CN104580518A - Load balance control method used for storage system - Google Patents

Abstract

The invention discloses a load balance control method used for a storage system. The load balance control method includes steps of (1) encoding by a multi-branch tree structure; (2) randomly generating an initial population comprising a plurality of individuals; (3) determining whether iteration end conditions are met or not, if not, entering a step (4), and if yes, outputting results; (4) calculating adaptability of each individual; (5) selecting the individual of the next step according to a roulette selecting method; (6) subjecting the selected individual to crossover operation; (7) subjecting gene positions of the population to heteromorphosis; (8) removing the individuals with the minimum adaptability in the population, guaranteeing the scale of the population, generating a new population and returning to the step (3). By applying tree code rules to the mapping relation of physical nodes and virtual nodes and redesigning crossover operators and heteromorphosis operators in genetic algorithm, a mapping scheme meeting load balance is obtained, system output, disk utilization efficiency and user access service quality are remarkably improved.

Description

A kind of control method for equalizing load for storage system

Technical field

The present invention relates to a kind of control method for equalizing load for storage system

Background technology

Along with increasing smart machine accessing Internet, cause the data volume on network increasing.Up to the present, on line, data are also with the mode rapid growth of Moore's Law, therefore must bring great challenge for the store and management of data.Distributed storage technology is solution popular at present, by merging the functions such as Clustering, distributed computing technology and file system, realizes the collaborative work between numerous memory device, and externally access distinct device provides unified access module.Distributed memory system is generally made up of physical storage layer, basic platform management level, application-interface layer and user's access layer, and wherein basic management layer is the part storing core.The functions such as basic management layer primary responsibility data storing, redundancy backup, data consistency, load balancing, High Availabitity, the load of its interior joint is a very important module.So-called load imbalance refers to that Data distribution8 between each memory node is uneven, the load imbalance between node will inevitably affect the throughput of system of whole storage system and response time.

Solving load balancing between Mass storage node is efficiently a more difficult problem, and the dynamic data that current mainstream technology comprises based on distributed computing technology is deposited and data tape technology.Such as: document " Ganger G R et al.Disk Subsystem Load Balancing:Disk striping vs Conventionaldata placement [C] .Proceedings of the 26th Hawaii International Conference onSystem Sciences, LosAlamitos:IEEE CS Press, 1993:40-49 " think that traditional dynamic data strategy easily causes disk size to tilt, because it requires to be forced to accept atomic data under quick change access module, thus cause load imbalance, disk tape is finally adopted to be turned to supplementary means, document " R.J.Honicky, E.L.Miller.Replication Under Scalable Hashing:A Family ofAlgorithms for Scalable Decentralized Data Distribution.the18th InternationalParallel and Distributed Processing Symposium (IPDPS2004), Santa Fe, NM, April2004 " propose RUSH race algorithm, based on the Data distribution8 of decentralization, the situation of load imbalance is there is mainly for file system system when increasing or delete data, different RUSH algorithm mutation has different features, respectively with solving different situations, document " Ni Yunzhu, Lv Guanghong, Huang Yanhui. solve based on the disk problem of load balancing [J] of disk-striping strategy by genetic algorithm. Chinese journal of computers " propose and adopt genetic algorithm to solve disk Dynamic Load-Balancing Strategy based on tape technology, mainly have employed tape technology to divide file and the file allocation algorithm for realizing load balancing, and the temperature of disk is analyzed, finally obtains best solution according to genetic algorithm, document " Dong Huanqing, Li Zhanhuai. based on the method [J] of disk load balancing in the RAID disk array of genetic algorithm. computer engineering and application " the IO feature of the logic magnetic disc of RAID disk array is analyzed, analyze according to the mapping between logic magnetic disc and physical disk and loading condition again on this basis, finally carry out load balance process from the inside of memory node, propose one memory node internal data migration fast scheme.

The method that above-mentioned document adopts can realize certain effect, but the high-efficient carrier being all difficult to realize memory node is balanced, is particularly difficult to the load reducing frequent access node.

Summary of the invention

Object of the present invention is just to provide a kind of control method for equalizing load for storage system that can reduce frequent access node load to solve the problem.

The present invention is achieved through the following technical solutions above-mentioned purpose:

For a control method for equalizing load for storage system, comprise the following steps:

(1) according to the mapping feature of system model physical node and dummy node, employing Multiway Tree Structure is encoded, the corresponding one tree of whole system structure, nodes all on the corresponding root node to leaf node path of gene position, root node and level 0 are defined as system, and ground floor represents physical node, and the second layer represents disk, the corresponding disk piecemeal of third layer, the 4th layer of corresponding dummy node;

(2) stochastic generation contains the initial population of several individualities;

(3) determine whether to meet iteration termination condition, if do not met, enter step (4), if met, Output rusults;

(4) according to the fitness of each individuality of following formulae discovery:

$F\left(s\right)=\frac{1}{M+N.f\left(s\right)}$

Wherein, F (s) fitness function, M, N are constant, and f (s) is determined by following formula:

$f\left(s\right)=d\left(s\right)=\sqrt{\frac{1}{\mathrm{MN}}\underset{i=1}{\overset{\mathrm{MN}}{\mathrm{\Σ}}}{(G_{\mathrm{ij}}-\stackrel{\‾}{\mathrm{AGL}})}^2}=\sqrt{\frac{1}{\mathrm{MN}}\underset{i=1}{\overset{\mathrm{MN}}{\mathrm{\Σ}}}{(G_{\mathrm{ij}}-\frac{S}{\mathrm{MN}})}^2}$

Wherein, G _ijrepresent the overall load of any one piece of disk in any one physical node, represent the average load of any one piece of disk, S represents total use amount of whole system;

(5) according to next step individuality of roulette selection method choice;

(6) crossing operation is carried out to the individuality selected;

(7) variation process is carried out to the gene position of population;

(8) remove the individuality that in population, fitness is minimum, guarantee the scale of population, produce new population, and return step (3).

Above-mentioned steps (1)-(8) come from genetic algorithm but are different from genetic algorithm, genetic algorithm is that Artificial smart field is for solving a kind of efficient, parallel search heuritic approach of optimization problem, use for reference the natural selection of living nature and natural genetic mechanism, there is the ability of the fast search overall situation.Genetic algorithm simulating nature select and occur in natural genetic process breeding, intersection and gene mutation phenomenon, all retain one group of candidate solution in each iteration, and it is preferably individual from treating to choose selected population by certain index, genetic operator (selection, crossover and mutation) is utilized to combine these individualities, produce the candidate solution group of a new generation, repeat this process, until meet certain convergence index.The present invention utilizes genetic algorithm to produce the solution meeting mapping condition, thus makes the load relative equilibrium of whole storage system.Classical genetic algorithm adopts binary coding, adopts character set 0,1 to form chromosome, to gene direct modeling, but there is mapping error.According to the feature of the corresponding relation of physical node, disk and dummy node, the present invention adopts Multiway Tree Structure to encode.

As preferably, in described step (5), described roulette selection method comprises the following steps:

1. select probability is calculated as follows:

$p_i=\frac{F(x,i)}{\underset{j=1}{\overset{N}{\mathrm{\Σ}}}F(x,j)}$

Wherein, p _ithe probability being genetic to colony of future generation for individual i is selected, F (x, i) is the fitness function of individual i;

2. according to select probability, be divided into by disk N number of fan-shaped, each fan-shaped central angle is 2 π p _i, in a disk border random selecting reference point, rotary disk, corresponding fan-shaped of final reference point is exactly the individual x be selected _i;

3. cumulative x _ithe number of times be selected;

2. and 3. N time 4. constantly repeat above-mentioned steps, can individuality be selected.

In described step (6), described crossing operation comprises the following steps:

A, from population, select two intersect individual, be expressed as S ₁and S ₂;

B, for S _iany one physical node, be denoted as namely a vector of physical node, disk and this three of dummy node is represented;

C, P _iall mappings are arranged in order according to virtual label order, are expressed as T _i;

D, re-establish physical node, relation between disk and dummy node, T _ielement be iteratively corresponding in turn to physical node P _iall disks in.

In described step (7), described variation refers to that some genic value in individual UVR exposure string is replaced with other genic value by foundation mutation probability, thus the individuality that formation one is new, described variation process comprises the following steps:

A, according to random selection two physical nodes of mutation probability;

B, when ensureing that dummy node mapping position is different, from these two physical nodes, selecting a disk respectively, these two disks are exchanged.

The determination of above-mentioned control method for equalizing load gets based on following analysis:

First, redundant data distribution rule is analyzed:

In present distributed storage system, usually adopt cheap PC to build bottom memory node, node failure is considered as normality, and in supposing the system, M physical node is expressed as P={P ₁, P ₂,, P _m, P _i(i≤M) represents i-th physical node, N number of disk on each physical node, P _ion disk be expressed as D _i={ D _i1, D _i2..., D _iN, D _ij(j≤N) represents P _ion a jth disk, the storage capacity of each disk is identical, represents with Q.M _ithe N of individual physical node _ithe use size of block disk is U _ij(i≤M, j≤N), each dummy node corresponding L block disk, V _i(l < Y), lays respectively at different physical nodes, and every block dish can corresponding multiple dummy node.A lot of storage system, in order to redundancy backup, being selected each data block to keep three parts, in order to reduce the complexity of algorithm, being got L=3.

According to foregoing description, be made up of the disk be positioned on different physical node as long as meet a dummy node, such as, dummy node V _iby laying respectively at P ₁on D _1N, P ₂on D ₂₃and P _mon D _mnblock disk is formed.Node which disk is made up of actually and decides according to the rule of default and load.

Secondly, physical node disk load analysis:

Consider the size fixation of C that dummy node takies in each disk, so each disk can a corresponding Q/C dummy node.At any one physical node P _iany one piece of disk D _ija corresponding bitmap T _ij, the dummy node that minute book local disk is corresponding, so disk D _ijuse size be:

$U_{\mathrm{ij}}=C\underset{k=0}{\mathrm{\&Sigma;}}{v}_k,v_k\&Element;T_{\mathrm{ij}}$

The use size of single physical node is the use size sum of each disk, uses PU _irepresent:

$P{U}_i=\underset{j=0}{\overset{N}{\mathrm{\&Sigma;}}}{U}_{\mathrm{ij}}=C\underset{j=0}{\overset{N}{\mathrm{\&Sigma;}}}\underset{k=0}{\mathrm{\&Sigma;}}{v}_k$

In order to calculate any one disk D _ijoverall load, total use amount of whole system must be calculated _s:

$S=\underset{i=1}{\overset{M}{\mathrm{\&Sigma;}}}{\mathrm{PU}}_i=\underset{i=1}{\overset{M}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{N}{\mathrm{\&Sigma;}}}{\mathrm{PU}}_{\mathrm{ij}}=C\underset{i=1}{\overset{M}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{N}{\mathrm{\&Sigma;}}}\underset{k=1}{\mathrm{\&Sigma;}}{v}_k$

So can calculate the overall load G of any one piece of disk in any one physical node _ij:

$G_{\mathrm{ij}}=\frac{U_{\mathrm{ij}}}{S}=\frac{C\underset{k=0}{\mathrm{\&Sigma;}}{v}_k}{C\underset{i=1}{\overset{M}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{N}{\mathrm{\&Sigma;}}}\underset{k=1}{\mathrm{\&Sigma;}}{v}_k}=\frac{\underset{k=0}{\mathrm{\&Sigma;}}{v}_k}{\underset{i=1}{\overset{M}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{N}{\mathrm{\&Sigma;}}}\underset{k=1}{\mathrm{\&Sigma;}}{v}_k}$

To the average load of any one piece of disk be so

According to above analysis, can determine a kind of standard whether load of the disk of whole storage system is balanced that judges, the mean square deviation of disk is expressed as:

$d\left(s\right)=\sqrt{\frac{1}{\mathrm{MN}}\underset{i=1}{\overset{\mathrm{MN}}{\mathrm{\&Sigma;}}}{(G_{\mathrm{ij}}-\stackrel{\&OverBar;}{\mathrm{AGL}})}^2}=\sqrt{\frac{1}{\mathrm{MN}}\underset{i=1}{\overset{\mathrm{MN}}{\mathrm{\&Sigma;}}}{(G_{\mathrm{ij}}-\frac{S}{\mathrm{MN}})}^2}$

If a kind of mapping scheme can be found, make d (s) minimum, so just can ensure load balancing between node.

To sum up, innovation of the present invention is just to find more reasonably mapping scheme, makes d (s) minimum.

Beneficial effect of the present invention is:

The present invention is by the research and analysis to main flow storage system framework, tree-like coding rule is applied in the mapping relations of physical node and dummy node, redesign the crossover operator in genetic algorithm and mutation operator, finally set up mapping model and Load Balancing Model that a kind of new storage system and genetic algorithm combine, draw a kind of mapping scheme meeting load balancing, and then carry out Data Migration, significantly improve the throughput of system, disk utilization and user's access services quality, solve the problem of load imbalance between storage node in distributed system, save industrial cost.After storage system is used control method for equalizing load of the present invention, the load of memory node can be at equilibrium by optimization, and the access load of all disks changes thereupon, and the load of the node of the frequent access of part reduces greatly.

Accompanying drawing explanation

Fig. 1 is physical node of the present invention, mapping relations schematic diagram between disk and dummy node;

Fig. 2 is the overall memory load schematic diagram of disk in embodiment;

Fig. 3 is iterations contrast schematic diagram in embodiment;

The memory load schematic diagram of disk after innovatory algorithm is introduced when Fig. 4 is β=0.3 in embodiment;

The memory load schematic diagram of disk after innovatory algorithm is introduced when Fig. 5 is β=0.4 in embodiment;

Fig. 6 is the access load schematic diagram of each disk in embodiment;

Fig. 7 is the contrast schematic diagram after the access load before optimization of disk in embodiment.

Embodiment

Below in conjunction with drawings and Examples, the invention will be further described:

The determination of control method for equalizing load of the present invention gets based on following analysis:

First, redundant data distribution rule is analyzed:

As shown in Figure 1, Fig. 1 illustrates physical node, mapping relations between disk and dummy node.In present distributed storage system, usually adopt cheap PC to build bottom memory node, node failure is considered as normality, and in supposing the system, M physical node is expressed as P={P ₁, P ₂,, P _m, P _i(i≤M) represents i-th physical node, N number of disk on each physical node, P _ion disk be expressed as D _i={ D _i1, D _i2..., D _iN, D _ij(j≤N) represents P _ion a jth disk, the storage capacity of each disk is identical, represents with Q.M _ithe N of individual physical node _ithe use size of block disk is U _ij(i≤M, j≤N), each dummy node corresponding L block disk, V _i(l < Y), lays respectively at different physical nodes, and every block dish can corresponding multiple dummy node.A lot of storage system, in order to redundancy backup, being selected each data block to keep three parts, in order to reduce the complexity of algorithm, being got L=3.

According to foregoing description, be made up of the disk be positioned on different physical node as long as meet a dummy node, such as, dummy node V _iby laying respectively at P ₁on D _1N, P ₂on D ₂₃and P _mon D _mnblock disk is formed.Node which disk is made up of actually and decides according to the rule of default and load.

Secondly, physical node disk load analysis:

Consider the size fixation of C that dummy node takies in each disk, so each disk can a corresponding Q/C dummy node.At any one physical node P _iany one piece of disk D _ija corresponding bitmap T _ij, the dummy node that minute book local disk is corresponding, so disk D _ijuse size be:

$U_{\mathrm{ij}}=C\underset{k=0}{\mathrm{\&Sigma;}}{v}_k,v_k\&Element;T_{\mathrm{ij}}$

The use size of single physical node is the use size sum of each disk, uses PU _irepresent:

$P{U}_i=\underset{j=0}{\overset{N}{\mathrm{\&Sigma;}}}{U}_{\mathrm{ij}}=C\underset{j=0}{\overset{N}{\mathrm{\&Sigma;}}}\underset{k=0}{\mathrm{\&Sigma;}}{v}_k$

In order to calculate any one disk D _ijoverall load, total use amount S of whole system must be calculated:

$S=\underset{i=1}{\overset{M}{\mathrm{\&Sigma;}}}{\mathrm{PU}}_i=\underset{i=1}{\overset{M}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{N}{\mathrm{\&Sigma;}}}{\mathrm{PU}}_{\mathrm{ij}}=C\underset{i=1}{\overset{M}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{N}{\mathrm{\&Sigma;}}}\underset{k=1}{\mathrm{\&Sigma;}}{v}_k$

So can calculate the overall load G of any one piece of disk in any one physical node _ij:

$G_{\mathrm{ij}}=\frac{U_{\mathrm{ij}}}{S}=\frac{C\underset{k=0}{\mathrm{\&Sigma;}}{v}_k}{C\underset{i=1}{\overset{M}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{N}{\mathrm{\&Sigma;}}}\underset{k=1}{\mathrm{\&Sigma;}}{v}_k}=\frac{\underset{k=0}{\mathrm{\&Sigma;}}{v}_k}{\underset{i=1}{\overset{M}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{N}{\mathrm{\&Sigma;}}}\underset{k=1}{\mathrm{\&Sigma;}}{v}_k}$

To the average load of any one piece of disk be so

According to above analysis, can determine a kind of standard whether load of the disk of whole storage system is balanced that judges, the mean square deviation of disk is expressed as:

$d\left(s\right)=\sqrt{\frac{1}{\mathrm{MN}}\underset{i=1}{\overset{\mathrm{MN}}{\mathrm{\&Sigma;}}}{(G_{\mathrm{ij}}-\stackrel{\&OverBar;}{\mathrm{AGL}})}^2}=\sqrt{\frac{1}{\mathrm{MN}}\underset{i=1}{\overset{\mathrm{MN}}{\mathrm{\&Sigma;}}}{(G_{\mathrm{ij}}-\frac{S}{\mathrm{MN}})}^2}$

If a kind of mapping scheme can be found, make d (s) minimum, so just can ensure load balancing between node.

Control method for equalizing load for storage system of the present invention, comprises the following steps:

(1) according to the mapping feature of system model physical node and dummy node, employing Multiway Tree Structure is encoded, the corresponding one tree of whole system structure, nodes all on the corresponding root node to leaf node path of gene position, root node and level 0 are defined as system, and ground floor represents physical node, and the second layer represents disk, the corresponding disk piecemeal of third layer, the 4th layer of corresponding dummy node;

(2) stochastic generation contains the initial population of several individualities;

(3) determine whether to meet iteration termination condition, if do not met, enter step (4), if met, Output rusults;

(4) according to the fitness of each individuality of following formulae discovery:

$F\left(s\right)=\frac{1}{M+N.f\left(s\right)}$

Wherein, F (s) fitness function, M, N are constant, and f (s) is determined by following formula:

$f\left(s\right)=d\left(s\right)=\sqrt{\frac{1}{\mathrm{MN}}\underset{i=1}{\overset{\mathrm{MN}}{\mathrm{\&Sigma;}}}{(G_{\mathrm{ij}}-\stackrel{\&OverBar;}{\mathrm{AGL}})}^2}=\sqrt{\frac{1}{\mathrm{MN}}\underset{i=1}{\overset{\mathrm{MN}}{\mathrm{\&Sigma;}}}{(G_{\mathrm{ij}}-\frac{S}{\mathrm{MN}})}^2}$

Wherein, G _ijrepresent the overall load of any one piece of disk in any one physical node, represent the average load of any one piece of disk, S represents total use amount of whole system;

(5) according to next step individuality of roulette selection method choice, described roulette selection method comprises the following steps:

1. select probability is calculated as follows:

$p_i=\frac{F(x,i)}{\underset{j=1}{\overset{N}{\mathrm{\&Sigma;}}}F(x,j)}$

Wherein, p _ithe probability being genetic to colony of future generation for individual i is selected, F (x, i) is the fitness function of individual i;

2. according to select probability, be divided into by disk N number of fan-shaped, each fan-shaped central angle is 2 π p _i, in a disk border random selecting reference point, rotary disk, corresponding fan-shaped of final reference point is exactly the individual x be selected _i;

3. cumulative x _ithe number of times be selected;

2. and 3. N time 4. constantly repeat above-mentioned steps, can individuality be selected;

(6) carry out crossing operation to the individuality selected, described crossing operation comprises the following steps:

A, from population, select two intersect individual, be expressed as S ₁and S ₂;

B, for S _iany one physical node, be denoted as namely a vector of physical node, disk and this three of dummy node is represented;

C, P _iall mappings are arranged in order according to virtual label order, are expressed as T _i;

D, re-establish physical node, relation between disk and dummy node, T _ielement be iteratively corresponding in turn to physical node P _iall disks in.

(7) carry out variation process to the gene position of population, described variation refers to that some genic value in individual UVR exposure string is replaced with other genic value by foundation mutation probability, thus the individuality that formation one is new, described variation process comprises the following steps:

A, according to random selection two physical nodes of mutation probability;

B, when ensureing that dummy node mapping position is different, from these two physical nodes, selecting a disk respectively, these two disks are exchanged.

(8) remove the individuality that in population, fitness is minimum, guarantee the scale of population, produce new population, and return step (3).

In order to verify correctness and the validity of control method for equalizing load of the present invention, we carry out applying and testing for the distributed file system of enterprise, and detailed process is as follows:

First set up a distributed storage system, one of them is as host node as hot standby metadata node to adopt two in this system, and four physical nodes are as data memory node, and each node three pieces of hard disks, each hard disk divides the block of size 64MB in logic.Client passes through the back end position and the file storage location that obtain data with host node alternately, then obtains data with back end direct interaction.Every number, according to there being three parts of copies, lays respectively at different physical node.All nodes run Red Hat Enterprise Linux X64 operating system, every platform machine 4G internal memory, every platform machine three pieces of 20G hard disks, Intel I5-2430 2.4GHz CPU.In experiment, add up the loading condition of all disks, as shown in Figure 2.After the load of certain disk reaches threshold values, system processes according to above-mentioned Load Balancing Model automatically, in order to be make above-mentioned algorithm have good convergence, relevant parameter is arranged as follows, in fitness function F (s), M is 1, probability of crossover and mutation probability α, β, from existing knowledge, when α ∈ [0,1], β ∈ (0,1), and the genetic algorithm of adoption rate selection algorithm can obtain globally optimal solution.Algorithm shows different performances when α, β value is different.When α=0.6, when 0.7, the performance of algorithm iteration number of times is as Fig. 3, and by observation experiment result, when β=0.3, when 0.4, iterations is minimum.When β=0.3, α is in value 0.7, and 0.8 uses this algorithm to carry out load balance optimization to system, Supported Comparison after process is as Fig. 4, and when β=0.4, α is in value 0.7,0.8 uses this algorithm to carry out load balance optimization to system, and the Supported Comparison after process is as Fig. 5.By comparison diagram 4 and Fig. 5, when α=0.8, when β=0.3, whole process iterates least number of times, and to the load optimized the best of system storage.Under above-mentioned value scheme, the access load between each disk also can effectively be improved, and Fig. 6 shows the access load of each disk, and Fig. 7 shows each disk access Supported Comparison adopting the method for the invention to optimize front and back.

As from the foregoing, after storage system uses such scheme, the load of memory node can be at equilibrium by optimization, and the access load of all disks changes thereupon, and the load of the node of the frequent access of part reduces greatly.Thus demonstrate correctness and the validity of algorithm herein.

Above-described embodiment is preferred embodiment of the present invention; it is not the restriction to technical solution of the present invention; as long as without the technical scheme that creative work can realize on the basis of above-described embodiment, all should be considered as falling within the scope of the rights protection of patent of the present invention.

Claims (4)

1. for a control method for equalizing load for storage system, it is characterized in that: comprise the following steps:

(1) according to the mapping feature of system model physical node and dummy node, employing Multiway Tree Structure is encoded, the corresponding one tree of whole system structure, nodes all on the corresponding root node to leaf node path of gene position, root node and level 0 are defined as system, and ground floor represents physical node, and the second layer represents disk, the corresponding disk piecemeal of third layer, the 4th layer of corresponding dummy node;

(2) stochastic generation contains the initial population of several individualities;

(3) determine whether to meet iteration termination condition, if do not met, enter step (4), if met, Output rusults;

(4) according to the fitness of each individuality of following formulae discovery:

$F\left(s\right)=\frac{1}{M+N.f\left(s\right)}$

Wherein, F (s) fitness function, M, N are constant, and f (s) is determined by following formula:

$f\left(s\right)=d\left(s\right)=\sqrt{\frac{1}{\mathrm{MN}}\underset{i=1}{\overset{\mathrm{MN}}{\mathrm{\&Sigma;}}}{(G_{\mathrm{ij}}-\stackrel{\&OverBar;}{\mathrm{AGL}})}^2}=\sqrt{\frac{1}{\mathrm{MN}}\underset{i=1}{\overset{\mathrm{MN}}{\mathrm{\&Sigma;}}}{(G_{\mathrm{ij}}-\frac{S}{\mathrm{MN}})}^2}$

Wherein, G _ijrepresent the overall load of any one piece of disk in any one physical node, represent the average load of any one piece of disk, S represents total use amount of whole system;

(5) according to next step individuality of roulette selection method choice;

(6) crossing operation is carried out to the individuality selected;

(7) variation process is carried out to the gene position of population;

(8) remove the individuality that in population, fitness is minimum, guarantee the scale of population, produce new population, and return step (3).

2. the control method for equalizing load for storage system according to claim 1, is characterized in that: in described step (5), and described roulette selection method comprises the following steps:

1. select probability is calculated as follows:

$p_i=\frac{F(x,i)}{\underset{j=1}{\overset{N}{\mathrm{\&Sigma;}}}F(x,j)}$

Wherein, p _ithe probability being genetic to colony of future generation for individual i is selected, F (x, i) is the fitness function of individual i;

2. according to select probability, be divided into by disk N number of fan-shaped, each fan-shaped central angle is 2 π p _i, in a disk border random selecting reference point, rotary disk, corresponding fan-shaped of final reference point is exactly the individual x be selected _i;

3. cumulative x _ithe number of times be selected;

2. and 3. N time 4. constantly repeat above-mentioned steps, can individuality be selected.

3. the control method for equalizing load for storage system according to claim 1, is characterized in that: in described step (6), described crossing operation comprises the following steps:

A, from population, select two intersect individual, be expressed as S ₁and S ₂;

B, for S _iany one physical node, be denoted as namely a vector of physical node, disk and this three of dummy node is represented;

C, P _iall mappings are arranged in order according to virtual label order, are expressed as T _i;

D, re-establish physical node, relation between disk and dummy node, T _ielement be iteratively corresponding in turn to physical node P _iall disks in.

4. the control method for equalizing load for storage system according to claim 1, it is characterized in that: in described step (7), described variation refers to that some genic value in individual UVR exposure string is replaced with other genic value by foundation mutation probability, thus the individuality that formation one is new, described variation process comprises the following steps:

A, according to random selection two physical nodes of mutation probability;

B, when ensureing that dummy node mapping position is different, from these two physical nodes, selecting a disk respectively, these two disks are exchanged.