CN101923496A - Method for achieving parallel reconstruction of data on RAID - Google Patents

Abstract

The invention discloses a method for achieving the parallel reconstruction of data on an RAID (redundant array of independent disk), particularly achieving the parallel reconstruction thereof in a sub-array distribution manner, wherein the sub-array comprises the blocks of different disks. The method comprises the following steps: respectively and simultaneously searching the other blocks constituting the same sub-array together with each block in a data disk; and reconstructing the data in the block according to the data in the other blocks. The method of the invention can shorten the process of data reconstruction by times, more particularly, all the operations of reconstruction, either reading or writing the related data, are parallel operations, and the requests from onstage users can be responded on a real-time basis in the reconstruction process, thus effectively reducing the average time for responding to the users, improving the reliability of the system and enhancing the performance thereof at the same time.

Description

The parallel reconstruction of data method of a kind of RAID

Technical field

The invention belongs to array data reconfiguration technique field, particularly relate to a kind of method of supporting reconstructing parallel data.

Background technology

Disk failure will make storage system be operated under the degraded mode, and the overall performance decline problem that system works will cause under this pattern therefore in order to guarantee the Performance And Reliability of system, must be reconstructed the data of failed disk.How can guarantee during data reconstruction that accelerating reconstruct speed again in the performance of system is two aspects that data reconstruction need be considered, many data reconstruction technology have been related to thus, data reconstruction is at reconstitution time and user's draw response time two indexs at present, mainly distribute and two aspects of data-flow analysis are done and optimized improvement, improve system performance with this from data or check block.

First kind of mode of data reconstruction mainly is to utilize the distribution of data or check block to improve the performance of reconstruct.(Distributed sparing) method of disperseing spare drive of having researched and proposed is arranged, the space availability ratio that this mode has improved Hot Spare spare dish is accelerated the speed of reconstruct to a certain extent, this method mainly also is applied to idle at ordinary times spare dish in the array in the DATA DISTRIBUTION of array RAID, in each disk in array all the reserve part clear area use as spare dish, the data of same band are through writing the clear area in the same band when data failure occurring after the verification, this method has effectively been avoided the bottleneck problem of writing of traditional spare dish, shortened the data reconstruction time, but owing to need a large amount of read operations in order to finishing data restore operation, still be a bottleneck to the read operation of single-deck.Checking data dispersion ParityDeclustering method is then balanced data reconstruction time and user's average response time two aspects, the key distinction of its checking data distribution mode and traditional RAID distribution (particularly RAID5) is, the checking data dispersion degree is littler than traditional RAID 5 modes, be that data are lacked than RAID 5 modes with the number of disks that its checking data distributes, this characteristics make also that after disk failure system can finish user's request that data reconstruction also can utilize the disk response foreground that does not participate in reconstruct simultaneously.Adopt the data reconstruction performance of this distribution mode few with respect to traditional reconstruct poor performance, just it is better than the former on user's average response time performance index.Utilize data flow characteristics, it is the another kind of mode of present data reconstruction that the storage distribution is optimized.Use and propose and be primarily aimed at Streaming Media class (Continuous-media Servers) based on data reconstruction (Pipelined Reconstruction) method of instruction stream, this application scenario is higher to the reliability requirement of system, if the disk failure situation must guarantee foreground user's request responding is not influenced or influences not quite, therefore the disk reconstruct of this method has made full use of the succession characteristics of media data, the read operation that is reconstructed is the unit with the magnetic track, and deposits unnecessary data in buffer memory in order to next reconstructed operation use.That in addition utilizes data flow characteristics also has based on the restructing algorithm (Popularity-based multithreaded Reconstruction Optimization) of hot spot data etc.Although said method all has more or less improvement or lifting on reconstitution time and average response time, reconstitution time is still very long, and this has also constituted threat to the reliability of system.

Summary of the invention

The present invention proposes the parallel reconstruction of data method of a kind of RAID, has reduced the disk reconstitution time at double, and active balance reconstitution time and user's average response time have improved the reliability of system.

The parallel reconstruction of data method of a kind of RAID, relate to plural data disks and at least one spare drive, each data disks is divided into the block of equal number, utilize each block to make up subarray, make each block of any one data disks belong to different subarrays and only belong to a subarray;

The data reconstruction of individual data dish carries out in the following manner: for each block in this data disks, seek other block respectively simultaneously with the same subarray of its formation, be reconstructed according to the data of other block data, and reconstruction result be saved in spare drive or do not participate in the data disks of this reconstruct this block.

As optimization, for any two subarrays, one of them subarray has only a block of a block and another subarray to belong to a data disks at most.

Described data disks adopts RAID 1～6 data storage rank, and described subarray is constructed in the following manner:

(11) data disks is divided into the c group, the 1st to c group data disks m ₀, m ₁..., m _C-1Block constitute the 1st, 2 respectively ..., c group block distribution matrix, if data disks can not be divided into the c group, then by adding the virtual disk of dividing with the data disks same block or reducing data disks and satisfy and divide equally requirement, c is greater than 2;

(12) numbering of each block in the initialization first block distribution matrix, the j row element of i+1 component cloth matrix be by the j row element ring shift right or the j-1 position that moves to left of i group matrix, i=1 ..., c-1, j=1 ..., n, n are the block counts that data disks is divided;

(13) choose subarray of block formation that each data disks has identical numbering.

Described data disks adopts RAID 10 data storage ranks, and the data disks number is to be not less than 4 even number, and described subarray is constructed in the following manner:

(21) data disks is divided into two groups, the first, two group of data disks m ₀, m ₁Block constitute first and second group block distribution matrix respectively;

(22) capable each block of x of first group of block distribution matrix is carried out the initialization numbering of shape such as x.y form, x=0,1, ..., m/2-1, y=0,1, ..., m/2-1, capable each the block numbering of the x of second group of block distribution matrix is by x capable each the block numbering ring shift right or the x position that moves to left of the first component cloth matrix;

(23) from each data disks, choose block and constitute a subarray with identical numbering.

Technique effect of the present invention is embodied in: by parallel reconfiguration algorithm disk is carried out subregion, select suitable R AID rank that disk is formed a plurality of sub-RAID according to partition table then, when in the array disk failure being arranged, this distribution mode is different from traditional array reconstruct, its reconstruct to disk will can not produce the read-write disk interference, therefore reconstruct speed will significantly improve, simultaneously in restructuring procedure, the disk that does not participate in reconstruct can also respond foreground user request, the present invention can active balance reconstitution time and user's average response time, thereby has improved the performance of system.

Description of drawings

Fig. 1 is the matrix distribution figure of 16 disks;

Fig. 2 is the matrix distribution figure of 10 disks;

Fig. 3 is the matrix distribution of 10 disks figure that optimizes structure;

Fig. 4 is parallel reconstruction RAID 5 data reconstruction structural drawing;

Fig. 5 is the logical mappings figure of parallel reconstruction RAID 5;

Fig. 6 is parallel reconstruction RAID 5 reconstruct process flow diagrams;

Fig. 7 is parallel reconstruction RAID 5 spare drive optimisation strategy figure

Fig. 8 is parallel reconstruction RAID 10 data reconstruction process flow diagrams;

Fig. 9 is test system structure figure;

Figure 10 is a parallel reconstruction reconstitution time test pattern (unit: second)

Figure 11 is a parallel reconstruction user average response time test pattern (unit: millisecond)

Embodiment

1, theoretical foundation

What the present invention mainly realized is the parallel reconstruction of RAID, the parallel reconstruction truly that traditional RAID DATA DISTRIBUTION form is not enough to realize, for example RAID 5 distribution forms of standard after system enters degraded mode with log-on data reconstruct, but because have the data write conflict in this process, therefore the time of whole data reconstruction will be very very long; Particularly for on-line reorganization, it also wants corresponding foreground user request in reconstruct, and this will further increase the time of reconstruct.And proposition of the present invention mainly is at this problem, by adopting parallel reconstruction data block Distribution Strategy and setting up the RAID structure of standard, by with disk zoning piece and with the block being the composition RAID of unit, the present invention of this RAID mode is referred to as subarray (also can be called subRAID), this RAID organizational form can realize the data reconstruction that walks abreast, and do not have read/write conflict in this restructuring procedure, thereby shortening reconstitution time that can be at double;

Table 1 symbol definition

At first introduce the distribution of parallel reconstruction with standard RAID 5 constituted modes, we carry out as giving a definition.

Below will be with regard to D=9, the situation of N=3 distributes to the parallel reconstruction data block and does a description.At first the initial distribution table is divided into 3 groups of sub-block tables, every group of 3 disks are the matrix form of 3 * 3 row with first group echo, use m ₀Represent first group of sub-block table, and to m ₀Carry out the initialization numbering, concrete distribution mode is as follows

${\mathrm{<figure-callout\; id="0"\; label="m"\; filenames="HDA0000023910420000012.png,HDA0000023910420000013.png"\; state="\{\{state\}\}">m</figure-callout>}}_0=\left(\begin{array}{ccc}0& 1& 2\\ 3& 4& 5\\ 6& 7& 8\end{array}\right),$ p _0，1＝(3?4?5)，p _0，2＝(6?7?8)；

M in the following formula ₀What the element numbering in the distribution matrix was represented is the block numbering of subarray, structure distribution matrix m ₁, m ₂Adopt

Function can finally determine the j of i+1 distribution matrix capable be p _{I, j}Distribution mode, promptly obtain i+1 the capable distribution form of distribution matrix j by the circulative shift operation that worked of the j to i distribution matrix, thus and the like can construct complete distribution table.The cyclic shift function here

Will be according to p _{I, j}Difference displacement figure place different, and direction of displacement must unify, and can be ring shift left or ring shift right, the m in this example ₁, m ₂Adopt the ring shift right operation to generate.

${\mathrm{<figure-callout\; id="1"\; label="m"\; filenames="HDA0000023910420000012.png,HDA0000023910420000013.png"\; state="\{\{state\}\}">m</figure-callout>}}_1=\left(\begin{array}{c}{\mathrm{SH}}_r^0\left({\mathrm{<figure-callout\; id="0"\; label="p"\; filenames="HDA0000023910420000012.png,HDA0000023910420000013.png"\; state="\{\{state\}\}">p</figure-callout>}}_{\mathrm{0,0}}\right)\\ {\mathrm{SH}}_r^1\left({\mathrm{<figure-callout\; id="0"\; label="p"\; filenames="HDA0000023910420000012.png,HDA0000023910420000013.png"\; state="\{\{state\}\}">p</figure-callout>}}_{\mathrm{0,1}}\right)\\ {\mathrm{SH}}_r^2\left({\mathrm{<figure-callout\; id="0"\; label="p"\; filenames="HDA0000023910420000012.png,HDA0000023910420000013.png"\; state="\{\{state\}\}">p</figure-callout>}}_{\mathrm{0,2}}\right)\end{array}\right)=\left(\begin{array}{ccc}0& 1& 2\\ 5& 3& 4\\ 7& 8& 6\end{array}\right)$

${\mathrm{<figure-callout\; id="2"\; label="m"\; filenames="HDA0000023910420000012.png,HDA0000023910420000013.png"\; state="\{\{state\}\}">m</figure-callout>}}_2=\left(\begin{array}{c}{\mathrm{SH}}_r^0\left({\mathrm{<figure-callout\; id="1"\; label="p"\; filenames="HDA0000023910420000012.png,HDA0000023910420000013.png"\; state="\{\{state\}\}">p</figure-callout>}}_{\mathrm{1,0}}\right)\\ {\mathrm{SH}}_r^1\left({\mathrm{<figure-callout\; id="1"\; label="p"\; filenames="HDA0000023910420000012.png,HDA0000023910420000013.png"\; state="\{\{state\}\}">p</figure-callout>}}_{\mathrm{1,1}}\right)\\ {\mathrm{SH}}_r^2\left({\mathrm{<figure-callout\; id="1"\; label="p"\; filenames="HDA0000023910420000012.png,HDA0000023910420000013.png"\; state="\{\{state\}\}">p</figure-callout>}}_{\mathrm{1,2}}\right)\end{array}\right)=\left(\begin{array}{ccc}0& 1& 2\\ 4& 5& 3\\ 8& 6& 7\end{array}\right)=\left(\begin{array}{c}{\mathrm{SH}}_r^0\left({\mathrm{<figure-callout\; id="0"\; label="p"\; filenames="HDA0000023910420000012.png,HDA0000023910420000013.png"\; state="\{\{state\}\}">p</figure-callout>}}_{\mathrm{0,0}}\right)\\ {\mathrm{SH}}_r^2\left({\mathrm{<figure-callout\; id="0"\; label="p"\; filenames="HDA0000023910420000012.png,HDA0000023910420000013.png"\; state="\{\{state\}\}">p</figure-callout>}}_{\mathrm{0,1}}\right)\\ {\mathrm{SH}}_r^4\left({\mathrm{<figure-callout\; id="0"\; label="p"\; filenames="HDA0000023910420000012.png,HDA0000023910420000013.png"\; state="\{\{state\}\}">p</figure-callout>}}_{\mathrm{0,2}}\right)\end{array}\right)$

Generally speaking if known initial distribution matrix m ₀Distribution mode, the distribution mode that then is numbered the matrix of i is

$m_i=\left[\begin{array}{c}{\mathrm{SH}}_r^0\left(p_{i-\mathrm{1,1}}\right)\\ {\mathrm{SH}}_r^1\left(p_{i-\mathrm{1,2}}\right)\\ \&CenterDot;\&CenterDot;\&CenterDot;\\ {\mathrm{SH}}_r^{N-1}\left(p_{i-1,N}\right)\end{array}\right]$ Or $m_i=\left[\begin{array}{c}{\mathrm{SH}}_l^0\left(p_{i-\mathrm{1,1}}\right)\\ {\mathrm{SH}}_l^1\left(p_{i-\mathrm{1,2}}\right)\\ \&CenterDot;\&CenterDot;\&CenterDot;\\ {\mathrm{SH}}_l^{N-1}\left(p_{i-1,N}\right)\end{array}\right],$ I ∈ [1, N-1] formula 1

In the formula 1 Function parameter r and l are expressed as right circulation and left side circulation respectively, are M=(m thereby obtain final matrix distribution table ₀, m ₁..., m _N-1).

Then the parallel reconstruction distribution mode of 9 disks is,

$N=3,M=\left[\begin{array}{ccccccccc}0& 1& 2& 0& 1& 2& 0& 1& 2\\ 3& 4& 5& 5& 3& 4& 4& 5& 3\\ 6& 7& 8& 7& 8& 6& 8& 6& 7\end{array}\right]$ Formula 2

Formation for subarray then is to number the identical numbering group subarray that constitutes among the matrix distribution table M that promptly chooses final generation by the same block of choosing among each distribution matrix m, the matrix distribution table M that generates in the formula 2 can constitute 9 subarrays, can adopt similar subRAID 0 to these 9 subarrays, subRAID 1, ..., subRAID 8 represents.

2, theoretical proof

The present invention can prove theoretically that data disks number (also the can be understood as block counts) N of the subarray of supposition formation in theory is a prime number, then can construct the array distribution form with parallel reconstruction characteristic by translation mode.The array that for example adopts 9 disks to constitute, per 3 disks constitute one group, and the set mode of its parallel distributed can be expressed as: regard each disk as row, the number table of each row is shown the block number of this disk, such as

Matrix m ₀Expression array initial distribution is wherein established set u _0,0, u _0,1, u _0,2Represent m respectively ₀The initial distribution matrix three row, then

Set be expressed as U ₀=(u _0,0, u _0,1, u _0,2), m in like manner ₁, m ₂Set be expressed as U ₁=(u _1,0, u _1,1, u _1,2), U ₂=(u _2,0, u _2,1, u _2,2).m ₁, m ₂The matrix distribution mode as noted above; If will realize that parallel reconstruction then must satisfy condition: u between group _{I, j}∩ u _{J, i}={ in e} and the group

I, j, k ∈ [0, N-1] promptly shows between the group and has only an identical element, and column element does not occur simultaneously in the group.

To prove that below for convenience the matrix subscript of proof this moment is since 1 with regard to above-mentioned set expression way, rather than above-mentioned 0.If symbol m _{I, p}The p row of representing i submatrix, m _{J, q}If the q row of representing j submatrix are known m _{I, p}P tabulation be shown

m _{J, q}Q tabulation be shown

a _{I, j}And b _{I, j}Represent distribution matrix m respectively _iThe capable j of i row because the colleague finishes by the circulation translation mutually, so we can use m _{I, p}Represent m _{J, q}

b _{1, q}=a _{1, q}, b _{2, q}=a _{2, (q+j-i) %N}, b _{R, q}=a _{R, [q+2 * (j-i)] %N}..., b _{R, [q+ (R-1) * (j-i)] %N}, promptly have following formula to set up.

$m_{j,q}=\left[\begin{array}{c}{b}_{1,q}\\ b_{2,q}\\ b_{3,q}\\ \&CenterDot;\&CenterDot;\&CenterDot;\\ b_{R,q}\end{array}\right]=\left[\begin{array}{c}{a}_{1,q}\\ a_{2,(q+j-i)\%N}\\ a_{3,[(q+2)\&times;(j-i)\%N]}\\ \&CenterDot;\&CenterDot;\&CenterDot;\\ a_{R,[q+(R-1)\&times;(j-i)]\%N}\end{array}\right],$ R ∈ [1, N] and R ∈ I

We know arbitrarily

M as can be known then _{I, p}And m _{J, q}With identical element is arranged in the delegation.

$\left[\begin{array}{c}{a}_{1,\mathrm{<figure-callout\; id="2"\; label="p\; a"\; filenames="HDA0000023910420000012.png,HDA0000023910420000013.png"\; state="\{\{state\}\}">p</figure-callout>}}& \\ a_{}{}_{2,p}\\ a_{3,p}\\ \&CenterDot;\&CenterDot;\&CenterDot;\\ a_{R,p}\end{array}\right]\&DoubleLeftRightArrow;\left[\begin{array}{c}{a}_{1,q}\\ a_{2,(q+j-i)\%N}\\ a_{3,[(q+2)\&times;(j-i)\%N]}\\ \&CenterDot;\&CenterDot;\&CenterDot;\\ a_{R,[q+(R-1)\&times;(j-i)]\%N}\end{array}\right]$

Suppose to exist two row elements identical, be made as m and n respectively, and m ≠ n, then have ready conditions

Set up, i.e. [(m-n) * (j-i)] %N=0, because

Then can draw N is prime number scarcely, and this is false for prime number contradiction event hypothesis with hypothesis N, so and m _{I, p}And m _{J, q}Common factor have only an identical element, must demonstrate,prove, so this distribution mode can realize conflict free parallel reconstruction.

3, other situation explanations

By top proof as can be known,, constitute the required block counts N of subarray and be necessary for prime number, and can to draw the data disks number that needs by the process that constitutes the parallel reconstruction distribution matrix be N if will realize optimum parallel reconstruction performance ²This also will increase the cost of system virtually, therefore the present invention also proposed still to carry out under a kind of situation of counting deficiency at data disks high performance parallel reconstruct the balance scheme, if will constitute the parallel reconstruction array, can adopt two kinds of strategies to realize, promptly reduce to constitute the block or the disk division block counts dual mode of subarray.

Illustrate the block strategy that reduces to constitute subarray below, if this moment, the data disks in the system was 16, because desired data dish several 16 is between 3 ²With 5 ²Between, in this case, the distribution mode of these 16 data disks can adopt the distribution mode of 25 data disks, promptly adds 9 virtual disks and helps to make up distribution matrixs, last only need be with the distributions of preceding 16 row of the matrix distribution table of 25 disks as 16 data disks.See specifically shown in the accompanying drawing 1 that 25 disks are divided into 5 groups, by to m ₀Initialization, and generate m by circulative shift operation ₁, m ₂, m ₃, m ₄The matrix distribution table.Owing to have only 16 disks at present, then adopt the distribution form of preceding 16 column distribution forms this moment as existing 16 disks.One has 25 subRAID among the figure, but constitute the prime number that is necessary for that the block counts of subRAID does not then have above-mentioned requirements this moment, for example subRAID 0 is numbered 0 block by 4 and constitutes among the figure, and subRAID1 is numbered 1 block by 3 and constitutes, and adopting this mode is to adopt to reduce subarray block counts mode and realize parallel reconstruction in essence.

Usually under the situation if the data disks number is m in this moment system, if m ∈ (a ², b ²], a, b are prime number and a greater than 2 ²And b ²The parallel reconstruction of individual disk is respectively M _aAnd M _b, constituted mode is obtained by formula 1, and then the distribution form of m disk adopts b at this moment ²The distribution M of individual disk _bPreceding m row constitute.But this distribution mode neither be fit to the situation of each m, such as the situation of m=10, satisfies 10 ∈ (3 this moment ², 5 ²], if still get 5 ²Preceding 10 row that individual disk distributes then can not constitute subarray, and as shown in Figure 2, because each subarray all has the minimum block counts of requirement, just as traditional RAID 5 needed 3 disks at least, in order to solve these special circumstances of m=10, the present invention had provided 4 ²The parallel reconstruction distribution form of individual disk, generating mode still adopt formula 1 to generate, but in order to realize that conflict free optimum parallel reconstruction taked to reduce disk here and divided the method that block counts and subarray block counts combine, are about to 4 ²Distribution mode the third line numbering remove, this moment, each disk was divided into three blocks, preceding 10 row of getting this distribution this moment constitute the distribution form of m=10, as shown in Figure 3, subRAID 0, subRAID 1, subRAID 6, subRAID 7, subRAID 14 and subRAID 15 constitute by three blocks, other subRAID then can not constitute standard RAID structure, therefore these block numberings do not constitute sub-permutation, 10 disks have only constituted 6 subarrays, space availability ratio has only accounted for half, this constituted mode is not recommended, and can adopt 32 distribution to solve this problem, but like this only in view of the situation of m=10.

The data organization mode of subarray of the present invention is except RAID 2～6 structures commonly used, the present invention has also provided the parallel reconstruction distribution mode of RAID 10, this mode requires data disks to count m and is not less than 4 even number, it requires harsh unlike other RAID ranks commonly used such as 5 pairs of structures of RAID parallel reconstruction disk array, parallel reconstruction distribution for RAID 10 does not then have such requirement, and its parallel multiple is only with to form this distribution array disk number relevant, for its half.To elaborate with regard to the block table construction process of RAID 10 distribution modes below.

Constitute submatrix with 8 disks and classify example as, each subarray adopts RAID 10 ranks, then in order to realize maximum parallel multiple, half that the disk block number need be divided into total disk number promptly 4, simultaneously all masters are divided into one group in logic, mirror drive is divided into another group, and block is carried out label.

The block distribution matrix of master group represents that mode is

Wherein the block that disk is divided is shown in each tabulation of this matrix, promptly be divided into 0.0,1.0,2.0 and 3.0 4 blocks, the x in shape such as the x.y form represents the numbering of subarray, y represents to constitute the block numbering of x subarray, 0.0 expression constitutes first block of subarray 0, can be understood as first disk that constitutes the traditional magnetic disk array, 0.1,0.2 0.3 is similar.The block distribution matrix of mirror set is represented number to carry out obtaining by formula 1 mode equally according to the block that constitutes subarray in the master group.

Be referred to as the original block table among the dish component cloth matrix the present invention here, if to original block table matrix m in the example ₁The mirror set distribution mode that adopts the ring shift right mode to constitute is

Here shape such as x.y ' are expressed as the mirror image of block numbering x.y, and then parallel reconstruction RAID 10 distributed architectures of 8 disk formations are

This table can be realized the reconstructing parallel data of 4 times of speed, these RAID 10 structures are made of four subarrays 0～4, the band span is 4, and there is not conflict in the reconstruct visit to disk, here the blockette number of disk also can be divided into less than 4, the block table of this moment only need be selected wherein corresponding line to reconstitute new block distribution table on the block table basis of 4 times of speed to get final product, but parallel reconstruction speed at this moment will reduce.Usually, be for the RAID 10 block distribution table forms of the composition of the parallel reconstruction that the individual disk of N (N is not less than 4 even number) is arranged,

Initial table

Mirror list

P ₀First row element of expression initial table, P ₀... P _(N/2)-1It is similar,

Represent a translation function, be about to (i+1) row element ring shift right n position in the initial table, wherein r also can replace with l and represents ring shift left.

4, implementation procedure

(1) at first suitable parallel reconstruction mode is selected according to the available data dish by system, can adopt RAID 10 or RAID 2～6 modes, and this implementation procedure is implemented with the RAID 5 of 9 disks.As shown in Figure 4, the block numbering of 9 disks among the figure distributes shown in formula 2, and each disk is divided into three blocks, and per three blocks constitute a subarray, one has 9 subarrays, and the form of shape such as x.y is expressed as the block that is numbered y of forming subarray x.SubRAID 6 is respectively by disk D0,6.0,6.1 and 6.2 blocks among D5 and the D7 constitute, in like manner subRAID 7 and subRAID 8 have three blocks to constitute equally, the subarray here (subRAID) adopts the data organization mode of standard RAID 5, and subarray of the present invention also can adopt other data organization modes.

(2) logical mappings table such as Fig. 5 of all these RAID, the mapping table information that 9 RAID are arranged among the figure, here we suppose that the mapping table size of each RAID is K data unit, in order to reach the purpose of DATA DISTRIBUTION equilibrium, we have carried out the distribution shown in the figure to each data cell, here storing the corresponding corresponding disk address information of each RAID in each data cell, this mapping mode can guarantee the load balancing of each disk.

(3) after parallel reconstruction RAID configuration is finished, there is difference in its response mode under normal mode with respect to traditional RAID, can know by the previous step mapping process, with respect to traditional RAID, the band span of parallel reconstruction mode is narrower than traditional RAID, and this structure improves for the read-write operation performance, particularly when RAID is in degradation or reconfiguration mode, concrete respective process is the same with traditional RAID, and detailed process is not done introduction.

(4) when disk failure occurring among the RAID, system will enter degraded mode, to carry out the data reconstruction operation subsequently, system will read corresponding disk according to the logical mappings relation and be reconstructed, because the parallel reconstruction that adopts the present invention to propose distributes, make reconstruction progress not have read/write conflict to the data reconstruction operation of failed disk.RAID 5 structures that this process will be formed with 9 disks are that example describes, D0～D9 is the numbering of these 9 disks, SD0～SD2 is the required backup diskette of system, as shown in Figure 6, the numbering of similar a.b among the figure, a represents the label of subRAID, because 9 RAID are arranged, here we are referred to as subRAID a, and b represents (b+1) individual Logical Disk of sub-RAID.Disk D4 lost efficacy among the figure, and the inefficacy that this will cause subRAID 1, subRAID 3, subRAID 6 relates separately to block 1.1,3.1 and 8.1.Reconstruct to block 1.1, system is with 1.2 block data information on 1.0 blocks on the reading disk D1 and the disk D7, to the result be write back to standby dish SD1 by xor operation then, till this process lasts till that always the data of whole block 1.1 are all finished by reconstruct, meanwhile, two other reconstruction progress is the reconstruct of finishing respectively block 3.1 and 8.1, and reconstruct is write back to standby dish SD0 and SD2 respectively.

(5) after restructuring procedure is finished, the information of the disk D4 of three disk storage of SD0～SD2, system will lead the information of three standby dishes in idle period of time and get back in the dish of replacing D4, revise relevant map information simultaneously.

(6) the too much problem of standby dish that needs in view of this parallel reconstruction structure, the present invention has also proposed solution, notice that disk D3 and D5 among Fig. 6 do not participate in reconstructed operation in restructuring procedure, therefore can be at the parallel reconstruction disk tell the information that a block is used to store reconstruct again, structure as shown in Figure 7, can be written to the reconstruct data to block 1.1 among the disk D3 this moment, the data reconstruction of block 3.1 is written among the D5, block 8.1 is written among the standby dish SD0, this distribution mode can realize parallel reconstruction equally, has also reduced the use of two standby dishes simultaneously.

(7) the sub-RAID of this parallel reconstruction mode can use common RAID 1, RAID5, RAID6 etc., for RAID 10 this array modes, Fig. 8 has provided the DATA DISTRIBUTION mode of RAID10 parallel reconstruction, distribution mode adopts shifting function to finish equally, what implement among the figure is reconstructed operation to disk D2, and concrete restructuring procedure and RAID 5 are similar.

(8) in order to assess the performance of this parallel reconstruction of the present invention, main is reconstitution time and user's average response time according to two indexs, Fig. 9 is the structural drawing of test, adopt the client-server mode to test, connect by ISCSI, the present invention is mainly used in the parallel RAID reconstructed module among the figure, test case has adopted three standard testing use-case (Fin1, Fin2, Web), contrast test is MD (Multiple Device) (not needing Chinese) to liking linux system soft RAID commonly used down.

(9) Figure 10 represents the reconstitution time test, and as can be seen from the figure parallel reconstruction RAID5 is with respect to the RAID 5 of standard MD formation, no matter its which test case, the reconstitution time of parallel reconstruction RAID 5 has several times raising; Aspect user's average response time, the result who provides from Figure 11 as can be seen, under the situation of same load, user's average response time of parallel reconstruction RAID 5 also has reduction in various degree.

The present invention adopts circulation translation method to make up just wherein a kind of optimal way of subarray, and the technician in field, place can adopt alternate manner to make up, and makes up and requires as long as satisfy subarray.

Claims (4)

1. the parallel reconstruction of data method of a RAID, relate to plural data disks and at least one spare drive, each data disks is divided into the block of equal number, utilize each block to make up subarray, make each block of any one data disks belong to different subarrays and only belong to a subarray;

The data reconstruction of individual data dish carries out in the following manner: for each block in this data disks, seek other block respectively simultaneously with the same subarray of its formation, be reconstructed according to the data of other block data, and reconstruction result be saved in spare drive or do not participate in the data disks of this reconstruct this block.

2. parallel reconstruction of data method according to claim 1 is characterized in that, for any two subarrays, one of them subarray has only a block of a block and another subarray to belong to a data disks at most.

3. parallel reconstruction of data method according to claim 1 and 2 is characterized in that, described data disks adopts RAID 1～6 data storage rank, and described subarray is constructed in the following manner:

(11) data disks is divided into the c group, the 1st to c group data disks m ₀, m ₁..., m _C-1Block constitute the 1st, 2 respectively ..., c group block distribution matrix, if data disks can not be divided into the c group, then by adding the virtual disk of dividing with the data disks same block or reducing data disks and satisfy and divide equally requirement, c is greater than 2;

(12) numbering of each block in the initialization first block distribution matrix, the j row element of i+1 component cloth matrix be by the j row element ring shift right or the j-1 position that moves to left of i group matrix, i=1 ..., c-1, j=1 ..., n, n are the block counts that data disks is divided;

(13) choose subarray of block formation that each data disks has identical numbering.

4. parallel reconstruction of data method according to claim 1 and 2 is characterized in that, described data disks adopts RAID 10 data storage ranks, and the data disks number is to be not less than 4 even number, and described subarray is constructed in the following manner:

(21) data disks is divided into two groups, the first, two group of data disks m ₀, m ₁Block constitute first and second group block distribution matrix respectively;

(22) capable each block of x of first group of block distribution matrix is carried out the initialization numbering of shape such as x.y form, x=0,1, ..., m/2-1, y=0,1, ..., m/2-1, capable each the block numbering of the x of second group of block distribution matrix is by x capable each the block numbering ring shift right or the x position that moves to left of the first component cloth matrix;

(23) from each data disks, choose block and constitute a subarray with identical numbering.

Images