CN1955912A - Method for distributing resource in large scale storage system - Google Patents

Abstract

A method for distributing resource in mass storage system includes evaluating resource in SAN according to volume rule, carrying out resource distribution as per evaluation value based on probability random manner for realizing resource utilization and performance optimization and simultaneously supporting addition and deletion of magnetic disc in SAN.

Description

Resource allocation methods in the mass storage system (MSS)

Technical field

Resource allocation methods in the mass storage system (MSS) belongs to field of storage area networks, relates in particular to wherein DATA DISTRIBUTION and load balance field.

Background technology

Along with fast development of information technology, people's data volume to be processed constantly increases; This has proposed great demand to storage system capacity and performance, and makes traditional storage mode can't satisfy user's needs.(Storage AreaNetwork SAN) is a kind of special storage networking to storage area network, and it can link together various memory devices, form the storage resources of unified management, and main frame can visit these resources by optical fiber or IP network.Simultaneously, for the disk of supporting different size in the storage networking and the front end main frame of various isomeries, Intel Virtualization Technology is introduced in the storage area network.The physical details of virtual shielding memory device, with they virtual be that a huge storage pool offers the front end main frame.In this storage architecture, virtualized storage resources is shared by the front end main frame by network, and the physical resource configuration can constantly change, to satisfy the demand of user to memory capacity or performance.

But, simple virtually can not solve all problems along with the continuous increase of storage networking scale.At first, in extensive storage area network, (Quality of Service QoS) requires to be not quite similar to the service quality of storage system capacity, bandwidth, time-delay for the bandwidth of memory device, capacity and front end main frame; How to be that the host assignment resource is a major issue with optimization that reaches performance of storage system and the requirement of satisfying the front end main frame of trying one's best in the memory device of differentiation.Secondly, when the access module of storage system capacity or front end main frame changes,, then can't reach the best performanceization of storage system as not carrying out redistributing of storage resources.Simultaneously, though present SCSI (Small Computer System Interface, small computer system interface) Mean Time Between Failures of disk can arrive hundreds thousand of hours, but in the SAN environment, there are thousands of disks, on average less than having a disk failures in 24 hours; How to replace the disk that goes wrong and not obliterated data become an important problem.

Summary of the invention

The objective of the invention is to design a kind of evaluation method based on capacity to the isomery disk, and realize the stochastic distribution of storage resources according to this evaluation result, promptly random data distributes.By this method, storage resources distributes from the striping of traditional rule and changes stochastic distribution according to capacity into, makes the resource utilization of each memory device reach balance substantially and load is balanced more; Simultaneously, when the physical layout of storage system changed, this method can guarantee to move minimum data to reach the new balance of system.

The present invention is structured on the Storage Virtualization.Virtualization system can be organized into a unified storage pool (Source Container is called for short SC) to all physical disks (Physical Disks is called for short PD).The storage administrator can optionally divide storage space according to user's needs from SC, thereby forms numerous virtual disk (Virtual Disks is called for short VD), uses for front-end server.The SCSI instruction that virtualization system receiving front-end server is sent is converted into actual equipment physical address to logical address and carries out read-write operation and return results.Like this, virtualization system can be realized the integration of storage resources.In realization, system is with the unit of physical disk according to a fixed size---and SG (Storage Granularity) is divided into several pieces, and according to the sequence of addresses number consecutively; Simultaneously, VD also is divided into several SG by this size, and according to the sequence of addresses number consecutively.We get SG in realization be 512KB.Be respectively PD and VD establishing resource use table and resource mapping table.Resource use table record should table correspondence each SG of PD whether used and specifically be mapped on which SG of which VD by certain VD.Establishing resource mapping table record is the particular location of a certain PD that is mapped to of the SG of corresponding VD with it.So the purpose of the resources allocation among the present invention is filled resource use table and resource mapping table exactly; In PD, find the SG of correspondence for each SG of VD.The SG of a VD can be mapped among a plurality of PD.

The disk evaluation is the basis of resource distribution, and it produces evaluation of estimate according to the capacity (Capacity is called for short C) of physical disk, and the data volume that is distributed on this physical disk is determined by its capacity.

Distributing based on virtualized storage resources is exactly to distribute SG among the corresponding PD to fill resource use table and resource mapping table for the SG among the VD.Generally speaking, will carry out resources allocation under following three kinds of situations: create new VD, add new PD and delete PD in SC from SC, its rudimentary algorithm is as follows.

N piece physical disk PD is arranged in the supposing the system ₁, PD ₂PD _n, evaluation of estimate (being capacity) is respectively C ₁, C ₂C _nSet of data objects S={SG to be distributed ₁, SG ₂... SG _m, m is the size of S set.Definition PD ₁Capacity region between be (0, C ₁), PD ₂Capacity region between be (C ₁, C ₁+ C ₂) ..., PD _nCapacity region between be

When creating VD or adding PD, S is a data object all among the SAN; When deletion PD, S is the data object of waiting to delete on the PD.For each the object SG that treats among the distributed data object set S _i, produce a random number r, when r drops on PD _jCapacity region between the time, with SG _iBe distributed in PD _jIn.

Particularly, when adding in the storage system or removing physical disk, must redistribute the storage resources of each VD, the part storage be moved in the new physical disk, with the proof load balance.In order to guarantee this load balance, the present invention has adopted based on the random device of probability and has carried out resources allocation.

The invention is characterized in, contain following step successively:

Step (1): on the virtual management node of described system, load a resources allocation and resource distribution module and a middle layer command processing module, wherein:

Resources allocation and configuration module contain following data structure:

Ax_pd_t: write down the attribute of some physical disk PD, comprise capacity, bandwidth, use space, free space and resource use table; This resource uses table corresponding to PD not, and whether each the storage administration elementary cell SG that has write down corresponding PD is used and be mapped on which SG of which VD by certain virtual disk VD; Described PD and VD all are divided into equal-sized several SG;

Ax_vd_t: write down the attribute of the VD of some virtual disks, comprise capacity and resource mapping table; This resource mapping table is corresponding to some VD, write down the particular location of a certain PD that each SG of corresponding VD is mapped to;

Axum_SC: thus total size of physical disk PD, virtual disk VD, total free space in the record total system, and comprise the chained list that a chained list of being made up of all ax_pd_t in the system and all ax_vd_t form;

Described resources allocation and configuration module are realized resources allocation in the following manner:

Set: n piece physical disk is arranged in the system, use PD ₁, PD ₂PD _nExpression, the evaluation of estimate of representing with capacity is respectively C ₁, C ₂C _n

Set of data objects S={SG to be distributed ₁, SG ₂... SG _m, m is the size of S set;

Definition PD ₁Capacity region between be (0, C ₁), PD ₂Capacity region between be (C ₁, C ₁+ C ₂) ..., PD _nCapacity region between be

Then: when creating VD or adding PD, S is a data object all among the SAN; When deletion PD, S is the data object of waiting to delete on the PD;

For each the object SG that treats among the distributed data object set S _i, produce a random number r, when r drops on PD _jCapacity region between the time, with SG _iBe distributed in PD _jIn;

Create PD _nFlow process as follows:

Step (1.0) is PD distribute data structure ax_pd_t, fills the base attribute that comprises capacity, available size, the resource allocation table of initialization PD; Defining variable P also is initialized as 0;

Step (1.1) joins PD in the virtual management node all PD is organized into a unified storage pool SC virtually;

Step (1.2), the VD that exists among the SC is carried out resource re-allocation according to the following steps:

Step (1.2.1) is to all physical disk PD that has distributed _i, i=1,2 ..., n-1, definition δ _i=C _i/ S '-C _i/ S, wherein $S=\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{C}_i$ For comprising PD _nIn the total volume of all interior physical disks, $S^{\′}=\underset{i=1}{\overset{n-1}{\mathrm{\Σ}}}{C}_i,$ Be all total volumies of the physical disk of Resources allocation;

Step (1.2.2) is for all SG _iProduce the random number r that (0,1) interval is evenly distributed, described SG _i∈ PD _iAnd all SG _iDistributed to VD, i=1,2 ..., n-1;

Step (1.2.3) is for this SG _iAffiliated PD _i, if r＜δ _i, this SG then _iMove to PD _nP SG in, and revise corresponding resource allocation table and resource mapping table, P adds 1; Otherwise this SG _iRemain unchanged;

Step (1.2.4), repeating step (1.2.3) finishes up to i=n-1;

Deletion PD _nFlow process as follows:

Step (2.0) is for this PD _nGo up the set of all SG that distributed, produce

The n-1 that the interval is evenly distributed random number r;

Step (2.1) is if random number r _iBe positioned at PD _iCapacity region between on, then this SG is moved to PD _iP _iOn the individual SG; Otherwise this SG remains unchanged; P _iBe PD _iUpward first is not assigned to the position of the SG of VD by sequence of addresses, i=1, and 2 ..., n-1;

Step (2.2), search PD _iMake P _iPoint to PD _iThe next SG that is not assigned with, and revise corresponding resource allocation table and resource mapping table;

Step (2.3) is PD _nFrom SC, delete;

The flow process of creating VD is as follows:

Step (3.0) is calculated remaining space sum among the SC whether greater than the capacity of the VD that will create.If not, then can not create this VD; Otherwise change next step;

Step (3.1), the data structure of establishment VD is also added among the SC;

Step (3.2) is as if the VD that is that will create _nIts step is as follows:

Step (3.2.1) is to this VD _nGo up the set of all SG, produce The n that the interval is evenly distributed random number r;

Step (3.2.2) is if r _iBe positioned at PD _iCapacity region between on, then this SG is assigned to PD _iP _iOn the individual SG; P _iBe PD _iUpward first is not assigned to the position of the SG of VD by sequence of addresses, i=1, and 2 ..., n;

Step (3.2.3), search PD _iMake P _iPoint to PD _iThe next SG that is not assigned with, and revise corresponding resource allocation table and resource mapping table;

Middle layer command processing module, the resource mapping table that provides according to ax_vd_t are carried out address mapping and are finished the order of the front end main frame of being accepted;

Step (2), described method are carried out resources allocation according to the following steps successively:

Adding PD _nThe time:

At first, be this PD _nDistribute and initialization ax_pd_t data structure and resource use table and it is joined among the global variable axup_SC;

Secondly, carry out resource re-allocation, heavily purchase resource use table and resource mapping table;

At last, wake background thread up, this thread carries out copying data according to resource use table after heavily purchasing and resource mapping table;

At deletion PD _nThe time:

The first step, deletion PD _nWhether in time, need shift the data of VD on it, have enough resources to use for VD behind the calculating deletion PD.If not, then this PD can not delete;

Second step, carry out resource re-allocation, the reconstruct resource is used table and resource mapping table;

The 3rd step, wake background thread up, finish the copying data after heavily purchasing;

In the 4th step, the ax_pd_t structure of this PD of deletion discharges its resource use table from axup_SC;

When deletion VD:

At first, whether there are enough spaces to create this VD among the calculating SC;

Secondly, for VD creates and initialization ax_vd_t data structure and resource allocation table, it is added among the axup_SC;

At last, be the VD Resources allocation, and the change resource is used table and resource mapping table.

Show that by experiment this method has reduced 14.8% with the mean access time of disk under the prerequisite that realizes the equilibrium of disk utilization factor, have good performance.

In order to solve the resource allocation problem in the magnanimity isomery storage networking environment, we have proposed the disk evaluation and the data stochastic distribution method based on capacity of Storage Virtualization in the SAN environment.It is estimated the disk in the SAN environment according to capacity, and carries out resources allocation with this, thereby reaches resource utilization and optimization in Properties; Simultaneously, support the interpolation and the deletion of disk among the SAN.This method has following characteristics:

The compatibility and the dirigibility of height: by the disk evaluation, this method is supported various isomery disks, and this has improved compatibility and the dirigibility of SAN greatly.

Resource management efficiently: the distribution of storage resources can be on each memory node executed in parallel, have very high performance.

The cost of resource management is little: owing to adopt at random DATA DISTRIBUTION, each resources allocation and the data volume that moves when heavily distributing reach optimum.

Highly scalable and reliability: heavily distribute by resource, realized the interpolation and the deletion of memory device in the SAN environment, improved extensibility and the reliability of SAN greatly.

Description of drawings

Fig. 1. the storage resources hierarchy chart.

Fig. 2. virtual mass memory mapping resources figure.

Fig. 3. system construction drawing.

Fig. 4. virtual mass storage module structural drawing.

Fig. 5. create the PD process flow diagram.

Fig. 6. the process flow diagram of deletion PD.

Fig. 7. create the process flow diagram of VD.

Embodiment

Core of the present invention is to safeguard a cover storage resource configuration information with various memory device resource in the centralized management storage networking by the virtualization process node of special use, and carries out resources allocation.This virtualization process node is the server of an operation (SuSE) Linux OS.Simultaneously, it carries out the order analysis processing by the SCSI middle layer provides transparent stores service fully to the front end main frame.Concrete realization framework comprises following two major parts:

1. safeguard the configuration information tabulation of whole memory devices (comprising physical equipment and virtual unit) on the virtualization process node: to each PD, write down its various attributes and space list, i.e. the SG table; Create VD, carry out resources allocation when adding deletion PD, and each VD is created a structure, write down various attributes and spatial mappings mode.

2. realize the SCSI software object device STML (SCSI TargetMid-level Layer) of a virtual support function on the virtualization process node, its work is that the front end main frame is converted into visit (being the physics order) to actual physical device to the visit (being dummy command) of virtual disk logical space, and the data processed result of physical equipment is fed back to the front end main frame.

Hardware device required for the present invention comprises front end main frame, virtual management node, I/O node and has the disk of sonet card interface; Network comprises fiber optic network (being the SAN network) and Ethernet.Wherein, the SAN network is connected by optical fiber switch and comprises front end main frame, I/O node and sonet card interface disk, carries out data transmission; Ethernet connects the virtual management node by the ether switch and the I/O node is managed for configuration, resources allocation.Wherein, virtual management node and I/O node operation (SuSE) Linux OS (the above kernel of 2.4.18), the front end main frame is any mainstream operation system, can be Linux, Windows, Solaris and OPEN BSD.

Define following data structure:

Ax_pd_t: write down the attribute of some physical disks, comprise capacity, bandwidth, use space, free space and resource use table;

Ax_vd_t: write down the attribute of some virtual disks, comprise capacity and address mapping table;

Axum_SC: total size of all physical disks, virtual disk, total free space in the record total system, and comprise the chained list that a chained list of being made up of all ax_pd_t in the system and all ax_vd_t form;

PD: i.e. physical disk in the system;

VD: i.e. virtual disk in the system;

1) load resources allocation source and configuration module (axup.o): this module is carried out virtual fusion to the resource in the storage area network environment, for preceding end subscriber provides a same resource view and hides the physics realization details.Simultaneously,, finish resources allocation according to the capacity of physical disk, and the interface that provides disk to add and delete.This module is finished following function set by step:

1. at character device of linux kernel registration.This character device name is called AXUP, and it number is communicated by letter with the keeper to respond various configurations and resource allocation request by master-slave equipment;

2. provide a global variable axup_SC register system whole resource distribution situation;

3. Linux is provided "/proc " kernel information interface under the catalogue, so that user or management software are inquired about existing resource allocation information;

4. generate complete resource allocation information structure (physical disk, the spatial organization of virtual disk and operating position) by resource searching and configuration, and obtain the specifications parameter of each physical equipment;

2) load middle layer command processing module (stml_target.o): this module is the processing module of actual command.The resource mapping table that it provides according to ax_vd_t carries out address mapping and finishes the order of being accepted.

1. the forward end main frame is reported available virtual disk in the storage area network.

2. starting one in kernel is responsible for handling the thread of dummy command and is thread of handling the physics order of each disk startup.

3. the dummy command to receiving is split as several physics orders and execution after map addresses; Merge after being finished and return to front end.

Resource allocation step is as follows:

1) adds PD: utility command ./pdcreate a b c d; Wherein four parameters of a b c d are respectively the host channelid lun number of PD to be added.Following work is finished in this order:

1. distribute for PD and initialization ax_pd t data structure and resource use table, and it is added among the axup_SC;

2. carry out resource re-allocation, the reconstruct resource is used table and resource mapping table;

3. wake the copying data after background thread background_process_thread finishes reconstruct up;

2) deletion PD: utility command ./pdremove id; Wherein id is a numbering for the treatment of the deletion of physically dish.Following work is finished in this order:

1. whether there are enough resources to use after calculating deletion PD for VD.If not, then this PD can not delete.

2. carry out resource re-allocation, the reconstruct resource is used table and resource mapping table;

3. wake the copying data after background thread background_process_thread finishes reconstruct up;

4. the ax_pd_t structure of this PD of deletion from axup_SC discharges its resource use table.

3) create VD: utility command ./vdcreate-n name-s size; Wherein name is this VD name, and size is its size.Following work is finished in this order:

1. whether calculate among the SC has enough spaces to create this VD;

2. for VD creates and initialization ax_vd_t data structure and resource allocation table, it is added among the axup_SC;

3. be the VD Resources allocation, and the change resource is used table and resource mapping table;

Claims (1)

1. the resource allocation methods in the mass storage system (MSS) is characterized in that, contains following step successively:

Step (1): on the virtual management node of described system, load a resources allocation and resource distribution module and a middle layer command processing module, wherein:

Resources allocation and configuration module contain following data structure:

Ax_pd_t: write down the attribute of some physical disk PD, comprise capacity, bandwidth, use space, free space and resource use table; This resource uses table corresponding to PD not, and whether each the storage administration elementary cell SG that has write down corresponding PD is used and be mapped on which SG of which VD by certain virtual disk VD; Described PD and VD all are divided into equal-sized several SG;

Ax_vd_t: write down the attribute of the VD of some virtual disks, comprise capacity and resource mapping table; This resource mapping table is corresponding to some VD, write down the particular location of a certain PD that each SG of corresponding VD is mapped to;

Axum_SC: thus total size of physical disk PD, virtual disk VD, total free space in the record total system, and comprise the chained list that a chained list of being made up of all ax_pd_t in the system and all ax_vd_t form;

Described resources allocation and configuration module are realized resources allocation in the following manner:

Set: n piece physical disk is arranged in the system, use PD ₁, PD ₂PD _nExpression, the evaluation of estimate of representing with capacity is respectively C ₁, C ₂C _n

Set of data objects S={SG to be distributed ₁, SG ₂SG _m, m is the size of S set;

Definition PD ₁Capacity region between be (0, C ₁), PD ₂Capacity region between be (C ₁, C ₁+ C ₂) ..., PD _nCapacity region between be

Then: when creating VD or adding PD, S is a data object all among the SAN; When deletion PD, S is the data object of waiting to delete on the PD;

For each the object SG that treats among the distributed data object set S _i, produce a random number r, when r drops on PD _jCapacity region between the time, with SG _iBe distributed in PD _jIn;

Create PD _nFlow process as follows:

Step (1.0) is PD distribute data structure ax_pd_t, fills the base attribute that comprises capacity, available size, the resource allocation table of initialization PD; Defining variable P also is initialized as 0;

Step (1.1) joins PD in the virtual management node all PD is organized into a unified storage pool SC virtually;

Step (1.2), the VD that exists among the SC is carried out resource re-allocation according to the following steps:

Step (1.2.1) is to all physical disk PD that has distributed _i, i=1,2 ..., n-1, definition δ _i=C _i/ S '-C _i/ S, wherein $S=\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{C}_i$ For comprising PD _nIn the total volume of all interior physical disks, $S^{\′}=\underset{i=1}{\overset{n-1}{\mathrm{\Σ}}}{C}_i,$ Be all total volumies of the physical disk of Resources allocation;

Step (1.2.2) is for all SG _iProduce the random number r that (0,1) interval is evenly distributed, described SG _i∈ PD _iAnd all SG _iDistributed to VD, i=1,2 ..., n-1;

Step (1.2.3) is for this SG _iAffiliated PD _i, if r＜δ _i, this SG then _iMove to PD _nP SG in, and revise corresponding resource allocation table and resource mapping table, P adds 1; Otherwise this SG _iRemain unchanged;

Step (1.2.4), repeating step (1.2.3) finishes up to i=n-1;

Deletion PD _nFlow process as follows:

Step (2.0) is for this PD _nGo up the set of all SG that distributed, produce The n-1 that the interval is evenly distributed random number r;

Step (2.1) is if random number r _iBe positioned at PD _iCapacity region between on, then this SG is moved to PD _iP _iOn the individual SG; Otherwise this SG remains unchanged; P _iBe PD _iUpward first is not assigned to the position of the SG of VD by sequence of addresses, i=1, and 2 ..., n-1;

Step (2.2), search PD _iMake P _iPoint to PD _iThe next SG that is not assigned with, and revise corresponding resource allocation table and resource mapping table;

Step (2.3) is PD _nFrom SC, delete;

The flow process of creating VD is as follows:

Step (3.0) is calculated remaining space sum among the SC whether greater than the capacity of the VD that will create.If not, then can not create this VD; Otherwise change next step;

Step (3.1), the data structure of establishment VD is also added among the SC;

Step (3.2) is as if the VD that is that will create _nIts step is as follows:

Step (3.2.1) is to this VD _nGo up the set of all SG, produce The n that the interval is evenly distributed random number r;

Step (3.2.2) is if r _iBe positioned at PD _iCapacity region between on, then this SG is assigned to PD _iP _iOn the individual SG; P _iBe PD _iUpward first is not assigned to the position of the SG of VD by sequence of addresses, i=1, and 2 ..., n;

Step (3.2.3), search PD _iMake P _iPoint to PD _iThe next SG that is not assigned with, and revise corresponding resource allocation table and resource mapping table;

Middle layer command processing module, the resource mapping table that provides according to ax_vd_t are carried out address mapping and are finished the order of the front end main frame of being accepted;

Step (2), described method are carried out resources allocation according to the following steps successively:

Adding PD _nThe time:

At first, be this PD _nDistribute and initialization ax_pd_t data structure and resource use table and it is joined among the global variable axup_SC;

Secondly, carry out resource re-allocation, heavily purchase resource use table and resource mapping table;

At last, wake background thread up, this thread carries out copying data according to resource use table after heavily purchasing and resource mapping table; At deletion PD _nThe time:

The first step, deletion PD _nWhether in time, need shift the data of VD on it, have enough resources to use for VD behind the calculating deletion PD.If not, then this PD can not delete;

Second step, carry out resource re-allocation, the reconstruct resource is used table and resource mapping table;

The 3rd step, wake background thread up, finish the copying data after heavily purchasing;

In the 4th step, the ax_pd_t structure of this PD of deletion discharges its resource use table from axup_SC; When deletion VD:

At first, whether there are enough spaces to create this VD among the calculating SC;

Secondly, for VD creates and initialization ax_vd_t data structure and resource allocation table, it is added among the axup_SC;

At last, be the VD Resources allocation, and the change resource is used table and resource mapping table.

Images