KR20030013619A - Data of insert and search method of parallel high-dimensional index structure - Google Patents

Abstract

PURPOSE: A method for inserting and searching data of a parallel higher index structure is provided to search a higher index effectively by transforming a higher index structure which uses a parallel property of a SAN(Storage Area Network) thereby increasing a fan out, reducing a height of a tree, and maximizing a parallel property of an input/output at searching a range in searching a similarity. CONSTITUTION: For performing a partial K-most access query in a main server and all sub servers simultaneously, if a K-access query is entered, all servers access to a root node(1300). It is judged whether the accessed root node is a non-terminal node or not(1301). If the accessed root node is a non-terminal node, all servers calculates a similarity with a query and an entry, and each entry is sorted in a list in order of similarity(1302). All servers access to a child node of the first entry stored in the list in parallel and allocate the current node as a root node(1303), and the current stage is returned to the above stage (1301) for judging a non-terminal node of not of the root node.

Description

Data of insert and search method of parallel high-dimensional index structure}

The present invention relates to a data insertion and retrieval method of a parallel high dimensional index structure, and more particularly, to insert and retrieve data of a parallel high dimensional index structure in order to efficiently store large amounts of high dimensional data in a storage area network (SAN) environment. And a computer-readable recording medium having recorded thereon a program for realizing the method.

As data becomes high-dimensional in a data transmission network environment, a method of inserting and searching data having various high-level index structures has been proposed to accommodate this.

However, the recently proposed high-dimensional index structure has a problem that the search performance decreases significantly as the fan-out is reduced as the dimension is increased. In particular, the parallel multi-dimensional index structure using the parallel environment has images, video, It was difficult to accommodate large amounts of high-dimensional data such as CAD data. In addition, since the K-nearest query, which is an important query form in similarity search, is not considered, there is a need for an appropriate index structure that can efficiently search a high-dimensional index.

Accordingly, the present invention has been proposed to meet the above requirements, and by modifying a high-dimensional index structure using parallelism such as a SAN structure to increase the fan-out and reduce the height of the tree, even when searching the range in similarity search With a computer that records and inserts data in a parallel high dimensional index structure in a data transmission network (SAN) environment to maximize the parallelism of the output so that the high dimensional index can be efficiently searched, and a program for realizing the method. The purpose is to provide a readable recording medium.

1 is a diagram for explaining the concept of a data transmission network (SAN) to which the present invention is applied.

2 is a configuration diagram of an embodiment of a data transmission network system to which the present invention is applied.

FIG. 3 is a detailed configuration diagram of the tree structure in each disk group shown in FIG.

4 is a diagram illustrating an insertion process of a parallel high-dimensional index structure in the system shown in FIG.

5 and 6 are diagrams illustrating one embodiment of a node partitioning process in a disk during the insertion process shown in FIG.

7 is a flow diagram of an embodiment of a data insertion method of a parallel high-dimensional index structure in a data transmission network (SAN) environment in accordance with the present invention.

8 is a diagram illustrating a range query search process in a disk group during data search of a parallel high-dimensional index structure in a data transmission network (SAN) environment according to the present invention.

FIG. 9 is another embodiment illustrating a range query search process in a disk group during data search of a parallel high-dimensional index structure in a data transmission network (SAN) environment according to the present invention. FIG.

10 is a flowchart of a first embodiment of a data searching method of a parallel high-dimensional index structure in a data transmission network (SAN) environment according to the present invention.

11 is a flowchart of a second embodiment of a data searching method of a parallel high-dimensional index structure in a data transmission network (SAN) environment according to the present invention.

12 is a flowchart of a third embodiment of a data searching method of a parallel high-dimensional index structure in a data transmission network (SAN) environment according to the present invention.

13 is a flowchart of a fourth embodiment of a data searching method of a parallel high-dimensional index structure in a data transmission network (SAN) environment according to the present invention.

<Description of the symbols for the main parts of the drawings>

200: primary server 201: part 1 server

202: Part 2 server 203: Storage Area Network (SAN)

204: first disk storage group 205: second disk storage group

206: third disk storage group

In order to achieve the above object, the present invention provides a data entry method of a parallel high dimensional index structure applied to a parallel high dimensional index system in a data transmission network (SAN) environment. A first step of allocating a new entry to the at least one secondary server determined by itself and including itself; A second step of the at least one secondary server and the primary server selecting a node through tree traversal for insertion of a received new entry, and determining whether there is free space for inserting an entry in the selected node; A third step of inserting a new entry into the page free space of the node to be inserted by the main server and the at least one sub server when there is free space in the node as a result of the determination of the second step; And as a result of the determination of the second step, when there is no free space in the node, the primary server and the at least one secondary server allocate a page to the disk where the pages of the level where the current partition occurs are distributed, and insert a new entry in the node. And a fourth step of reflecting the change of the minimum bounding region in the parent node.

In addition, the present invention is a data searching method of a parallel high dimensional index structure applied to a parallel high dimensional index system in a data transmission network (SAN) environment, the primary server and at least one secondary server input / output scheduler A first step of determining whether there is information managed by the delivered I / O scheduler;

As a result of the determination in the first step, when there is no information managed by the delivered I / O scheduler, at least one secondary server delivers the range query result to the primary server, and the primary server delivers the at least one secondary server. A second step of collecting the received range query result and the range query result found by the user and delivering the collected query result to the client; And as a result of the determination in the first step, when there is information managed by an input / output scheduler, the primary server and at least one secondary server access several nodes in parallel from the input / output scheduler so that the current node and the terminal node are the same. And a third step of performing a range query search according to whether or not there is a request.

On the other hand, the present invention is a data search method of a parallel high-dimensional index structure applied to a parallel high-dimensional index system in a data transmission network (SAN) environment, the main server and at least one secondary server to access the root node, the root accessed Determining whether the node is a non-terminal node; As a result of the determination in the first step, when the accessed root node is a non-terminal node, the primary server and at least one secondary server calculate similarities between the query and the entries, and sort the calculated similarities in the order of similarity in the list. And a second step of accessing the child nodes of the first entry stored in the list in parallel and allocating the current node as the root node; As a result of the determination of the first step, when the approaching root node is not a non-terminal node, the primary server and at least one secondary server calculate the similarity between the query and the object so that only a value less than or equal to the K-nearest distance from the node is calculated. Extracting and storing the result set in a result set; A fourth step of determining whether the number of results stored in the result set is greater than or equal to K (k is a natural number) and the K th result is less than or equal to the similarity value of the first entry of the list; As a result of the determination in the fourth step, if the number of stored results is greater than or equal to K and the Kth result is less than or equal to the similarity value of the first entry in the list, the at least one secondary server returns K primary results. A fifth step in which the primary server collects the results K and the results found by the primary server from the at least one secondary server, sorts them in order of similarity, and selects K final results and returns them to the client; And if the number of stored results is not greater than or equal to K or the Kth result is not less than or equal to the similarity value of the first entry in the list, the primary server and the at least one secondary server are determined. And a sixth step of allocating the current node as the root node by accessing the child nodes of the first entry stored in the list in parallel.

In addition, the present invention, in a data search method of a parallel high-dimensional index structure applied to a parallel high-dimensional index system in a data transmission network (SAN) environment, the main server inputs a K-access query to the root node, Determining whether the accessed root node is a non-terminal node; As a result of the determination of the first step, if the accessed root node is a non-terminal node, the main server calculates the similarity between the query and the entries, sorts each similarity in the order of similarity in the list, and stores the first entry in the list. Accessing the child nodes of the second node in parallel and allocating the current node as a root node; As a result of the determination in the first step, when the accessed root node is not a non-terminal node, the main server calculates the similarity between the query and the object, extracts only a value less than or equal to the K-nearest distance from the node, and stores the result in the result set. Determining whether the number of stored results is greater than or equal to K; As a result of the determination in the third step, when the number of results stored in the result set is greater than or equal to K, the main server calculates the similarity between the K th result object and the query and converts it into a range query form. Transmitting a range query to at least one secondary server; At least one sub-server receiving the converted range performs a range query to deliver a search result of maximizing parallelism to the main server, and collects the search results received from the main server and the obtained results and sorts them in similarity order. Obtaining a final K results and returning the final K results to the client; And when the number of results stored in the result set is not greater than or equal to K as a result of the determination in the third step, the main server accesses the child nodes of the first entry stored in the list in parallel and makes the current node the root node. And a sixth step of allocating.

In addition, the present invention provides a data search method of a parallel high dimensional index structure applied to a parallel high dimensional index system in a data transmission network (SAN) environment, wherein a main server and at least one department access a root note for a K-access query. A first step of determining whether the root node is a non-terminal node; As a result of the determination of the first step, when the root note is a non-terminal node, the primary server and at least one secondary server calculate similarities between the query and the entries, sort each similarity in the order of similarity in the list, and store the same in the list. A second step of accessing the child nodes of the first entry in parallel and allocating the current node as the root node; If the root node is not the non-terminal node as a result of the first step, the primary server and at least one secondary server calculate the similarity between the query and the object, and extract only the value less than or equal to the K-nearest distance from the node. Storing in a result set and determining whether the number of stored results is greater than or equal to K; If the number of results is not greater than or equal to K as a result of the determination in step 3, the primary server and at least one secondary server access the child nodes of the first entry stored in the list in parallel to access the current node as the root node. Assigning a fourth step; As a result of the determination in the third step, when the result value is greater than or equal to K, the primary server and at least one secondary server calculate the similarity between the K th result object and the query and convert the similarity to the range query form. A fifth step of sending, by the secondary server, the converted range query to the primary server; The main server receiving the converted range query selects a range query having the smallest range among the ranged queries and the range query converted by the main server, and transmits the selected range query to at least one sub-server, and at least one received the range query. A sixth step of the secondary server performing a range query and delivering a result of the range query to the primary server; And a seventh step in which the main server, which has received the result of performing the range query from the at least one secondary server, collects the results found by the main server, selects K results, and returns the selected K results to the client. do.

On the other hand, the present invention, in order to insert a query of a parallel high-dimensional index structure, in a parallel high-dimensional index system having a processor, at least one determined by including a new entry that is input by the main server to each disk group, including itself A first function of forwarding a new entry to the secondary server; The at least one secondary server and the primary server select a node through tree traversal for insertion of a new entry, and determine whether there is free space for inserting an entry in the selected node; A third function of inserting a new entry into a page free space of a node to be inserted by the main server and the at least one sub server when there is free space in the node as a result of the determination of the second function; And as a result of the determination of the second function, when there is no free space in the node, the primary server and the at least one secondary server allocate a page to the disk where the pages of the level where the current partition occurs are distributed, and insert a new entry in the node. A computer readable recording medium having recorded thereon a program for realizing a fourth function of reflecting a change in a minimum bounding region on a parent node is provided.

In addition, the present invention, in order to search the data of the parallel high-dimensional index structure, to a parallel high-dimensional index system having a processor, the main server and at least one secondary server delivers the root node information to the input / output scheduler, A first function of determining whether there is information managed by an input / output scheduler; As a result of the determination in the first function, if there is no information managed by the transmitted I / O scheduler, at least one secondary server transmits the range query result to the primary server, and the primary server delivers the at least one secondary server. A second function of collecting the received range query result and the range query result found by the user and delivering the collected query result to the client; And when there is information managed by an input / output scheduler as determined by the first function, the primary server and at least one secondary server access several nodes in parallel from the input / output scheduler so that the current node and the terminal node are the same. A computer readable recording medium having recorded thereon a program for realizing a third function of performing a range query search depending on whether or not the present invention is provided.

In addition, the present invention, in order to search the data of the parallel high-dimensional index structure, in the parallel high-dimensional index system having a processor, the main server and at least one secondary server access the root node, the access root node is a non-terminal A first function of determining whether a node is a node; As a result of the determination of the first function, when the accessed root node is a non-terminal node, the primary server and at least one secondary server calculate the similarity between the query and the entry, and sort each calculated similarity in the order of similarity in the list. And a second function of accessing the child nodes of the first entry stored in the list in parallel and allocating the current node as the root node; As a result of the determination of the first function, when the approaching root node is not a non-terminal node, the primary server and at least one secondary server calculate the similarity between the query and the object so that only a value less than or equal to the K-nearest distance from the node is calculated. A third function of extracting and storing the result set; A fourth function of determining whether the number of results stored in the result set is greater than or equal to K (k is a natural number) and the K th result is less than or equal to the similarity value of the first entry of the list; As a result of the determination of the fourth function, when the number of stored results is greater than or equal to K and the Kth result is less than or equal to the similarity value of the first entry in the list, at least one secondary server returns the K results in the primary server. A fifth function of collecting the result K and the result found by the primary server from the at least one secondary server, sorting them in order of similarity, and selecting and returning K final results to the client; And the primary server and the at least one secondary server if the number of stored results is not greater than or equal to K or the Kth result is less than or equal to the similarity value of the first entry of the list. A computer-readable recording medium having recorded thereon a program for realizing a sixth function of allocating a child node of a first entry stored in a list in parallel and allocating a current node as a root node is provided.

In addition, the present invention, in order to search the data of the parallel high-dimensional index structure, in the parallel high-dimensional index system having a processor, the K-access query inputted by the main server to the root node, the accessed root node is A first function of determining whether the terminal is a non-terminal node; As a result of the determination of the first function, if the accessed root node is a non-terminal node, the main server calculates the similarity between the query and the entry, sorts each similarity in the order of similarity in the list, and stores the first entry in the list. A second function of allocating a current node as a root node by accessing the child nodes in parallel; As a result of the determination in the first function, when the accessed root node is not a non-terminal node, the main server calculates the similarity between the query and the object, extracts only a value less than or equal to the K-nearest distance from the node, and stores the result in the result set. And determining whether the number of stored results is greater than or equal to K numbers; As a result of the third function, if the number of results stored in the result set is greater than or equal to K, the main server calculates the similarity between the K th result object and the query and converts it into a range query form. A fourth function of forwarding a range query to at least one secondary server; At least one sub-server receiving the converted range performs a range query to deliver a search result that maximizes parallelism to the main server, and collects the search results received from the main server and the obtained results and sorts them in similarity order. A fifth function of obtaining the final K results and returning the final K results to the client; And when the number of results stored in the result set is not greater than or equal to K as a result of the third function, the main server accesses the child nodes of the first entry stored in the list in parallel and makes the current node the root node. A computer readable recording medium having recorded thereon a program for realizing the sixth function to be allocated is provided.

In addition, the present invention provides a parallel high-dimensional indexing system having a processor for searching data of a parallel high-dimensional index structure, wherein a main server and at least one department access a root note for a K-access query so that a root node is a non-terminal. A first function of determining whether a node is a node; As a result of the determination of the first function, when the root note is a non-terminal node, the primary server and at least one secondary server calculate similarities between the query and the entry, sort each similarity in the list in the order of similarity, and store the list in the list. A second function of allocating the current node as the root node by accessing the child nodes of the first entry in parallel; As a result of the determination of the first function, if the root node is not a non-terminal node, the primary server and at least one secondary server calculate the similarity between the query and the object and extract only the value less than or equal to the K-nearest distance from the node. A third function of storing in a result set and determining whether the number of stored results is greater than or equal to K; As a result of the determination of the above three functions, if the number of results is not greater than or equal to K, the primary server and at least one secondary server access the child nodes of the first entry stored in the list in parallel, and make the current node the root node. A fourth function of allocating; As a result of the determination of the third function, when the result value is greater than or equal to K, the main server and the at least one sub server calculate the similarity between the K th result object and the query and convert the result into a range query form. A fifth function for the secondary server to forward the converted range query to the primary server; The main server receiving the converted range query selects a range query having the smallest range among the ranged queries and the range query converted by the main server, and transmits the selected range query to at least one sub-server, and at least one received the range query. A sixth function of the secondary server performing a range query and delivering a result of the range query to the primary server; And a program for realizing a seventh function in which the main server, which has received the result of executing the range query from at least one secondary server, collects the results found by the main server, selects K results, and returns the selected K results to the client. Provide a computer-readable recording medium for recording.

The objects, features and advantages described above will become more apparent from the following detailed description taken in conjunction with the accompanying drawings. Hereinafter, exemplary embodiments of the present invention will be described in detail with reference to the accompanying drawings.

First, a storage area network (SAN) environment to which the present invention is applied will now be described before explaining data insertion and search of a parallel high-dimensional index structure according to the present invention.

SAN is a technology that enables high-speed transmission, long-range connectivity, and multi-protocol capabilities that are advantages of Fiber Channel (FC) over a network, just as a host computer can quickly exchange data with storage devices over a small computer system interface (SCSI). The Storage Networking Industry Association (SNIA), the organization for standardizing SANs, defines "high-speed networks that can transfer large amounts of data between separate connected storage devices regardless of the type of host computer."

In other words, the background of the emergence of SANs is that the data to be managed is growing exponentially, while the devices for storing them are dependent on different servers, which are local area network (LAN) or wide area network (WAN). It is connecting with) to limit the sending and receiving of large amount of data at high speed.

Therefore, SAN has emerged as one of the ways to solve this limitation, which can improve scalability, increase data transmission speed and transmission distance, reduce the cost of sharing data and equipment, and backup server-independent data. There is this.

1 is a view for explaining the concept of a data transmission network (SAN) to which the present invention is applied.

As shown in FIG. 1, a SAN structure is a high-speed network that allows a large amount of data to be exchanged between remotely distributed storage devices 104 and 105 regardless of a host server 100, 101, or 102 to which a large amount of data is connected. The 104 and 105 may be shared by a plurality of servers out of a master relationship with the host servers 100, 101, and 102, and may be used for the joint use of disk or tape equipment, clustering of replication functions, and high availability.

2 is a diagram illustrating an embodiment of a system for inserting and searching a query having a parallel high-dimensional structure in a data transmission network environment to which the present invention is applied.

As shown in FIG. 2, the system includes a primary server 200, a first department 201, a second department 202, a SAN 203, a first disk storage group 204, and a second disk. Storage group 205 and third disk storage group 206.

The main server 200 performs calculation and search related to the index configuration, and performs integrated management of the search results received from the first sub-server 201 and the second sub-server 202. The first sub-server 201 and the second sub-server 202 perform an operation and search on the index configuration, and transmit the search result to the main server 200.

Here, the index structure to which the present invention is applied is an nP-n × mD form in which nP-nD and 1P-nD structures are mixed (where n and m are one or more natural numbers, P is a server and D is a storage device). In general, each server uses parallelism of inputs and outputs.

In addition, such an index structure assumes that the number of disks participating in image retrieval in a SAN environment is much larger than the number of servers. One server manages one disk storage group, and each disk group is indexed separately. The tree exists.

FIG. 3 is a diagram of an embodiment of a tree structure in each disk group shown in FIG. 2, and the nodes shown in FIG. 2 logically represent terminal nodes 310, 314, 318 and non-terminal nodes 303 in an index structure. Include. The page 304 is an input / output unit of a physical disk.

The terminal nodes 310, 314 and 318 are divided and stored in the other disks 300, 301 and 302 in the group, and one entry of each node has a minimum bounding region (hereinafter abbreviated as "MBR") of the lower node to which the node points. Contains pointers to the pages (distributed on each disk) that make up the.

That is, as shown in FIG. 3, the first entry 304 of the root node 303 points to node 1 310, 314, 318. Node 1 (310,314,318) is composed of the first page 310 of disk A 300, the first page 314 of disk B 301 and the first page 318 of disk C (302) In the first entry of the root node 303, an MBR including a pointer to the pages and an object stored in the pages is stored.

That is, the index structure as described above can obtain a declustering effect because one node is divided and stored in pages of several disks. In addition, since the size of one node of the index tree (number of disks in the group x page size) is increased, the height of the index tree is lowered. And although the size of the node is large, this is not a problem because it can be read in parallel at one input / output time. Therefore, the index structure of the data partitioning technique increases the height of the tree as the fan-out of the node decreases as the high-dimensional data becomes higher, which causes a chain effect in which the overlap region also increases.

However, in the above structure, the number of fan-outs can be adjusted according to the number of disks, so that the height of the tree can be kept low. As the dimension increases, the number of nodes that need to be accessed when the search is performed increases, so that pages that participate in the query are higher in dimension. The number of will be large.

Therefore, it is more effective to utilize the uniformly active disk (unispread) property rather than the minimum active disk (Minimum Load) property which is an important property in the parallel high-dimensional index structure. Therefore, to read a node, it is necessary to read the disk on which the page exists, so that the uniformly active disk (unispread) property can be maximized, which is effective in processing range queries and K-nearest queries.

Hereinafter, the process of inserting an entry into the proposed index structure and the process of dividing and assigning when the free space is insufficient after the insertion will be described in detail.

FIG. 4 is a diagram for explaining a data insertion process of a parallel high-dimensional index structure in the system shown in FIG.

As shown in Fig. 4, when trying to insert each entry (a, b, c, d, e, f, g, h) 400 into a disk group in turn, the main server 401 inserts a entry. In the first disk group 404, the entry b is assigned to the second disk group 405 and the entry c to the third disk group 406. Here, it is preferable to use a round robin method for allocating an entry to be inserted in the main server 401. The reason for this is that other declustering methods do not show a great effect at higher levels, and there are many differences when compared to the round robin method. Therefore, the round robin method is less expensive and easy to implement.

5 and 6 are diagrams illustrating an embodiment of a node partitioning process in a disk during the entry insertion process illustrated in FIG. 4. When the free space is insufficient after insertion, nodes are divided and allocated.

That is, as shown in FIG. 5, if overflow occurs due to an attempt to insert a new entry when node 2 505 of each disk is full, the node may assume that disk D 503 and disk E 504 impersonate it. It allocates from disk D 503 and disk E 504 because it maintains a small number of pages. Thereafter, as shown in FIG. 6, node 2 505 divides into node 2 608, 610 and node 3 (605, 606, 607, 609). Here, the allocation process first allocates node 2 608, 610 to disk D 603 and disk E 604, and node 3 605, 606, 607, 609 in turn from disk A 600.

By allocating the partitioned nodes to different disks, multiple nodes can be read in parallel during range queries, the most basic form of queries in high-dimensional index structures.

7 is a flow diagram of an embodiment of a method for inserting a parallel high dimensional index structure on a data transmission network (SAN) in accordance with the present invention.

As shown in FIG. 7, first, according to the declustering method, the main server 401 determines a server to insert and forwards each new entry to the server to be inserted 700.

Then, the main server 401 and each sub-server 402, 403 selects the most suitable node through tree traversal for insertion of the new entry received (701), and selects an entry when the appropriate node is selected through this tree traversal. It is determined whether there is free space in the node to be inserted (702).

As a result of the determination in step 702, if there is free space in the node to be inserted, the main server 401 and each of the sub-servers 402 and 403 determine whether there is free space in the page of the node to be inserted (705).

As a result of the determination in step 705, if there is free space in the node but there is no free space in the node, the main server 401 and each sub-server 402, 403 allocates a page from the disk where no page is allocated to the current node. If the page has free space, a new entry is inserted and the main loop ends after reflecting the MBR change in the parent node (707).

On the other hand, as a result of the determination in step 702, if there is no free space in the node to be inserted, the main server 401 and each sub-server 402, 403 call a node splitting function to perform node splitting and allocate a new node. (703). Then, a new entry is inserted into the allocated node and the MBR change is reflected in the parent node (704). In other words, a new entry is inserted into a node by first assigning a page to a disk in which pages with the lowest level of partitioning are distributed the least, and ending this loop by reflecting the MBR change in the parent node.

FIG. 8 is a diagram illustrating a range query search process in a disk group during data search of a parallel high-dimensional index structure in a data transmission network (SAN) environment according to the present invention.

That is, FIG. 8 illustrates that the main server 401 and each of the sub-servers 402 and 403 approach and search one by one in order to access several nodes during query processing. An entry included in the range query in the root node is 2, 4, The case at 7 will be explained. This range discovery method occurs equally in all disk groups configured in a data transmission network (SAN) environment. In FIG. 8, assuming that the disk group is disk group 1 404 and a server managing disk group 1 404 is assumed to be a main server 401.

First, the main server 401 managing the disk group 1 404 accesses the root node and reads from the inserted entry 1 805 to the entry 7 811, and the entry 2 806 entry 4 (corresponding to the range). 808 and entry 7 811 are selected. When accessing the child nodes 2 812, 816, 818, the child nodes 4 813, 819, and the child nodes 7 814, 815, 817 respectively indicated by the selected entry 2 806, entry 4 808, and entry 7 811, one node. Because we are accessing, first we access the child node 2 (812, 816, 818).

Here, since the child nodes 2 812, 816, 818 are distributed and stored in the disk A 800, the disk D 803, and the disk E 804, the child nodes 2 812, 816, and 818 may access the disk A 800, the disk D 803. And disk E 804 are accessed simultaneously.

Next, the child nodes 4 813 and 819 are distributed and stored in the disk A 800 and the disk E 804 so that when the child nodes 4 813 and 819 are accessed, the child nodes 4 813 and 819 are simultaneously stored in the disk A 800 and the disk E 804. Approach The child nodes 7 (814, 815, 817) are distributed and stored in the disk B 801, the disk C 802, and the disk D 803, and the disk B 801 and the disk C 802 are accessed when the child node 7 is accessed. And disk D 803 are accessed simultaneously.

Therefore, data is accessed and read in the order of child nodes 2 (812, 816, 818), child nodes 4 (813, 819), and child nodes 7 (814, 815, 817), where the total number of disk accesses is the number of first access to the root node and the terminal node disk. Can be calculated as the sum of the number of accesses. Here, since the number of accesses of the root node R is 1 and the number of disk accesses for the terminal nodes 2, 4, and 7 is 3, the total number of disk accesses becomes 4.

In this case, the number of disk accesses can be reduced by combining the positions of the pages of the terminal nodes 2, 4, and 7 nodes.

9 is a diagram illustrating another example of a range query search process in a disk group during data search of a parallel high-dimensional index structure in a data transmission network (SAN) environment according to the present invention.

That is, FIG. 9 is a method of searching and accessing pages of several nodes at the same time when processing a range query. When the entries included in the range query are 2, 4, and 7 at the root node, the range search method is a data transmission network ( The same happens for all disk groups configured in a SAN) environment.

Here, it is assumed that the disk group is assumed to be disk group 1 404, and the server managing disk group 1 404 is assumed to be the main server 401.

First, the main server 401 managing the disk group 1 404 accesses the root node and reads from the inserted entry 1 905 to the entry 7 911, and the entry 2 906 entry 4 (corresponding to the range). 908 and entry 7 911, all the pointer information for the page pointed to by each entry is collected to make a disk I / O plan. Here, the pointers A3, D1, and E1 of the entry 2 mean the third page 912 of the disk A 900, the first page 916 of the disk D 903, and the disk E ( 904, the first page 918.

Here, the disk I / O plan means that pages that can be read at the same time in order to reduce the number of disk accesses are input / output at one time regardless of the node to which they belong. Input / output at once. By properly planning the pages of the nodes to be input and output, you will be able to input and output as many pages as possible at once. That is, A3 912, B3 914, C3 915, D4 917, and E1 918 are all pages present on different disks. Thus, they can be processed with one input / output. Next, A5 913, D4 917, and E2 919 are simultaneously input / output. Therefore, the terminal node access time is 2, and the number of access including the root node is 3, thereby reducing the number of accesses than the first method.

10 is a flowchart of a first embodiment of a data searching method of a parallel high-dimensional index structure in a data transmission network (SAN) environment according to the present invention.

As illustrated in FIG. 10, all servers 401, 402, and 403 transmit root node information to the I / O scheduler 1000, and determine whether there is information managed by the transferred I / O scheduler 1001.

As a result of the determination in the process 1001, when there is information managed by the input / output scheduler, all servers 401, 402, and 403 access several nodes in parallel from the input / output scheduler (1003), and the current node and the terminal node are the same. Determine (1004).

As a result of the determination in step 1004, if the current node and the terminal node are not the same, all servers 401, 402, 403 select an entry included in the range, and transmit information on the entry to the input / output scheduler (1006). In step 1001, it is determined whether there is information managed by the transferred I / O scheduler, and the loop is repeated.

On the other hand, if it is determined in step 1004 that the current node and the terminal node are the same, all the servers (401, 402, 403) calculates the similarity of the object and query included in the node, and stores the result in the result set (1005), The process returns to the process of determining whether there is information managed by the transferred I / O scheduler (1001) and repeats the loop.

On the other hand, as a result of the determination in the process 1001, if there is no information managed by the input / output scheduler, the result of the range query in each sub-server 402, 403 is transmitted to the main server 401, to the main server 401 It combines the range query result found by each sub-server 402 and 403 with the range query result found by the main server 401, returns the final result to the client, and terminates this loop (1002).

11 is a flowchart of a second embodiment of a data searching method of a parallel high-dimensional index structure in a data transmission network (SAN) environment according to the present invention.

That is, FIG. 11 is a search method according to the execution of an independent K-nearest query in all servers 401, 402 and 403. First, all servers 401, 402 and 403 access the root node (1100), and the accessed root node is non-discovered. It is determined whether the terminal node (1101).

As a result of the determination in step 1101, when the accessed root node is a non-terminal node, all servers 401, 402, and 403 calculate similarities between the query and the entry, and sort each similarity in the order of similarity in the list (1102). Then, all the servers 401, 402, 403 access the child nodes of the first entry stored in this list in parallel, assign the current node as the root node (1103), and determine whether the root node is a non-terminal node (1101). It returns and repeats the loop.

On the other hand, if the root node is not a non-terminal node as a result of the process 1101, all the servers (401, 402, 403) calculates the similarity between the query and the object to extract only the value less than or equal to the K- nearest distance from the node Stored in the result set (1404), all servers 401, 402, 403 determine if the number of stored results is greater than or equal to K and the K th result is less than or equal to the derived value of the first entry in the list (1105).

If the number of stored results is greater than or equal to K and the K th result is less than or equal to the similarity value of the first entry in the list, the sub-servers 402 and 403 select K results. Forward to primary server 401 (1106). Then, the main server 401 collects the result K delivered from each sub-server 402 and 403 and the result obtained by the main server 401, sorts them in the order of similarity, and selects K results and returns them to the client (1107). ).

On the other hand, if it is determined in step 1105 that the number of stored results is not greater than or equal to K or the Kth result is not less than or equal to the similarity value of the first entry in the list, the process proceeds to step 1103. All servers 401, 402, 403 access the child nodes of the first entry stored in the list in parallel, assign the current node as the root node, and return to the process of determining whether the root node is a non-terminal node 1101 to repeat the loop. do.

12 is a flowchart of a third embodiment of a data searching method of a parallel high-dimensional index structure in a data transmission network (SAN) environment according to the present invention.

That is, FIG. 12 is a mixture of a range query and a K-nearest query, and it is possible to maximize parallelism at the input and output than the search method shown in FIG.

First, when a K-access query comes in, the main server 401 approaches 1200 the root node, and determines whether the accessed root node is a non-terminal node (1201).

In this case, when it is determined in step 1201 that the accessed root node is a non-terminal node, the main server 401 calculates the similarity between the query and the entry, and sorts each similarity in the order of similarity in the list (1202). ). Then, the main server 401 accesses the child node of the first entry stored in this list in parallel, assigns the current node to the root node (1203), and determines whether the root node is a non-terminal node (1201). It returns and repeats the loop.

On the other hand, if the root node is not a non-terminal node, as determined in the process 1201, the main server 401 calculates the similarity between the query and the object and extracts only a value less than or equal to the K-nearest distance from the node. It is stored in the result set (1204), and it is determined whether the number of stored results is greater than or equal to K (1205).

As a result of the determination in step 1205, if the number of results is greater than or equal to K, the main server 401 calculates the similarity between the K th result object and the query and converts the result into a range query form. The range query is forwarded to all secondary servers 402 and 403 (1206). That is, the main server 401 performs a partial K-nearest search by obtaining the distance between the object corresponding to the K th and the query, and the sub-servers 402 and 403 and the main server 401 that receive the converted range The range query of the proposed method (FIG. 10) is performed to maximize parallelism, and the findings of each sub server 402 and 403 are transmitted to the main server 401 (1207).

Then, the main server 401, which has received the search results from the sub-servers 402 and 403, collects the search results and the results obtained in the main server 401, sorts them in the order of similarity, and obtains the final K results. Results are returned to the client and the loop ends (1208).

On the other hand, as a result of the determination in step 1205, if the number of results is not greater than or equal to K, the main server 401 proceeds to step 1203 to parallelly execute the child nodes of the first entry stored in the list. The method returns to the process of determining whether the root node is a non-terminal node and assigns the current node to the root node, and repeats the loop.

13 is a flowchart of a fourth embodiment of a data searching method of a parallel high-dimensional index structure in a data transmission network (SAN) environment according to the present invention.

That is, FIG. 13 simultaneously executes a partial K-nearest query on the primary server 401 and all secondary servers 402, 403. When a K-access query comes in, all servers 401, 402, 403 access the root note ( 1300, It is determined whether the accessed root node is a non-terminal node (1301).

As a result of the determination in step 1301, when the root note is a non-terminal node, all the servers 401, 402, and 403 calculate similarities between the query and the entry, and sort each similarity in the order of similarity in the list (1302). Then, all the servers 401, 402, 403 access the child nodes of the first entry stored in this list in parallel to allocate the current node as the root node (1303), and determine whether the root node is a non-terminal node (1301). It returns and repeats the loop.

On the other hand, if the root node is not a non-terminal node as a result of the process 1301, all the servers (401, 402, 403) calculates the similarity between the query and the object to extract only the value less than or equal to the K- nearest distance from the node It is stored in the result set (1304), and it is determined whether the number of stored results is greater than or equal to K (1305).

As a result of the determination in step 1305, if the result value is greater than or equal to K, all servers 401, 402, 403 calculate the similarity between the K th result object and the query and convert it to a range query form. ) Transmits the converted range query to the main server 401 (1306). Then, the main server 401 selects the smallest of the range queries from the sub servers 402 and 403 and the range queries of the main server 401 and transmits the smallest one to all the sub servers 402 and 403 (1307). Then, all secondary servers 402 and 403 execute the range query with the received range query, and each secondary server 402 and 403 transmits the result to the primary server 401 (1308). At this time, the main server 401 selects K results by combining the results received from all the sub-servers 402 and 403 and the results found by the main server 401, and the main server 401 selects the K results. Return to the client and end this loop (1309).

The method of the present invention as described above may be implemented as a program and stored in a computer-readable recording medium (CD-ROM, RAM, ROM, floppy disk, hard disk, magneto-optical disk, etc.).

The present invention described above is not limited to the stated embodiments and the accompanying drawings, and various substitutions, modifications, and changes are possible in the art without departing from the technical spirit of the present invention. It will be apparent to those of ordinary skill.

The present invention as described above, by dividing the disk group by the number of servers, each server manages the disk group to maximize the parallelism of the input and output and processor at the same time, change the structure of the node to increase the fan-out has a high-dimensional feature The data can be indexed efficiently, and the range query can maximize the parallelism of I / O and apply it to the K-nearest query to retrieve the data faster.

Claims (13)

In the data insertion method of a parallel high dimensional index structure applied to a parallel high dimensional index system, A first step in which the primary server allocates a new entry to each disk group and forwards the new entry to at least one secondary server determined including itself; A second step of the at least one secondary server and the primary server selecting a node through tree traversal for insertion of a received new entry, and determining whether there is free space for inserting an entry in the selected node; A third step of inserting a new entry into the page free space of the node to be inserted by the main server and the at least one sub server when there is free space in the node as a result of the determination of the second step; And As a result of the determination of the second step, if there is no free space in the node, the primary server and at least one secondary server allocate a page to the disk where the pages of the level where the current partition occurs are distributed and insert a new entry in the node. Fourth step to reflect the change in the minimum bounding region on the node Data insertion method of a parallel high-dimensional index structure comprising a. The method of claim 1, The third step, A fifth step of determining whether there is free space in pages of the node to be inserted by the primary server and the at least one secondary server when there is free space in the node; A sixth step in which the main server and at least one sub-server insert a new entry and reflect the change of the minimum boundary area in the parent node when there is free space in the page as a result of the determination in the fifth step; And As a result of the determination of the fifth step, when there is no free space in the page, the seventh step of the primary server and at least one secondary server receives and inserts the page from the disk that the page is not allocated to the current node Data insertion method of a parallel high-dimensional index structure comprising a. In the data search method of a parallel high dimensional index structure applied to a parallel high dimensional index system, Transmitting root node information from the primary server and at least one secondary server to the input / output scheduler; A second step of determining whether there is information managed by the transferred I / O scheduler; A third step of performing a range query search according to whether the current node and the terminal node are identical by accessing several nodes in parallel by the primary server and at least one secondary server when there is information managed by the input / output scheduler. ; And If there is no information managed by an input / output scheduler, the at least one secondary server transmits the range query result to the primary server, and the primary server receives the range query result received from the at least one secondary server and its own. Fourth step of collecting and outputting the range query result Data search method of a parallel high-dimensional index structure comprising a. 4. The method of claim 3, wherein the third step is repeated until there is no information managed by the scheduler as a result of the determination of the second step. The method according to claim 3 or 4, The third step, A fifth step of determining whether the current node and the terminal node are the same by accessing several nodes in parallel from the input / output scheduler by the primary server and the at least one secondary server; As a result of the determination in the fifth step, when the current node and the terminal node are not the same, the primary server and the at least one secondary server select an entry included in the query range, and the information on the selected entry is input / output scheduler. A sixth step of delivering to; And As a result of the determination in the fifth step, when the current node and the terminal node are the same, the sixth step in which the primary server and the at least one secondary server calculate the similarity between the object and the query included in the node and store the calculated result. Data search method of a parallel high-dimensional index structure comprising a. In the data search method of a parallel high dimensional index structure applied to a parallel high dimensional index system, A first step in which the primary server or at least one secondary server accesses the root node to determine whether the accessed root node is a non-terminal node; As a result of the determination in the first step, when the accessed root node is a non-terminal node, the primary server and at least one secondary server calculate similarities between the query and the entries, and sort the calculated similarities in the order of similarity in the list. And a second step of accessing the child nodes of the first entry stored in the list in parallel and allocating the current node as the root node; As a result of the determination of the first step, when the approaching root node is not a non-terminal node, the primary server and at least one secondary server calculate the similarity between the query and the object so that only a value less than or equal to the K-nearest distance from the node is calculated. Extracting and storing the result set in a result set; A fourth step of determining whether the number of results stored in the result set is greater than or equal to K (k is a natural number) and the K th result is less than or equal to the similarity value of the first entry of the list; As a result of the determination in the fourth step, if the number of stored results is greater than or equal to K and the Kth result is less than or equal to the similarity value of the first entry in the list, the at least one secondary server returns K primary results. A fifth step in which the primary server collects the results K and the results found by the primary server from the at least one secondary server, sorts them in order of similarity, and selects K final results and returns them to the client; And As a result of the determination in the fourth step, if the number of stored results is not greater than or equal to K or if the Kth result is not less than or equal to the similarity value of the first entry in the list, the primary server and at least one secondary server are listed. Step 6 of allocating the current node as the root node by accessing the child nodes of the first entry stored in parallel in parallel Data search method of a parallel high-dimensional index structure comprising a. In the data search method of a parallel high dimensional index structure applied to a parallel high dimensional index system, A first step of accessing a root node by a K-access query inputted by the main server, and determining whether the accessed root node is a non-terminal node; As a result of the determination of the first step, if the accessed root node is a non-terminal node, the main server calculates the similarity between the query and the entries, sorts each similarity in the order of similarity in the list, and stores the first entry in the list. Accessing the child nodes of the second node in parallel and allocating the current node as a root node; As a result of the determination in the first step, when the accessed root node is not a non-terminal node, the main server calculates the similarity between the query and the object, extracts only a value less than or equal to the K-nearest distance from the node, and stores the result in the result set. Determining whether the number of stored results is greater than or equal to K; As a result of the determination in the third step, when the number of results stored in the result set is greater than or equal to K, the main server calculates the similarity between the K th result object and the query and converts it into a range query form. Transmitting a range query to at least one secondary server; At least one sub-server receiving the converted range performs a range query to deliver a search result of maximizing parallelism to the main server, and collects the search results received from the main server and the obtained results and sorts them in similarity order. Obtaining a final K results and returning the final K results to the client; And As a result of the determination in the third step, if the number of results stored in the result set is not greater than or equal to K, the main server accesses the child nodes of the first entry stored in the list in parallel and allocates the current node as the root node. 6th Step Data search method of a parallel high-dimensional index structure comprising a. In the data search method of a parallel high dimensional index structure applied to a parallel high dimensional index system, A first step in which the primary server and at least one department access the root note for the K-access query to determine whether the root node is a non-terminal node; As a result of the determination in the first step, when the root note is a non-terminal node, the primary server and at least one secondary server calculate similarities between the query and the entries, sort each similarity in the list in the order of similarity, and store them in the list. A second step of accessing the child nodes of the first entry in parallel and allocating the current node as the root node; If the root node is not the non-terminal node as a result of the first step, the primary server and at least one secondary server calculate the similarity between the query and the object, and extract only the value less than or equal to the K-nearest distance from the node. Storing in a result set and determining whether the number of stored results is greater than or equal to K; If the number of results is not greater than or equal to K as a result of the determination in step 3, the primary server and at least one secondary server access the child nodes of the first entry stored in the list in parallel to access the current node as the root node. 4th step of allocating As a result of the determination in the third step, when the result value is greater than or equal to K, the primary server and at least one secondary server calculate the similarity between the K th result object and the query and convert the similarity to the range query form. A fifth step of sending, by the secondary server, the converted range query to the primary server; The main server receiving the converted range query selects a range query having the smallest range among the ranged queries and the range query converted by the main server, and transmits the selected range query to at least one sub-server, and at least one received the range query. A sixth step of the secondary server performing a range query and delivering a result of the range query to the primary server; And A seventh step in which the primary server, which has received the result of executing the range query from at least one secondary server, collects the results found by the primary server, selects K results, and returns the selected K results to the client; Data search method of a parallel high-dimensional index structure comprising a. In a parallel high dimensional indexing system with a processor for inserting a query of a parallel high dimensional index structure, A first function in which a primary server allocates a new entry to each disk group and forwards the new entry to at least one secondary server determined including itself; The at least one secondary server and the primary server select a node through tree traversal for insertion of a new entry, and determine whether there is free space for inserting an entry in the selected node; A third function of inserting a new entry into a page free space of a node to be inserted by the main server and the at least one sub server when there is free space in the node as a result of the determination of the second function; And As a result of the determination of the second function, if there is no free space in the node, the primary server and the at least one secondary server allocate a page to the disk where the pages of the level where the current partition occurs are distributed, and insert a new entry in the node. Fourth feature to reflect changes in the minimum bounding region on the node A computer-readable recording medium having recorded thereon a program for realizing this. In order to search the data of the parallel high dimensional index structure, in a parallel high dimensional index system with a processor, A first function for the primary server and the at least one secondary server to transmit root node information to the input / output scheduler, and to determine whether there is information managed by the transferred input / output scheduler; As a result of the determination in the first function, if there is no information managed by the transmitted I / O scheduler, at least one secondary server transmits the range query result to the primary server, and the primary server delivers the at least one secondary server. A second function of collecting the received range query result and the range query result found by the user and delivering the collected query result to the client; And As a result of judging by the first function, if there is information managed by an input / output scheduler, the primary server and at least one secondary server access several nodes in parallel from the input / output scheduler to determine whether the current node and the terminal node are the same. Third function to perform range query search depending on whether A computer-readable recording medium having recorded thereon a program for realizing this. In order to search the data of the parallel high dimensional index structure, in a parallel high dimensional index system with a processor, A first function of the primary server and at least one secondary server accessing the root node to determine whether the accessed root node is a non-terminal node; As a result of the determination of the first function, when the accessed root node is a non-terminal node, the primary server and at least one secondary server calculate the similarity between the query and the entry, and sort each calculated similarity in the order of similarity in the list. And a second function of accessing the child nodes of the first entry stored in the list in parallel and allocating the current node as the root node; As a result of the determination of the first function, when the approaching root node is not a non-terminal node, the primary server and at least one secondary server calculate the similarity between the query and the object so that only a value less than or equal to the K-nearest distance from the node is calculated. A third function of extracting and storing the result set; A fourth function of determining whether the number of results stored in the result set is greater than or equal to K (k is a natural number) and the K th result is less than or equal to the similarity value of the first entry of the list; As a result of the determination of the fourth function, when the number of stored results is greater than or equal to K and the Kth result is less than or equal to the similarity value of the first entry in the list, at least one secondary server returns the K results in the primary server. A fifth function of collecting the result K and the result found by the primary server from the at least one secondary server and sorting the results found by the primary server, and selecting and returning K final results to the client; And As a result of the determination of the fourth function, when the number of stored results is not greater than or equal to K or the Kth result is not less than or equal to the similarity value of the first entry of the list, the primary server and at least one secondary server are listed. The sixth function of assigning the current node as the root node by accessing the child nodes of the first entry stored in parallel in parallel. A computer-readable recording medium having recorded thereon a program for realizing this. In order to search the data of the parallel high dimensional index structure, in a parallel high dimensional index system with a processor, A first function of accessing a root node through a K-access query input by the main server, and determining whether the accessed root node is a non-terminal node; As a result of the determination of the first function, if the accessed root node is a non-terminal node, the main server calculates the similarity between the query and the entry, sorts each similarity in the order of similarity in the list, and stores the first entry in the list. A second function of allocating a current node as a root node by accessing the child nodes in parallel; As a result of the determination in the first function, when the accessed root node is not a non-terminal node, the main server calculates the similarity between the query and the object, extracts only a value less than or equal to the K-nearest distance from the node, and stores the result in the result set. And determining whether the number of stored results is greater than or equal to K numbers; As a result of the third function, if the number of results stored in the result set is greater than or equal to K, the main server calculates the similarity between the K th result object and the query and converts it into a range query form. A fourth function of forwarding a range query to at least one secondary server; At least one sub-server receiving the converted range performs a range query to deliver a search result of maximizing parallelism to the main server, and collects the search results received from the main server and the obtained results and sorts them in similarity order. A fifth function of obtaining the final K results and returning the final K results to the client; And As a result of the determination of the third function, if the number of results stored in the result set is not greater than or equal to K, the main server accesses the child nodes of the first entry stored in the list in parallel and allocates the current node as the root node. 6th function to do A computer-readable recording medium having recorded thereon a program for realizing this. In order to search the data of the parallel high dimensional index structure, in a parallel high dimensional index system having a processor, A first function for determining whether the root node is a non-terminal node by accessing the root note for the K-access query by the primary server and the at least one department; As a result of the determination of the first function, when the root note is a non-terminal node, the primary server and at least one secondary server calculate similarities between the query and the entry, sort each similarity in the list in the order of similarity, and store the list in the list. A second function of allocating the current node as the root node by accessing the child nodes of the first entry in parallel; As a result of the determination of the first function, if the root node is not a non-terminal node, the primary server and at least one secondary server calculate the similarity between the query and the object and extract only the value less than or equal to the K-nearest distance from the node. A third function of storing in a result set and determining whether the number of stored results is greater than or equal to K; As a result of the determination of the above three functions, if the number of results is not greater than or equal to K, the primary server and at least one secondary server access the child nodes of the first entry stored in the list in parallel, and make the current node the root node. A fourth function of allocating; As a result of the determination of the third function, when the result value is greater than or equal to K, the main server and the at least one sub server calculate the similarity between the K th result object and the query and convert the result into a range query form. A fifth function for the secondary server to forward the converted range query to the primary server; The main server receiving the converted range query selects a range query having the smallest range among the ranged queries and the range query converted by the main server, and transmits the selected range query to at least one sub-server, and at least one received the range query. A sixth function of the secondary server performing a range query and delivering a result of the range query to the primary server; And The seventh function that the main server, which has received the result of executing the range query from at least one secondary server, collects the results found by itself and selects K results and returns the selected K results to the client. A computer-readable recording medium having recorded thereon a program for realizing this.