JP4891657B2 - Data storage system, file search device and program - Google Patents

Description

  The present invention relates to a data storage system in which a plurality of nodes are connected in a grid, and a file search apparatus and program used in this system.

  Due to the increase in the number of files exchanged on the network and the size of individual files, the storage capacity required for database construction is increasing year by year. However, enormous costs are required to construct a storage area network (SAN) or the like to ensure a sufficient storage capacity.

Therefore, there is a demand for constructing a large-capacity storage system using a plurality of relatively inexpensive servers or personal computers. In such a system, even if the storage capacity of the individual nodes constituting the system is not sufficient, it is required to logically integrate them to store data as if it were a single storage area, and to retrieve the data. Is done.
JP 7-129450 A

  In a system using a plurality of nodes as described above, when there is some regularity in the connection between the plurality of nodes and the nodes, index data including a pointer indicating a file storage position is prepared. . This index data is used when retrieving data. If the correspondence between the index data storage area and the actual data storage area is improperly designed, it will not be scalable, it will take a long time to search for files, or the data volume of the index data and the actual data volume There may be a problem that the balance is not achieved.

  The present invention has been made in view of such a situation, and an object of the present invention is to manage files held in a plurality of nodes each having a storage device by using a common index to function as a single database. The object is to provide a technique for constructing a storage system.

  One embodiment of the present invention is a data storage system that manages files held in a plurality of nodes each having a storage device using a common index to function as a single database. A plurality of nodes are arranged in a lattice pattern, and each node is connected to be communicable with the front, rear, left and right nodes. The file storage node that actually stores the file and the index node that stores the index data of the file storage node are each divided so as to form a square partial lattice. Information on the tree structure for managing files and index data is stored in the index node by associating the file storage node with the leaf of the tree structure and the index node with the root or node of the tree structure. Then, the address information of the file storage node is stored in the index node that is uniquely determined based on the file specifying information for specifying the file held in the file storage node.

  According to this aspect, the storage areas provided in each of the plurality of nodes can be logically integrated and used as one storage area. Therefore, an inexpensive computer or server can be combined instead of a large-capacity storage device to replace the large-capacity storage device. In addition, since the address information of the file storage node is held in the index node determined based on the file identification information, even if the address of the file storage node where the file is actually stored is unknown, the file identification information is even If so, the storage location of the desired file can be easily specified. The “file specifying information” may be information unique to a file, and includes, for example, a file name, a file creation time, an update time, a file creator, a computer name that created the file, and a combination thereof.

  The number of nodes arranged in the vertical direction and the number of nodes arranged in the horizontal direction in the grid-like array are relatively prime, and the grid-like array is divided into a plurality of square sub-lattices using the Euclidean algorithm. Also good. According to this, it is possible to easily divide arbitrary m × n nodes that are relatively prime into a plurality of square sub-lattices. In addition, when square division that is naturally derived from Euclidean mutual division is used, nodes in a parent-child relationship or sibling relationship are located close together in the tree structure, and the parent-child direction or sibling during file search or file storage The access time in the direction can be reduced. In addition, if the file storage node is a square sub-grid, multiple paths to the destination node will be created when transferring the file, so the file can be transferred even if part of the connection between the nodes is broken. Can be realized.

  The file specifying information may be coded according to a predetermined rule, and a node that should store a file corresponding to the file specifying information may be determined according to the obtained code. In this case, a hash function may be obtained by applying a hash function to the code, and an index node that should hold the address information of the file storage node may be determined according to the hash value.

  Another aspect of the present invention is a file search program in the above-described data storage system. The file search program operates on the index node, receives a file search request, encodes file specifying information of the file according to a predetermined rule, and stores a file corresponding to the file specifying information according to the obtained code. A function for determining a file storage node that is stored, a function for obtaining a hash value by applying a hash function to the code, and a function for determining an index node that holds the address information of the file storage node according to the hash value .

  According to still another aspect of the present invention, when a plurality of nodes each having a storage device are arranged in a grid pattern and each node is connected to the front, rear, left, and right nodes so as to be able to communicate with each other, This is a data storage system managed with a B-tree structure. The roots, nodes, and leaves of the B-tree structure are associated with any one of a plurality of nodes arranged in a grid. The node corresponding to the leaf actually stores the file, the node corresponding to the node stores address information indicating the node corresponding to the leaf included in the subtree, and the node corresponding to the root Stores address information indicating the node corresponding to the node.

  According to this aspect, a plurality of nodes arranged in a grid and a known B-tree can be combined to efficiently manage files arranged in a plurality of nodes by using the B-tree.

  It should be noted that any combination of the above-described components and a representation of the present invention by a method, apparatus, system, recording medium, and computer program are also effective as an aspect of the present invention.

  According to the present invention, files held in a plurality of nodes each having a storage device can be collectively managed and function as a single database.

  One embodiment of the present invention is a data storage system that manages a set of storage devices arranged in each node in a system in which a plurality of nodes each including a processor are arranged in a grid pattern. In particular, it is characterized by an indexing technique for searching a storage location of a file. Hereinafter, this embodiment will be described in detail with reference to the drawings.

  FIG. 1 is an overall configuration diagram of a data storage system 100 and a client terminal 12 connected to the data storage system 100 according to an embodiment of the present invention. The “data storage system” targeted by this embodiment refers to a system in which a plurality of nodes each having a processor, such as a server or a personal computer, are connected in a grid pattern, and data is distributed and arranged on the plurality of nodes. .

  As shown in FIG. 1, the data storage system 100 includes a grid array 10 that searches and provides a file stored in the system in response to a search request issued from a client terminal 12. In the grid array 10, nodes 20 are arranged so as to form a grid of a plurality of columns and multiple rows (5 rows and 7 columns in FIG. 1). Each node in FIG. 1 corresponds to one server or personal computer, and each node is represented by a white square. These nodes arranged in a lattice shape are configured to be able to communicate with nodes located on the top, bottom, left, and right. In FIG. 1, the grid array 10 has 5 rows and 7 columns, but it is needless to say that it may be composed of a larger number or a smaller number of nodes. This will be described later.

  Each node 20 in the grid array 10 is connected to a network 14 such as the Internet, a LAN, and a WAN via a router 16. The data storage system 100 is arranged in a company data center or the like, and can respond to a large number of search requests simultaneously.

  The grid array 10 functions as one database as a whole. Information such as various programs and IP addresses necessary for exhibiting this function is given to each node in advance by the system administrator. In another embodiment, a program and various data stored in a specific node may be transmitted to each node when the role of each node described later is determined.

  The client terminal 12 may be a personal computer including an input device such as a keyboard and a mouse and an output device such as a display, or a mobile phone including an input / output device equivalent thereto. However, in the case of a mobile phone, wireless communication is assumed. The user issues a search request to the data storage system 100 using a web browser or the like on the client terminal 12.

  FIG. 2 is a hardware configuration diagram of each node constituting the grid array 10. The node includes a processor 92 that executes various processes according to a program, a memory 94 that temporarily stores data and programs, a storage device 96 that does not lose recorded contents even if the node is restarted, A network interface 98 for executing various input / output processes connected to the nodes, and a bus 90 for interconnecting them. Each node includes a maximum of four network interfaces 98 for connecting to other nodes located on the top, bottom, left, and right as shown in FIG. As the storage device 96, a hard disk device, an optical disk device, a magnetic disk device, a nonvolatile memory, or any other device can be used. Each node may have an input device such as a keyboard and a mouse and an output device such as a display as necessary.

  Each node may be a blade server equipped with a compact shape suitable for configuring the data storage system 100, that is, a processor, a memory, a hard disk, a bus, and the like. The grid array 10 is preferably arranged so that a large number of blade servers are arranged in a rack in order to link the servers to each other, but other modes may be used.

  FIG. 3 is a diagram illustrating a state in which each node 20 is linked to the upper, lower, left, and right nodes. As shown in the figure, the node 20 is connected to the upper, lower, left, and right nodes so as to be in a peer-to-peer connection, and an IP address is given in advance for each link. Therefore, each node needs to have the same number of network interfaces as neighboring nodes.

  Each node 20 is configured so that a plurality of OSs can be started simultaneously using a known virtualization technology such as hyperthreading or VMware. If the virtualization technology is used, it is not necessary to have two or more CPUs for simultaneously starting a plurality of OSs. One OS manages application execution (hereinafter, this OS is referred to as “application OS”), and the other OS manages routing (hereinafter, this OS is referred to as “autonomous disk OS”). Considering each node 20 virtually divided into an application node 22 that executes the application OS and an autonomous disk node 24 that executes the autonomous disk OS, the application node 22 is connected to the other node 20 with the autonomous disk node 24. By leaving the determination of the communication path, communication between any nodes 20 in the grid array becomes possible. Since the routing function is realized by the autonomous disk node 24 and the network interface, it is not necessary to arrange a switch or a router between the nodes. However, this embodiment can be realized even with a conventional network configuration in which a plurality of nodes are grouped in units of rows and columns, and routers are arranged in each of them.

  In the following description, it is assumed that the application OS and the autonomous disk OS are running on all nodes, and that routing and application execution are possible. It is assumed that storage and retrieval of data files in the storage device, hash calculation described later, and the like are executed by the application node 22, and file transfer processing and routing are executed by the autonomous disk node 24, and their roles are not particularly distinguished. I will explain it.

  By the way, assuming a data storage system composed of a plurality of nodes arranged in a grid as shown in FIG. 4, for example, conventionally, how many nodes should be allocated according to the storage capacity expected by the system administrator. Decide what is good. For example, as shown in FIG. 4A, assume that 2 × 2 = 4 nodes are initially allocated. If the subsequent operation requires a storage capacity that exceeds the initial expectation, the system administrator needs to increase the number of nodes to be allocated. At this time, it is assumed that the additional node is assigned as shown in FIG. Then, it becomes impossible to maintain consistency between the arrangement rule of the data file stored in each node and the arrangement rule of the data file stored after adding the node, and the index data becomes different. Therefore, in order to use a large number of nodes arranged in a grid as a database, it is necessary to determine nodes where data is arranged according to a certain rule and maintain the consistency of index data.

  Further, as shown in FIG. 5, assuming that the storage capacity is increased from the state in which the index node of 2 rows and 1 column and the actual data storage node of 2 rows and 2 columns are allocated, the index data is increased as the actual data increases. Will also increase. For this reason, it is necessary to distribute the index data to a plurality of nodes, and an index data division rule is required. For example, as shown in FIGS. 5 (a) and 5 (b), if the node for storing the index data is determined to be the left column of the block, the allowable amount of the actual data is a quadratic expression that is a product of length and width. However, since the allowable amount of index data increases only by the ratio of a linear expression in a vertical column, the storage area of the index data is restricted and access to actual data is restricted. As a result, relocation of index data is inevitably required. If this rearrangement is neglected, all nodes having an index must be queried in order to search for actual data.

  Therefore, there is a need for a method that can appropriately determine the arrangement of the actual data and the index data when a certain grid arrangement is given. In the present embodiment, the above-described problem is solved by configuring the number of nodes in the vertical and horizontal directions of the lattice-like array by a pair of prime integers. More specifically, the physical arrangement of the nodes in the grid array and the management of the index data by a well-known B-tree for hierarchizing the index data are used in combination.

  As a premise for managing the index data according to the present embodiment, the nodes constituting the grid array are stored as “file storage nodes” that actually store files, and index data including address information of the file storage nodes is stored. It is necessary to divide it into “index nodes”. Here, “index data” is data necessary to specify a file storage node, and in the embodiment described later, is a set of a node number and address information of the file storage node.

  With reference to the flowchart of FIG. 6, a procedure for dividing the grid array 10 will be described. First, the system administrator constructs a grid array in which m and n are prime when m is the number of nodes in the vertical direction and n is the number of nodes in the horizontal direction (m, n is a natural number). Then, the IP address of each node is set (S10). The reason why the number of nodes in the vertical and horizontal directions is relatively prime is that it is guaranteed that the index nodes can be hierarchically formed by dividing into squares by the Euclidean mutual division method described later.

  Next, the system administrator divides the grid array into a plurality of square partial grids having different sizes (S12). In order to divide the lattice-like array 10 in which the number of nodes in the vertical and horizontal directions are relatively prime into a plurality of square sub-lattices, a well-known Euclidean algorithm is used. Hereinafter, a method of dividing the nodes into the square partial regions having the same number of nodes in the vertical and horizontal directions by applying the Euclidean mutual division method to the nodes arranged in a grid in the present embodiment will be described.

1. A latticed array having a natural number m and n nodes such that m> n is denoted as K (m, n).
2. From the left side of K (m, n), the square partial lattice K (n, n) is packed. If the number of square and _{q 0,} the quotient obtained by dividing m by n is _{q 0,} the remainder can be expressed as _{(m-q 0 · n)} .
3. When the remainder (mq ₀ · n) = r ₁ is expressed, the remaining part filled with the square partial lattice K (n, n) can be expressed as K (r ₁ , n).
4). From the lower side of the rectangle K (r ₁ , n), this time, the square partial lattice K (r ₁ , r ₁ ) is packed. When the number of square sublattices and _{q 1,} the quotient obtained by dividing the n in _{r 1} is _{q 1,} remainder expressed as _{(n-q 1 · r 1} ).
5. When the above operation is repeated, the grid array K (m, n) is divided into a plurality of square partial grids in a finite number of times.

  The method of dividing an arbitrary area into square sub-lattices using the Euclidean algorithm is, for example, “Two Theorems by Geometrical Dane of Division”, Nihon Crihonsha, Toshikazu Sunada, p.34-p.38. Since it is well-known as described in (1), further detailed explanation is omitted.

With reference to FIG. 7, the specific example of the said procedures 1-5 is shown.
1. The grid array 10 can be expressed as K (7, 5).
2. If a square sub-grid K (5,5) is packed from the left side of K (7,5), only one square is entered, so q ₀ = 1 and the remainder is (7−5 · 1) = 2. .
3. Therefore, rectangular K (2,5) remains.
4). If the square sub-lattice K (2, 2) is packed from the lower side of K (2, 5), two squares can be arranged, so q ₁ = 2, and K (2, 1) is left behind.
5. Finally, two nodes with K (1,1) remain.
As a result, the grid array 10 is divided into five square partial grids surrounded by thick squares in FIG.

  Returning to FIG. 6, the system administrator designates the divided square partial grid as either the file storage node or the index node (S14). In FIG. 7, a node included in the left partial grid is defined as a “file storage node”, and nodes included in the remaining partial grids are defined as “index nodes”. The system administrator assigns a number for identifying each node (S16). The result is shown in FIG. As shown in the figure, there are 25 file storage nodes c0 to c24, four secondary index nodes b0 to b3, and one primary index node a0. A method of using the nodes a0 'and b0' to b3 'in the drawing will be described later.

  In this embodiment, files are distributed and arranged in file storage nodes, and a pointer to the storage location is managed by a known B-tree as index data for knowing the storage location. In this embodiment, the IP address or MAC address shown in FIG. 3 is used as a pointer. Each file storage node has a one-to-one correspondence with a “leaf” in the B-tree, and each index node has a one-to-one correspondence with a “node” in the B-tree. The 1 × 1 node obtained by the last square division is made to correspond to the “root” in the B-tree structure. As a result, the configuration of the B-tree is as shown in FIG. As shown in the figure, the primary index node a0 has four branches connecting itself and the secondary index nodes b0 to b3. The secondary index nodes b0 to b3 have some of the file storage nodes as leaves. The relationship between the file storage node and the secondary index node will be described later.

  The system administrator installs a program for executing a search process described later on the primary index node, the secondary index node, and the file storage node.

  FIG. 10 is a functional block diagram of the primary index node in a state where the search program is executed. The file receiving unit 102 receives a file name for an inquiry about a file storage location, or receives a transferred file. The search unit 104 includes a coding unit 106 and a hash calculation unit 108. The encoding unit 106 encodes (numerizes) the file name by a method described later, and converts the file name into a “storage location code” representing a file storage node that should hold the file. Here, “encoding” refers to converting an arbitrary file name into an integer equal to or less than the number of file storage nodes, as will be described later with a specific example. The hash calculation unit 108 applies a hash function to the storage location code and calculates a hash value. The file transfer unit 110 transfers the file to another node according to the address information. The table holding unit 112 holds the address information of the node corresponding to the storage location code or the hash value in a table format. The information acquisition unit 114 acquires node number and IP address information given by the system administrator.

  The configuration of the secondary index node and the file storage node is the same as that of the primary index node, but the function of the search unit 104 and the index data stored in the table holding unit 112 are different as will be described later.

  Next, with reference to the flowchart of FIG. 11, a process for determining a file storage node to store the file based on the file name and further determining an index node to store the index data of each file storage node will be described. To do. Through this process, a file in the data storage system can be searched using the B-tree as shown in FIG.

Consider the case where a B-tree is constructed according to the procedure of FIG. 6 and then a file is transmitted from the outside to the data storage system and stored in an appropriate node. In the primary index node, the address information of the secondary index nodes b0 to b3 and the file storage nodes c0 to c24 is recorded in advance by the system administrator. The data is recorded in the unit 112 (S20). External files are sent by the router to the primary index node. The file receiving unit 102 of the primary index node receives the file and passes it to the encoding unit 106. The encoding unit 106 encodes the file name of the received file according to a predetermined rule, and calculates the storage location code x (S22). The file transfer unit 110 transfers the file to the file storage node specified by the storage location code x (S24). Next, the hash calculator 108 calculates a hash value by applying a hash function to the storage location code x (S26). When this hash function is represented as h (x) and the number of secondary index nodes is represented as r ₁ ² , h (x) = 0,. . . , (R ₁ ² −1), the hash function is selected. The file transfer unit 110 transmits the address information of the file storage node to the secondary index node specified by the hash value (S28). The secondary index node records the node number of the file storage node in association with the address information in its own table holding unit 112 (S30).

  Hereinafter, each step of FIG. 11 will be described with a specific example. In this example, for the sake of simplicity, it is assumed that the file names are all given in hiragana, and encoding is executed based on the vowels of each character of hiragana.

  The encoding unit 106 of each node holds a vowel code table as shown in FIG. Then, according to the vowel code table, the encoding unit 106 sets “1” if the vowel of each character in the file name is “A”, “2” if the vowel is “Yes”, and sets the vowel “ "3" is given for "U", "4" is given if the vowel is "E", and "5" is given if the vowel is "O". “0” is given for prompting sound, stuttering sound, long sound and “n”.

  FIG. 13 shows a specific example of a file name and a storage location code calculation method based on the file name. When the file name is “Seminasan?”, The encoding unit 106 performs the code “4” of the vowel “e” of “se”, the code “2” of the vowel “i” of “mi”,. . . In this way, the hiragana code of the file name is converted according to the vowel code table. After converting all the characters, add them together. In this example, codes “4”, “2”, “1”, “1”, “0”, and “1” respectively corresponding to “se”, “mi”, “na”, “sa”, “n”, and “ka” are added, The storage location code “9” is obtained. This number indicates the number of the file storage node (that is, c9) where the file is to be stored. The other file names “Keihi”, “Yosan”, “Kashi” and “Taisaku” are calculated in the same procedure, and the respective file storage nodes are c8, c6, c3 and c7. If the storage location code is equal to or greater than “25”, which is the total number of file storage nodes, as a result of the code addition, the remainder when the storage location code is divided by 25 is used as the storage location code of the file.

Further, with reference to FIG. 13, a procedure for determining a node to store the index data of the file storage node will be described. The hash calculator 108 calculates a remainder modulo the number of secondary index nodes for the storage location code as a hash value. Then, the hash value is determined as the number of the secondary index node that should hold the index data. For example, if the storage location code is “9”, since 9 = 4 · 2 + 1, the secondary index node is b1. Therefore, the file storage nodes c0, c4, c8,. . . , C24 index data is stored in the secondary index node b0 and file storage nodes c1, c5, c9,. . . , C21 is stored in the secondary index node b1, and the file storage nodes c2, c6, c10,. . . , C22 index data is stored in the secondary index node b2, and the file storage nodes c3, c7, c11,. . . , C23 is stored in the secondary index node b3.
Note that the hash function may be another function as long as it outputs different hash values for successive storage location codes.

  FIG. 14 shows each node and data stored in them as a B-tree structure constructed according to the above specific example. In the primary index node, index data (that is, a set of node number and address information) of all secondary index nodes and file storage nodes is arranged. On the other hand, the index data of the file storage node corresponding to the leaf of the B-tree is arranged in the secondary index node. The secondary index node does not hold the index data of all the file storage nodes, but holds only the index data of the file storage node storing the file whose hash calculation result matches the node number of itself. In this specification, this is referred to as “holding index data of a subtree”.

  As shown in FIG. 14, in this embodiment, an actual file is held in a file storage node, and index data necessary for searching for the file is stored in a secondary index node and a primary index node. Thus, the file storage location and the index data storage location are set to different nodes.

  In addition, since the files can be managed with the B-tree structure regardless of the number of the entire file storage nodes, the number of arrangements of the nodes in the vertical and horizontal directions is scalable.

The division into sublattice of the above-mentioned square, consists of disjoint, the file storage node and also always disjoint two the number of nodes of primary index node of the ratio r ₀ ² and r ₁ ² and r ₀ and r ₁ . For this reason, applying a hash function to the hierarchy following the index is expected to increase the effect of distributing the index and actual data, compared to the case where there is a common divisor in the number of secondary index nodes and the number of file storage nodes. .

  When a file request comes from the outside to the data storage system, the request is passed to the primary index node. The primary index node encodes the file name, searches the table for the address of the file storage node having the same number as the storage location code, and passes the file request to the searched file storage node. The file storage node that has received the file request retrieves the file from the storage device and transmits the retrieved file to the request source address.

  After the B-tree structure is constructed, the files originally stored in the nodes can be rearranged according to the B-tree. This process will be described with reference to the flowchart of FIG.

  First, in each node, the file name of the data file currently stored is acquired, and each storage location code is calculated (S40). It is determined whether or not the calculated storage location code matches the node number assigned to each node (S42). If they match (Y in S42), the data file having the file name is to be stored in the node, so this flow ends. If they do not match (N in S42), the data file with that file name is to be stored in another node. Therefore, the file storage node inquires of the higher-order index node about the address of the file storage node where the file is to be stored (S44).

Upon receiving the file name, the secondary index node calculates a storage location code and further calculates a hash value (S46). If the calculated hash value matches the node number assigned to the secondary index node (Y in S48), the file storage node address that matches the storage location code exists in the table, so an inquiry has been made. Address information is transmitted to the file storage node (S50). When the calculated hash value does not match the node number assigned to the secondary index node (N in S48), the secondary index node inquires of the primary index node about the file name (S52).
Since the address information of all the nodes is recorded in the primary index node, the address of the file storage node corresponding to the storage location code is searched, and the address information is transmitted to the file storage node that has inquired (S54). ).

  The inquiring file storage node directly sends the file to the file storage node according to the transmitted address information and requests storage (S56). The file storage node that has received the file encodes the file name and confirms that the file is to be stored therein, and then stores the file in the storage device (S58).

  The above procedure is for the file storage node, but is slightly different for the index node. If the secondary index node knows that it is a file storage node in the subtree managed by the coding and hash value calculation, the secondary index node searches the address and requests storage of the file. If it is determined that the data is not in the subtree managed by the user, the primary index node is inquired about the address of the file storage node and then requested to store the file.

  Since the primary index node holds the addresses of all nodes, it searches for the address of the file storage node having the same number as the storage location code, and requests storage of the file. This makes it possible to rearrange the files held in the storage device of each node before the creation of the B-tree.

  The procedure shown in FIG. 15 can also be applied to a case where an application is executed in each node in the data storage system and a file stored in the system is needed and the file is searched. The process until the file name is inquired is the same as in FIG. 15. When the address of the node storing the desired file is found, a file request is issued to that address. When the file storage node receiving the file request encodes the file name and confirms that the file is stored in itself, the file storage node transmits the file to the requesting file storage node.

  The addresses of all nodes in the grid array are at the primary index node. Therefore, if only the address of the primary index node is notified in advance to each file storage node, each file storage node issues a query for the storage location of the necessary file to the primary index node, thereby storing the file. You can know the address of the node. However, with this configuration, all queries are concentrated on the primary index node. On the other hand, in the present embodiment, the address of a part of the query from the file storage node can be known by the secondary index node if the file is in a subtree managed by the secondary index node. The search load on the node can be reduced.

  The address information of all secondary index nodes as well as the addresses of secondary index nodes in the subtree may be transmitted in advance to each file storage node. In this way, when searching for a file from the file storage node, if encoding and hash value calculation are performed and the file does not exist in the storage device of its own node, the subtree including the node storing the file The secondary storage node that manages the file can be directly inquired about the storage location of the file. Therefore, since the secondary index node does not make an inquiry to the primary index node as shown in FIG. 15, the processing load on the primary index node can be reduced.

  Instead of storing the addresses of all nodes in the primary index node, only the addresses of the secondary index nodes may be stored. In this case, when there is an external file request, the primary index node calculates the encoding and hash value, and manages the file storage node storing the file as a subtree. Can know the address. The primary index node forwards the file request to the secondary index node. The secondary index node retrieves the file storage node address from the table by extracting the file name from the file request and obtaining the storage location code. Then, the file request is passed to the corresponding file storage node. By doing so, the search load of the primary index node can be further reduced by passing the process of searching the final file storage node address to the secondary index node.

  As shown in FIG. 8, when nodes arranged in a grid are divided into squares and given the roles of file storage nodes and index nodes, it is inevitable that nodes that do not constitute a B-tree will occur. In FIG. 8, the nodes a0 'and b0' to b3 'do not constitute a B-tree. Therefore, these unused nodes may be used as a backup area for search data.

  FIG. 16 illustrates index node replication. When unused nodes are generated by the Euclidean mutual division method described above, their sizes are always the same as the square sub-lattices constituting the index nodes. Therefore, replication of index data stored in the index node of the same size is executed to create a backup node. In FIG. 16, the node a0 ′ replicates the primary index node a0, and the nodes b0 ′ to b3 ′ replicate the secondary index nodes b0 to b3 (hereinafter, the original B-tree is “primary”). The person who replicates is called "backup"). By duplicating the index data in this way, the reliability of the index can be improved. This corresponds to duplication of a part of the hierarchy of the B-tree.

  In the procedure using the Euclidean mutual division method, two or more 1 × 1 nodes are finally formed. Therefore, replication of the primary index node can be executed as long as the vertical and horizontal directions of the grid array are relatively prime. After securing the backup node, the primary index that is the root is sent by distributing external file requests to multiple primary index nodes (primary and backup) using a known algorithm such as round robin or hash by the router. It is also possible to alleviate the concentration of access to the nodes and reduce the processing load on the primary index node when the number of accesses increases. The same effect can be obtained by distributing the file request from the file storage node to the secondary index node to the primary and backup secondary index nodes.

  Also, when divided into square sub-lattices by the Euclidean mutual division method, the direction in which the primary and the backup are arranged is alternate between the hierarchies. That is, as shown in FIG. 17, if the primary index node primary and backup are arranged in the horizontal direction, the secondary index node primary and backup are arranged in the vertical direction. Therefore, a data flow for synchronizing index data between the primary and the backup (in the direction of the hatched arrow in FIG. 17), and a data flow for exchanging file requests and address information during file search. It is normally orthogonal to (the direction of the white arrow in FIG. 17). Therefore, these flows do not compete for one link, and it is difficult for a situation where a part of the links become a bottleneck and the processing performance of the entire system deteriorates.

  As described above, in the present embodiment, nodes arranged in a lattice and a B-tree structure are applied, and the elements corresponding to “leaves”, “nodes”, and “roots” of the B-tree are applied to the lattice system. All nodes are assigned one-on-one. Then, the actual data is stored in the node corresponding to the leaf as in the B-tree, and the index data for routing to the leaf node is arranged in the node corresponding to the node and the root. A conventionally known B-tree is characterized in that both index data and actual data are closed in one computer, but it is spread over a plurality of nodes. Another feature is that a node area division for creating a B-tree and index data assignment are paired via a hash.

  In addition, by using the Euclidean mutual division method, data can be managed in principle regardless of the number of nodes as long as the number of nodes in the vertical and horizontal directions of the grid array is relatively prime. Therefore, the expandability of the system is high.

  According to the present embodiment, when nodes arranged in a grid are used as a data storage system, files and index nodes are hierarchized and distributedly arranged using a known B-tree structure. The index node serves as a pointer indicating the address of the node where the file is actually stored. Further, the index data can be distributed to a plurality of index nodes by utilizing the fact that the number of index nodes and the number of actual file storage nodes are relatively prime. Also, since the number of actual file storage nodes and the number of secondary index data storage nodes are relatively prime, it is possible to disperse the bias of data arranged in the actual file storage nodes. Since the ratio of the number of nodes between hierarchies is a square of numbers that are relatively prime, the ratio of the number of nodes is relatively prime in each hierarchy, so that the distribution of data can be reduced. In addition, since the file search load at each leaf node is relatively distributed, it is possible to prevent the processing load of a specific node from protruding and increasing.

  In the embodiment, for the sake of simplicity, the encoding is calculated by “vowels of Japanese file names”, and therefore the length of the file name is actually limited, so that the effect of dispersion is small. However, if encoding is performed using, for example, an ASCII code or the like, the value of the storage location code is predicted to be a more dispersed value, so that the effect of this dispersion is expected to be greater. In addition to this, in order to code character strings such as file names, such as simply counting the number of characters in the file name, assigning a number to the file name alphanumeric characters in advance, or applying an encryption method Any technique can be used.

In the present embodiment, the following effects are also obtained by dividing the nodes arranged in a lattice shape into square partial regions. That is, branch nodes having a parent-child relationship and a sibling relationship are located close to each other in the lattice system. Parent and child nodes have a dependency relationship with index data, and sibling nodes have a complementary relationship with index data. Therefore, the access time in the parent-child direction and sibling direction can be shortened by being adjacent to each other.
In addition, since the file storage node is secured as a square area, when transferring a file, multiple paths to that node are made, and the file is transferred even when some of the links between the nodes are disconnected. Can be realized.

  The present invention has been described based on some embodiments. Those skilled in the art will understand that these embodiments are exemplifications, and that there may be various modifications to the combinations of the respective constituent elements and processing processes, and such modifications are also within the scope of the present invention. By the way.

  It should also be understood by those skilled in the art that the functions to be fulfilled by the constituent elements described in the claims are realized by the individual functional blocks shown in the present embodiment or their linkage.

  In the embodiment, it has been described that the index node that stores the index data for searching the address of the file storage node is determined using the hash value obtained from the number of the file storage node. However, the above-described B-tree cannot cope with a search specifying a range. Therefore, index data may be managed using an n-ary search tree. An n-ary search tree can be constructed by forming nodes and leaves using the storage location code that appears when calculating the hash value. The index data uses this storage location code as it is, while the actual data determines the leaf to be stored by the hash value. In the range search, the n-number search tree is traced to obtain actual data stored in the leaves.

  The file name was used to calculate the storage location code, but other file-specific information such as file creation time, update time, file size, file creator, computer name that created the file, and combinations of these The storage location code may be calculated from Two or more types of indexes, that is, storage location codes may be used in combination.

  If the system administrator needs to increase the file storage node or the index node when the data to be stored increases, it is possible to increase the number of vertical and horizontal nodes in the grid array afterwards. In this case, the primary index node may reconstruct the B-tree according to the above-described procedure based on the number of nodes in the new grid array and the address information of each node. As a result of the reconstruction, the ratio between the number of file storage nodes and the number of secondary index nodes changes, so the number of secondary index nodes that should store index data and the number of file storage nodes increases, and the file storage that should store data files When a node changes, each node may perform data rearrangement according to the above-described rearrangement procedure.

  As described in the embodiment, the number of nodes in the grid array is determined so that the system administrator can cut into a desired square sub-grid according to the specifications, and the grid array It is possible to construct a storage system having a preferred logical configuration in which there are no unused nodes and a node for replication is secured. However, square division can be executed by a program without depending on the system administrator. In this case, the number of nodes in the vertical and horizontal directions of the target grid-like array 10 may be arbitrary, and the program executed on any node sequentially executes the steps described in FIG. You may make it cut out to.

  Depending on the shape of the grid, a hierarchy higher than the tertiary index may be provided in the B-tree. However, since the method of the present invention can determine the file storage node until the search of the secondary index node, no higher order hierarchy is required. It is.

  A plurality of servers or personal computers may be arranged under one node 20 in FIG. 1 by connecting a plurality of servers or personal computers to one router.

1 is an overall configuration diagram of a data storage system according to an embodiment of the present invention and a client terminal connected thereto. FIG. It is a hardware block diagram of each node which comprises a lattice-like arrangement | sequence. It is the figure which showed a mode that each node was linked with the node of an up, down, left and right. (A), (b) is a figure explaining the prior art in the data storage system which has a lattice-like arrangement | sequence. (A), (b) is a figure explaining the prior art in the data storage system which has a lattice-like arrangement | sequence. It is a flowchart which shows the procedure which divides | segments a grid | lattice-like arrangement | sequence into several square partial grids. It is a figure which shows the specific example of the division | segmentation of a grid | lattice-like arrangement | sequence. It is a figure which shows the division | segmentation result to the square partial lattice of a grid | lattice-like arrangement | sequence, and the node number allocated to each node. It is a figure which shows the structure of B-tree. It is a functional block diagram of a primary index node that executes a search program. 6 is a flowchart illustrating a process for determining a file storage node and an index node. It is a figure which shows a vowel code table. It is a figure which shows the calculation method of the code based on the specific example of a file name, and a file name. It is a figure which shows the B-tree and the data stored in each node. It is a flowchart which shows the process of reorganizing a file according to B-tree. It is a figure which shows replication of an index node. It is a figure which shows the data flow when synchronizing index data, and the data flow at the time of a file search.

Explanation of symbols

  10 grid array, 12 client terminal, 14 network, 16 router, 20 node, 96 storage device, 98 network interface, 100 data storage system, 102 file receiving unit, 104 search unit, 106 encoding unit, 108 hash calculation unit, 110 file transfer unit, 112 table holding unit, 114 information acquisition unit.

Claims (7)

A data storage system in which files held in a plurality of nodes each having a storage device are managed using a common index to function as a single database,
The plurality of nodes are arranged in a grid pattern, and each node is connected to be communicable with front, rear, left and right nodes, and a file storage node that actually stores a file and an index node that stores index data of the file storage node Each is divided to form a square sub-grid,
The file storage node is made to correspond to the leaf of the tree structure, the index node is made to correspond to the root or node of the tree structure, and the information of the tree structure for managing the file and the index data is held in the index node,
Address information of the file storage node is stored in an index node that is uniquely determined based on file specifying information for specifying a file held in the file storage node ;
The number of nodes arranged in the vertical direction and the number of nodes arranged in the horizontal direction are relatively prime in the lattice-like arrangement, and the lattice-like arrangement is divided into a plurality of square sub-lattices using the Euclidean algorithm. A data storage system.   Of the partial grids obtained by dividing the nodes arranged in the grid, a node included in the largest partial grid is the file storage node, and a node included in another partial grid is the index node. The data storage system of claim 1, wherein: The file specifying information is encoded according to a predetermined rule, and a file storage node to store a file corresponding to the file specifying information is determined according to the obtained code.
3. The data according to claim 1, wherein a hash value is obtained by applying a hash function to the code, and an index node that should hold address information of the file storage node is determined according to the hash value. Storage system. 4. The data storage system according to claim 3 , wherein a name assigned to the file is used as the file specifying information. The case that could portions grid of the same size by dividing the data storage system of claim 1, characterized in that to replicate the index data into a plurality of partial grid to create a backup of the index nodes. In a data storage system in which files held in a plurality of nodes each having a storage device are managed using a common index and function as a single database,
The plurality of nodes are arranged in a grid pattern, and each node is connected to be communicable with front, rear, left and right nodes, and a file storage node that actually stores a file and an index node that stores index data of the file storage node Each is divided to form a square sub-grid,
The file storage node is made to correspond to the leaf of the tree structure, the index node is made to correspond to the root or node of the tree structure, and the information of the tree structure for managing the file and the index data is held in the index node When
A file search program executed in the index node is
The ability to receive file search requests;
A function of encoding file specifying information of the file according to a predetermined rule, and determining a file storage node in which a file corresponding to the file specifying information is stored according to the obtained code;
A function for determining a hash value by applying a hash function to the code, and determining an index node in which address information of the file storage node is held according to the hash value;
A file search program in a data storage system comprising: In a data storage system in which files held in a plurality of nodes each having a storage device are managed using a common index and function as a single database,
The plurality of nodes are arranged in a grid pattern, and each node is connected to be communicable with front, rear, left and right nodes, and a file storage node that actually stores a file and an index node that stores index data of the file storage node Each is divided to form a square sub-grid,
The file storage node is made to correspond to the leaf of the tree structure, the index node is made to correspond to the root or node of the tree structure, and the information of the tree structure for managing the file and the index data is held in the index node When
The index nodes functions as a file search unit,
A file receiver for receiving a file search request;
A coding unit that codes the file identification information of the file according to a predetermined rule, and determines a file storage node in which a file corresponding to the file identification information is stored according to the obtained code;
A hash calculation unit that applies a hash function to the code to obtain a hash value, and determines an index node that holds address information of the file storage node according to the hash value;
A file search apparatus in a data storage system comprising:

Images