machine with a size of 4 comes, if the virtual machine is placed on the host, the virtual machine will tend to exceed the allowed resource range of the host, so that the virtual machine cannot be placed on the host, and other hosts need to be searched, namely the host is in the virtual machine with the state of 4, and the state of the host remains unchanged; when the remaining resources of the host can not accommodate any additional virtual machine any more, the host state is called as a termination state; the best state is a special termination state, meaning that all resources of the host are fully utilized. In order to reduce the number of active hosts, the present invention expects the hosts to be able to maintain an optimal state for as long as possible during the virtual machine resource adjustment. The host state Size (P) is obtained by the following formula_i(t)), wherein α_dThe weight of the resource representing the d-th dimension is set to 1 in this embodiment, and may also be adjusted according to actual needs:

based on these four different states, scoring the host state should comply with the following three rules:

rule 1: the score of the end state depends on its own state Size (P)_i(t))；

Rule 2: the score of an intermediate state depends on the Size (P) of the state it can reach_i(t)) and its corresponding arrival probability magnitude, i.e., the virtual machine state probability distribution. The former means that the score closer to the optimal host state should be higher, and the latter means that the host state score is also limitedThe distribution probability of the host state of the virtual machine.

Rule 3: a zero state score of 0 indicates that the virtual machine will prefer an active host over a newly-started host.

Based on these rules, the present invention proposes a host state scoring algorithm (shown as pseudo-code in FIG. 2). The algorithm inputs a global mapping scoreMap of host state-score, an initialized zero host state S, a set of virtual machine states S_VAnd its corresponding virtual machine state probability distribution set P_r(S_V) (ii) a The scoreMap is output, which contains each host state and its score. The specific working method of the algorithm is as follows:

step 1 (line 1): it is determined whether the current state is one of the key values of scoreMap. If yes, the state is processed and returned directly;

step 2 (lines 3-7): for each dimension of resources, a new host state is obtained by adding a value of delta (delta represents the size of the host resource utilization interval, 5%), and the new host state is taken as the input of the algorithm to continue recursive calling, so that the algorithm can generate all effective host states through the recursive calling process;

step 3 (lines 8-9): if the current state is a zero state, setting the state score to 0; otherwise, further executing step 4;

step 4 (line 11): initializing a nextHops variable for recording the number of states which can be reached by the current host state after the virtual machine is placed;

step 5 (lines 12-13): the current state is obtained by accumulating a virtual machine state S_vA new state S of the host can be obtained_next；

Step 6 (lines 14-17): if the host state S_nextLess than or equal to the host capacity limit (after isomorphism, the capacity limit should be the virtual host resource capacity), and the score of the current host state S is added up to S_nextScore multiplied by S_vAnd nextHops add 1 cumulatively; otherwise, further executing step 7;

step 7 (line 19): the score of the current host state S is summed with its own state size multiplied by S_vThe distribution probability of (2);

step 8 (lines 20-21): if the nextHops value is 0, the host with the current state of S can not accommodate any existing virtual machine, namely the state S is a termination state, and at the moment, the score of S is assigned to the state size of the host;

step 9 (line 22): the current state S and its corresponding score are added to scoreMap.

The algorithm comprises two parts in total, namely, the construction of a host state transition diagram and the scoring of the host state. The present invention further illustrates this process with an example, as shown in fig. 4. Suppose there are currently two different types of virtual machine states S_vWith {3,4}, the two virtual machine state distributions are 0.5 and 0.5, respectively, so that the host states {0,3,4,6,7,8,9} can be obtained. In this example, assume that s (x) represents a score for a host state of x: the first step is as follows: the state scores of 7,8, and 9 are calculated, and the state scores are represented by the state Size (calculated from Size (pi (t)) according to rule 1), and therefore, the state scores are 7,8, and 9, respectively; the second step is that: the algorithm will calculate the scores of

states

6, 4,3 in turn; state 6 score: according to rule 2, state 6 can become state 9 by placing virtual machine 3, but virtual machine state 3 has only a probability of 1/2, so state 6 can only have a probability of 1/2 to transition to state 9, i.e., can only inherit 1/2 of the state 9 score; on the other hand, a state 4 virtual machine cannot be placed on the host (beyond capacity 9), which means that state 6 remains unchanged under state 4 virtual machine with a probability of 1/2, as shown in algorithm step 7, and therefore inherits a 1/2 score from its own size (6). So the score for state 6 is: s (6) ═ 1/2 × S (9) +1/2 × Size (state 6) ═ 1/2 × 9+1/2 × 6 ═ 7.5; score for state 4: state 4 may be converted to state 7 by placing the virtual machine in state 3, or may be converted to state 8 by placing the virtual machine in state 4, and the distribution probabilities of virtual machines states 3 and 4 are both 1/2, so: s (4) ═ 1/2 × S (7) +1/2 × S (8) ═ 7.5, the same way as: s (3) ═ 1/2 × S (6) +1/2 × S (7) ═ 6.75, S (0)) 0, rule 3.

According to the host state scoring algorithm, the scores of the

states

7,8 and 9 can be obtained as the sizes of the states of the host, wherein the scores are respectively 7,8 and 9; state 6 may transition to 9 in virtual machine state 3 and remain unchanged in virtual machine state 4, i.e., its score is (9+ 6)/2-7.5; similarly, we can get scores of 6.75 and 7.5 for host states 3 and 4, respectively; the score for state 0 is 0.

4. Virtual machine placement algorithm

After the host state transition diagram is constructed and the host state score is obtained, the invention provides a corresponding virtual machine placement algorithm (shown as pseudo code in fig. 3) according to which the virtual machine selects the optimal migration destination host. The algorithm inputs a host state-score mapping table scoreMap and a virtual machine V to be migrated_kThe output destination host operates in detail as follows, wherein the virtual machine V to be migrated_kThis may be achieved by randomly selecting a virtual machine from the overloaded host until the host is not overloaded or all virtual machines on the underloaded host are considered as virtual machines V to be migrated_k：

Step 1 (line 1): initializing an alternative active host list candidateList, an idle host list idlehastList and a maximum host state score variable maxScore;

step 2 (lines 3-4): traversing each host in turn, for any host PM_iIf the host is a free host, add it to idleHostList; otherwise, checking whether the host is an overloaded host or an underloaded host, and if not, entering the step 3; the host overload means that: the resources actually requested by all virtual machines on the host exceed the host's allowed resource capacity. If multiple resource types are considered, the host is considered overloaded as long as one of the resources (such as memory) exceeds the sum of the actual requests of all the virtual machines on the host. Conversely, it can be understood that the actual amount of requests on all types of resources of the host is low, for example, each resource has a capacity of 1, and then the host underrun can be determined as follows: i.e. the sum of the virtual machine requests is below 20% for each resource.This is an empirical value of 20%.

Step 3 (lines 6-7): obtaining the state S of the current host under the resource capacity of the virtual host according to the isomorphism processing of the host, and calculating the state S in the virtual machine state S_VkNew state S under action_nextIf S is_nextIf not, entering step 4, otherwise, continuing to execute the for loop of the second line, and judging whether the next host is suitable;

step 4 (lines 9-12): if S is_nextIf the score of (1) is higher than the current maximum score maxScore, the candidate list canditeList is cleared and the current host PM is assigned_iAdding the candidate list as an alternative and updating the maximum score maxsore; otherwise, entering step 5;

step 5 (lines 13-14): if S is_nextIf the state score is equal to the currently known maximum score, directly adding the current host to the candidate list;

step 6 (lines 15-17): if the candidate list size is not equal to 0, there is an active host in which to place virtual machine V_kIf the host has the highest score, selecting the host with the largest state size from the list for output; otherwise, entering step 7;

step 7 (lines 19-21): by this it is meant that no active host is able to accommodate V_kThen, try to select the first host with enough resources from the idlehost list, and output the return;

step 8 (line 22): if again no host can accommodate V in idleHostList_kThen it indicates that no proper placement method exists and exits.

The present invention also takes fig. 4 as an example to illustrate how a virtual machine selects a host under the virtual machine placement algorithm. The state S of the virtual machine can be obtained according to a host state scoring algorithm_vWhen the probability distributions each account for 1/2, {3,4}, scoreMap may be expressed as { 0: 0,3: 6.75,4: 7.5,6: 7.5,7: 7,8: 8,9: 9}. Assuming that there is currently a virtual machine state of 3 and the optional host state is {0,3,4, 6}, since 6+3 is 9, 9 state scores 9, and the score is highest, there is a virtual machine state of 3, and the optional host state is {0,3,4, 6}, the virtual machine state is a virtual machine state of 9, and the virtual machine state of 9 scores 9, and thus the virtual machine state of the virtual machine state is the highestThe virtual machine is migrated to the host machine with state 6; similarly, if the current virtual machine state is 4, it will migrate to a host with a host state of 4, rather than a host with a state of 3.

The following are system examples corresponding to the above method examples, and this embodiment can be implemented in cooperation with the above embodiments. The related technical details mentioned in the above embodiments are still valid in this embodiment, and are not described herein again in order to reduce repetition. Accordingly, the related-art details mentioned in the present embodiment can also be applied to the above-described embodiments.

the host state Size (P) is obtained by the above formula_i(t)), wherein α_dWeight, P, of resource in d-th dimension_i ^d(t) to represent the host P_iThe utilization rate of the D-dimensional resource at the time t, D is the total resource type number。

{Max.Cap(P^d)|d∈[1,D]}

Claims

1. A heuristic virtual machine allocation method based on rules of a cloud data center is characterized by comprising the following steps:

step 4, judging whether the residual resources of the host can accommodate additional virtual machines, if not, the host is in a termination state and the score of the host is set as the state size of the host, if so, the host is in an intermediate state, and the state size of the host is multiplied by the probability distribution of the resource request state of the virtual machine to be used as the score of the host;

step 9, if the size of the candidate active host list is not equal to 0, selecting a host with the largest state size from the candidate active host list as a virtual machine allocation result, otherwise, selecting a host from the idle host list as a virtual machine allocation result;

the step 4 further comprises:

step 41, dividing the number of virtual machines in the host in the resource request state of each virtual machine by the total number of virtual machines in the host to obtain the probability distribution of the resource request state of each virtual machine;

step 42, adding the state size of the host to the score of the virtual machine to obtain a plurality of first intermediate values, sequentially judging whether the first intermediate values are larger than a preset value, if so, multiplying the state size of the host by the probability distribution of the virtual machine resource request state corresponding to the score of the virtual machine to obtain a second intermediate value, otherwise, multiplying the first intermediate values by the probability distribution of the virtual machine resource request state corresponding to the score of the virtual machine to obtain a second intermediate value;

step 43, summing all the second intermediate values as the score of the host.

2. The method as claimed in claim 1, wherein the host state vector in step 2 comprises: CPU utilization, memory utilization, disk utilization, and network bandwidth.

3. The method of claim 1, wherein the to-be-migrated virtual machine is obtained by randomly selecting a virtual machine from overloaded hosts until the host is not overloaded.

4. The method for heuristically assigning virtual machines based on rules for a cloud data center of claim 1, wherein the step 3 comprises:

5. The method for distributing the heuristic virtual machine based on the rules in the cloud data center according to any one of claims 1 to 4, wherein the step 3 further comprises a host isomorphism processing step of:

{Max.Cap(P^d)|d∈[1,D]}

recalculating while taking into account virtual machine host resource capacityGet the virtual resource occupation of each host, { Max^d)-Cap(P_i ^d)|d∈[1,D]And determining the state of each host according to the resource occupation condition.

6. A heuristic virtual machine distribution system based on rules of a cloud data center is characterized by comprising:

module 9, if the size of the candidate active host list is not equal to 0, selecting a host with the largest status size from the candidate active host list as a virtual machine allocation result, otherwise selecting a host from the idle host list as a virtual machine allocation result;

the module 4 further comprises:

the module 41, the number of virtual machines in the host in the resource request state of each virtual machine is divided by the total number of virtual machines in the host to obtain the probability distribution of the resource request state of each virtual machine;

the module 42 adds the state size of the host to the score of the virtual machine to obtain a plurality of first intermediate values, and sequentially judges whether the first intermediate values are larger than a preset value, if so, the state size of the host is multiplied by the probability distribution of the resource request state of the virtual machine corresponding to the score of the virtual machine to obtain a second intermediate value, otherwise, the first intermediate values are multiplied by the probability distribution of the resource request state of the virtual machine corresponding to the score of the virtual machine to obtain a second intermediate value;

and a module 43 for summing all the second intermediate values as the score of the host.

7. The rule-based heuristic virtual machine allocation system of cloud data center of claim 6, wherein the host state vector in module 2 comprises: CPU utilization, memory utilization, disk utilization, and network bandwidth.

8. The rule-based heuristic virtual machine allocation system of the cloud data center of claim 6, wherein the virtual machine to be migrated is obtained by randomly selecting a virtual machine from an overloaded host until the host is not overloaded.

9. The rule-based heuristic virtual machine distribution system of cloud data center of claim 6, wherein the module 3 comprises:

10. The rule-based heuristic virtual machine distribution system of the cloud data center of any of claims 6-9, wherein the module 3 further comprises a host isomorphism processing module:

{Max.Cap(P^d)|d∈[1,D]}