首页 > 移动平台 > 详细

Method and Apparatus for Providing Highly-Scalable Network Storage for Well-Gridded Objects

时间:2014-08-07 03:01:28      阅读:377      评论:0      收藏:0      [点我收藏+]

标签:des   style   src   Go   com   log   ui   http   

An apparatus comprising a plurality of storage nodes comprising a plurality of corresponding storage disks and configured to store data in a distributed manner between the storage disks that achieves a Redundant Array of Independent Disks-0 (RAID0) like performance based on positioning information and without indexing the distributed data. A network component comprising a storage disk configured to maintain a plurality of physical files for different user data that are mapped to different volumes, wherein the volumes are distributed between the storage disk and a second storage disk based on a?RAID0?like data distribution scheme without being indexed.

BACKGROUND

Cloud storage is a model of network online storage where data is stored on multiple virtual servers, generally hosted by third parties, rather than being hosted on dedicated servers. A customer that requires data hosting services can buy or lease storage capacity from cloud storage hosting providers that operate large data centers. The data center operators provide resources in the background, referred to as virtual resources, according to the requirements of the customer and expose them as storage pools, which customers can use to store files or data objects. The resource may be distributed across multiple servers. Typically, cloud storage services may be accessed through a web service application programming interface (API) or through a Web-based user interface. In some cloud storage systems, service providers provide block based storage services to customers or users. Since a provider typically owns the infrastructure of the storage service, it is beneficial for the provider to pursue efficient and highly-scalable design and implementation for providing the storage service to customers. Thus, one important design and implementation objective for providers is to maximize throughput and scalability, such as in the case of relatively large systems.

SUMMARY

In one embodiment, the disclosure includes an apparatus comprising a plurality of storage nodes comprising a plurality of corresponding storage disks and configured to store data in a distributed manner between the storage disks that achieves a Redundant Array of Independent Disks-0 (RAID0) like performance based on positioning information and without indexing the distributed data.

In another embodiment, the disclosure includes a network component comprising a storage disk configured to maintain a plurality of physical files for different user data that are mapped to different volumes, wherein the volumes are distributed between the storage disk and a second storage disk based on a?RAID0?like data distribution scheme without being indexed.

In a third aspect, the disclosure includes a method comprising receiving a request to write data to a storage system, wherein the request comprises positioning information, mapping the positioning information into a data volume, data block, and data sector combination, hashing the data volume, data block, and data sector combination, placing a first strip of the data at a location in the storage system based on the hash, and distributing one or more second strips of the data at one or more subsequent locations in the storage system that are separated from the location by an offset.

These and other features will be more clearly understood from the following detailed description taken in conjunction with the accompanying drawings and claims.

DETAILED DESCRIPTION

It should be understood at the outset that although an illustrative implementation of one or more embodiments are provided below, the disclosed systems and/or methods may be implemented using any number of techniques, whether currently known or in existence. The disclosure should in no way be limited to the illustrative implementations, drawings, and techniques illustrated below, including the exemplary designs and implementations illustrated and described herein, but may be modified within the scope of the appended claims along with their full scope of equivalents.

Disclosed herein is a system and method for providing a block based storage system with improved efficiency and scalability in comparison to other cloud storage systems. The design and implementation of the block based storage system may be suitable for handling file systems that are maintained on a plurality of storage units, and thus provides better efficiency and scalability than other local hard disk drive (HDD) based systems. The block based storage system may scale the storage services efficiently with increasing user demand and without the need for substantially large data volume redistribution among different serving or storage nodes, e.g. in a network. The block based storage system may be managed with relatively less operational requirements and thus less operational expense (OPEX) in comparison to other storage systems. The system scalability may be improved using a self-managed distributed hash table (DHT) based method, which may leverage the topology or hierarchy of the physical infrastructure of the storage system and the characteristics of target storage layout. The system throughput efficiency may also be improved using a data distribution mechanism, which may achieve?RAID0?like performance in a relatively highly-distributed and highly-scalable manner. Additionally, a file layout may be used to improve the HDD or storage unit read-write efficiency.

FIG. 1?illustrates an embodiment of a storage system topology?100, which may be used to provide storage services to a plurality of customers or users. The storage system topology?100?may correspond to a "sea of disk" (SoD) topology that is introduced by HUAWEI Core Networks Business Unit (BU) and which is suitable for providing data storage for cloud computing purposes. The storage system topology?100?may comprise a plurality of cases?110, which may each comprise a plurality of disks?120. The disks?120?may comprise a plurality of ATOM, ARM, and/or other processor type based computers. Each of the computers may also comprise other components, such as a Central Processing Unit (CPU), a random access memory (RAM), a Flash/Solid State Drive (SSD), a HDD, a one Gigabit per second (1G) Ethernet card, or combinations thereof, as described below.

bubuko.com,布布扣

As shown in?FIG. 1, each case?110?may comprise multiple disks?120. For instance, a case?110?may correspond to a shelf cabinet that comprises a stack of disks?120. The disks?120?and the cases?110?correspond to two hierarchical levels of the storage system topology?100, which may comprise other levels (not shown) that provide a hierarchical storage nodes cluster architecture. For instance, a plurality of cases?110?may be part of a unit, such as a unit of multiple cabinets, and a plurality of units may be part of a container. A plurality of containers may also be part of a data center. The different storage levels of the storage system topology?100?may correspond to or may be accessed and maintained by a plurality of storage nodes, for example in a network. For instance, each of the cases?110?may correspond to a storage node. Alternatively, a plurality of cases?110, or less than all of the disks?120?in the case110, may correspond to a single storage node.

The storage services provided by the storage system topology?100?may include a Virtual Block Service (VBS), which is a block based storage service. In the case of well-gridded object architecture, such as in the storage system topology100, the VBS may support APIs or API functions similar to:


put(start_position_key<address_bits>,data<bytes>), and


get(start_position_key<address_bits>,number_of_items<int>).

The get method above may be supported due to the well-gridded object architecture, for example where data objects may be distributed about proportionally over a plurality of storage nodes. In the put method, the size of data may be used to define or indicate a data object.

When a user or customer requests to read and/or write some files, the storage system topology?100?supported file system may redirect such request to the VBS using an Advanced Technology Attachment (ATA) over Ethernet (AoE), for instance to access the storage system devices over an Ethernet network. Using the AoE protocol to access the storage system devices in a VBS may be optional and other protocols or schemes may be used instead.

In a VBS system, the storage system devices may be accessed using positioning information, such as <Volume, Block, Sector>. The Block data and Sector data sizes may be fixed and the Volume size may be expanded according to user‘s demand. On the user side, a local file system may manage the mapping of user visible positioning file information, such as <file path, offset>, to the appropriate<Volume, Block, Sector> addressing information for placement of the file‘s data in the VBS system. Specifically, the user‘s local file system may be configured to manage mapping files into locations in the data Volume, e.g. in the storage target node(s) or disk(s), which may be divided into a grid of blocks and sectors within blocks of known and fixed sizes.

To address the scalability issue of the storage system, a self-managed DHT scheme may be used to leverage or take advantage of the system‘s physical infrastructure topology, such as the storage system topology?100, and other special characteristics that may be related to the target storage layout. For example, in the case of the storage system topology100, it may be difficult to achieve non-blocking Input/Output (I/O) over each pair of disks?120?without substantial investment in the switching and routing infrastructure for the system. To avoid such cost, a two-layered DHT may be used, which may handle an inter-case level (e.g. between the disks?120) and an intra-case level (e.g. between the cases?110). The two-layer DHT may comprise a first DHT for handling the inter-case level, which may be self-managed or maintained by a storage target or node associated with each case?110. The two-layer DHT may also comprise a second DHT for handling the intra-case level, which may be distributed between multiple storage nodes that correspond to different cases?110.

The DHT may be a decentralized distributed system (among different storage nodes) that provides a lookup service similar to a hash table. The DHT may comprise a plurality of (key, value) pairs, which may be used by the storage nodes to retrieve a value associated with a given key. The mapping from keys to values in the DHT may be distributed among the storage nodes, where a change in the set of storage nodes may not substantially change the DHT entries. As such, the DHT may scale to substantially large number of storage nodes and handle the arrivals, departures, and/or failures of storage nodes.

To address the throughput efficiency, a suitable data distributor may be used. The data distributor may be configured to achieve a?RAID0?like performance in a distributed manner (between the storage nodes). The data distributor may also be configured to determine how to split and distribute the data of one volume to a number of storage nodes to achieve a relatively high throughput. Typically, existing DHT schemes may use hash functions, such as Secure Hash Algorithm (SHA)-1, to randomize the data distribution. In this case, the output key resulting after performing the function hash(<Volume, Block, Sector>) may be evenly distributed without ordering data blocks and/or sectors, e.g. according to a block number and/or a sector number. In implementing such distribution scheme without ordering blocks and/or sectors, the data placement and retrieval operations may consume a substantial amount of network bandwidth, and cause a lot of random access to each storage node.

One alternative for data distribution is using the Cassandra‘s "order preserved partitioner" method. However, in a relatively large scale data center, this method may lead to storing one specific volume on one or two storage nodes, which may result in low performance when users attempt random read/write access. Such method may negatively impact volume access (e.g. uneven volume distribution) by multiple applications or a single application that comprises multiple threads. The negative impact on volume access may increase as the level of concurrency and volume increases.

FIG. 2?illustrates an embodiment of a data distributor scheme?200, which may be used to improve random data distribution and volume access in a storage system. The data distribution scheme?200?may be used with the DHT scheme above, e.g. based on the storage system topology?100. The data distribution scheme?200?may also support an order preserving partitioner (OPP), such as the OPP included in the Cassandra version 0.6 application. The data distribution scheme?200may use a Volume identifier (ID), or a user ID if the Volume is binding to a user, as an input parameter of a hash function, F. The data distributor?200?may also use a function that defines the range of the hash function F, Range(F). The function Range(F) may define a "ring" of cyclical values that starts with a minimum value, Min(Range(F)), and ends with a maximum value, Max(Range(F)). Both the hash function F and the range function Range(F) may be known quantities used in the data distributor scheme?200.

bubuko.com,布布扣

The hash function F may be used to generate a random number that indicates a random position on a DHT ID space. The DHT may be cyclical or have algebraic cyclical properties (recurring values), where ID values may belong to a range of cyclical or recurring values (a ring). The data distribution scheme?200?may comprise a DHT cluster that corresponds to the DHT ID space. The DHT cluster may comprise a plurality of nodes?210?(e.g. storage or host nodes) that may be arranged logically in a ring configuration, also referred to herein as a cluster node configuration (NC). The nodes?210?may correspond to the ID values in the DHT ID space.

FIG. 2?illustrates a case where a user may designate a Volume, V, using the Volume‘s ID, V_id, for example to read/write or access the Volume V. Thus, the function F(V_id) may be used to generate a random position for the Volume V on the DHT ID space. When the function F(V_id) is computed, the location of V_id (or Volume V) may be determined. Specifically, the node?210?that has the least largest value with respect to F(V_id) (e.g. equal or closest higher value) may be designated to store a first strip of Volume V, Strip-0.

A "strip placement function", Node, may be used to map a strip of a volume to a host node as follows:


Node(G(V_id,F,d,strip_number),NC),

where G is a "magic step function", strip_number(s) is a function of a block_id and the number of blocks that may fit in a strip (or block_id plus sector_id and the number of sectors that may fit in a strip), NC is the node configuration, and Node is the strip placement function. The strip placement function Node may provide the target or host node for a strip of the volume. The "magic step function" may start from F(V_id) and make a step size d around the ring of nodes?210, such as:


G(V_id,F,d,s)=F(V_id)+(d*s)/Range(F).

The Node may be a function of node configuration(NC), a hash function(F) that selects the "starting" position of a volume (e.g. Volume V), a step size(d), and a strip_number(s) function.

The data distribution scheme?200?may introduce a system parallelism parameter N that mimics the number of disks in aRAID0?system, where one volume may be distributed and stored on N disks. As such, the data distribution scheme?200?may store a plurality of strips of one volume on N nodes. The strip_number for each strip may depend on the block_id and N, where multiple blocks may fit into one strip. This relationship between the strip_number, the block_id, and N may be described as:


strip_number=(block_id/s)/N.

The strip_number of the first strip of the volume may be set to zero (Strip-0), the strip_number of the second strip of the volume may be set to one (Strip-1), and strip_numbers of subsequent strips may be set similarly. As described above, each strip_number may be mapped to a node?210?using the Node function. For example, the Strip-1‘s hosting node may be denoted by:


Node(G(V_id,F,d,1),NC).

For purposes of brevity, the "magic step function" G may be dropped from the expression above and hence the target or host node may be denoted as:


Node(V_id,F,d,strip_number,NC).

The number of host nodes N for a volume, N, the step size or offset d, and the number of blocks in each strip(s) may be defined in different ways. These parameters may be defined for a volume as a fixed header to a first strip, Strip-0. Alternatively, the parameters may be dynamically assigned as a configuration file for a volume at a node?210?where the first strip may be expected, for example using Node(V_id, F, d, 0, NC). The second approach may be more attractive since the second approach may separate the volume stripping meta data from the volume‘s actual data. The function Node(V_id, F, d, 0, NC) to obtain the first strip may be independent of d, which may be used when calculating the host nodes from the nodes?210?except the first node?210?that is designated for the first strip (Strip-0).

In one strip placement algorithm, the number of nodes N that may be designated for the volume‘s different strips may be fixed. For example, N is equal to four in?FIG. 2. The offset d may also be fixed, for example d=(Range(F))/N), where N and Range(F) are fixed. The strip placement algorithm may also use a cyclic pattern, e.g. using the strip placement functions above, to distribute the blocks within a volume to appropriate target strips on host nodes. As such, a plurality of strips may be mapped on one or more nodes?210. For example, the first node (labeled 0) may store Strip-0 and Strip-4, a second node (labeled 2) at offset d may store Strip-1 and Strip-5, a third node (labeled 4) may store Strip-2 and Strip-6, and a fourth node (labeled 6) may store Strip-3 and Strip-7.

In another strip placement algorithm, an offset d value may be varied and the Range(F) may or may not be divisible by d. Thus, N may be varied for different data volumes to maintain an about equal storage consumption or data distribution on each strip hosting node. Other algorithms or combinations of the two algorithms above or some of their features may also be used for volume strip placement on the nodes?210.

The data distribution scheme?200?may determine the host node for a <Volume, Block, Sector> combination for requested data using relatively simple calculations, as described above. As such, the user may fetch multiple strips in different nodes if needed in a parallel fetching scheme, where the strips may be obtained at about the same time from the different nodes210. A local file system at the user environment may perform the parallel fetching scheme, for instance, for a volume that comprises a plurality files, which may be accessed by a multi-threaded application. The data distribution scheme?200?may place or access the distributed data over the host nodes without using an index for data placement by taking advantage of the representation of a volume as a group of fixed-sized and well-gridded blocks and sectors.

The distribution scheme?200?and the DHT scheme that matches the storage system topology?100?may be applied to any well-gridded objects, for example in the case of a VBS, where the objects may be treated as volumes of distributed data. In well-gridded object architectures, the schemes above may allow data distribution without indexing, which may improve object placement and retrieval and avoid unnecessary index calculation and manipulation. Unlike other DHT based VBS storage systems or storage systems based on target storage objects that use indexing, the DHT, and data distribution schemes above may use logical indexing by implementing the hashing functions above based on the objects and other parameters to substantially improve storage system performance. The data distribution scheme?200?may be used with different DHT schemes that may be currently available. For example, additional functions or plug-ins may be added to some existing DHT schemes to implement the DHT and data distribution schemes above.

FIG. 3?illustrates an embodiment of a file layout?300, which may be used to improve random data distribution and volume access in a storage system, e.g. according to the storage system topology?100. The file layout?300?may be used in combination with the DHT and the data distribution schemes described above to improves the storage system read-write or throughput efficiency. The file layout?300?may be correspond to N=2 in the data distribution scheme?200. As such, the file layout?300?may comprise about two node storing strips?310?(labeled N1?and N2). Each node storing strip?310?may comprise about two physical files?320?(labeled F1, F2, F3, and F4), which may each comprise about two data blocks?330?(labeled B1to B8). Each data block?330?may also comprise about two data sectors?340.

bubuko.com,布布扣

The size of each physical file?320?may be fixed, for example equal to about 10 Megabytes (MB). The size of the physical file320?(e.g. data file) may be determined according to the number of data blocks?330?in the physical file?320, which may correspond to the number of strips within a data volume strip_number or s. The fixed data file size may simplify operations and improve sequential access efficiency in the node storing strips?310?(e.g. HDDs). The blocks?330?may be placed sequentially according to the corresponding block IDs inside each physical file?320. The block IDs may be distributed by a cyclic policy among the node storing strips?310. The sectors?340?may also be placed logically in a sequential manner inside each block?330?and may be ordered by corresponding sector IDs. The sequential placement of blocks and sectors in the file layout?300?may lead to a?RAID0?like performance and increase the storage system throughput linearly. Further, the physical files?320?may be created according to a Lazy Creation scheme, where a physical file?320?may not be created until one sector?340?of that file is accessed, for example by a write operation to that sector from a user.

FIG. 4?illustrates an embodiment of a data distributor method?400, which may be used to distribute data in a storage system. The data distribution method?400?may be implemented by the storage system in the data distribution scheme?200and support the DHT scheme based on the storage system topology?100?and the file layout?300. The data distribution method?400?may be used to place or fetch user requested data on or from a target storage node. The method?400?may begin at block?410, where a request to write or read data to a storage system may be received including positioning information. The request may be a VBS supported get or put API function that may be sent by a user‘s local file system. At block?420, the positioning information may be mapped into a data volume, data block, and data sector combination. The user‘s local file system may translate the user request from user visible positioning file information, such as <file path, offset>, to appropriate <Volume, Block, Sector> addressing information for placement of the data or file in the VBS system.

bubuko.com,布布扣

At block?430, the data volume, data block, and data sector combination may be hashed. At block?440, a first strip of the data may be placed or fetched at a location in the storage system based on hashing. For instance, the data volume, data block, and data sector information may be hashed using the hashing functions and parameters described in the data distribution scheme?200?to obtain a location to begin storing the data. The data volume, data block, and data sector and the hashing value (e.g. location) may be saved as a value and key pair, respectively, in an entry in a DHT, which may be stored locally at the storage node associated with the location. At block?450, one or more second strips of the data may be stored or retrieved at one or more subsequent locations in the storage system that are separated by an offset. The data strips may be distributed in a cyclical manner between the storage nodes, as described above. The data may be distributed in different nodes in a plurality of strips that comprise a sequential order of blocks and sectors, as indicated by the positioning information. Thus, the data or file may be accessed or retrieved in the order indicated by the corresponding strips, blocks, and sectors in the nodes. The method?400?may then end.

The network components described above may be implemented on any general-purpose network component, such as a computer or network component with sufficient processing power, memory resources, and network throughput capability to handle the necessary workload placed upon it.?FIG. 5?illustrates a typical, general-purpose network component?500?suitable for implementing one or more embodiments of the components disclosed herein. The network component?500?includes a processor?502?(which may be referred to as a central processor unit or CPU) that is in communication with memory devices including secondary storage?504, read only memory (ROM)?506, RAM?508, I/O devices?510, and network connectivity devices?512. The processor?502?may be implemented as one or more CPU chips, or may be part of one or more application specific integrated circuits (ASICs).

bubuko.com,布布扣

The secondary storage?504?is typically comprised of one or more disk drives or tape drives and is used for non-volatile storage of data and as an over-flow data storage device if RAM?508?is not large enough to hold all working data. Secondary storage?504?may be used to store programs that are loaded into RAM?508?when such programs are selected for execution. The ROM?506?is used to store instructions and perhaps data that are read during program execution. ROM506?is a non-volatile memory device that typically has a small memory capacity relative to the larger memory capacity of secondary storage?504. The RAM?508?is used to store volatile data and perhaps to store instructions. Access to both ROM506?and RAM?508?is typically faster than to secondary storage?504.

SRC=https://www.google.com.hk/patents/US20120005424

Method and Apparatus for Providing Highly-Scalable Network Storage for Well-Gridded Objects,布布扣,bubuko.com

Method and Apparatus for Providing Highly-Scalable Network Storage for Well-Gridded Objects

标签:des   style   src   Go   com   log   ui   http   

原文:http://www.cnblogs.com/coryxie/p/3896095.html

(0)
(0)
   
举报
评论 一句话评论(0
0条  
登录后才能评论!
© 2014 bubuko.com 版权所有 鲁ICP备09046678号-4
打开技术之扣,分享程序人生!
             

鲁公网安备 37021202000002号