JP2000330959A - Inquiry parallel processing system and machine-readable recording medium where program is recorded - Google Patents

Abstract

PROBLEM TO BE SOLVED: To increase the total throughput of the inquiry parallel processing system by leveling the loads on respective computer nodes for partial inquiry processing when inquiries are processed in parallel on a cluster system composed of the computer nodes. SOLUTION: An inquiry analyzing means 11 analyzes a received inquiry to derive a processing procedure divided into partial inquiries. An inquiry integrating processing means 14 when distributing the partial inquiries to respective computer nodes 2-1 to 2-n determines the distribution number of partial inquiries to the respective computer nodes 2-1 to 2-n by using a distribution number determining means 13 so that the respective computer nodes are equal in the number of partial inquiries. The computer nodes at distribution destinations process the distributed partial inquiries and send the results back and an inquiry integrating process means 14 on a computer node at a distribution source derives the whole result while merging the returned results of the partial inquiries.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a technology for parallel processing of query statements describing data retrieval, grouping, and totaling operations for a database using a plurality of computer nodes, and more particularly, to processing of a plurality of computer nodes. The present invention relates to a query parallel processing system that improves the performance of the entire system by balancing loads.

[0002]

2. Description of the Related Art A method for processing queries to a database in parallel has been proposed by many researchers in the past ("Query Evaluation Techiniques for Large Datab").
ases ", Goetz Graefe, ACM Computing Surveys, Vol. 2
5, No.2 June 1993). Query processing for a huge database is a bottleneck in reading data from the secondary storage device. For this reason, a basic method is to divide and store a database in a plurality of secondary storage devices and read and process the data in parallel.

[0005] A process of performing a search on one database by designating conditions and extracting a corresponding record is referred to as a selection process. For example, the process of designating an account number in a bank account database to inquire the balance, or giving a fingerprint image to a fingerprint database and collating the same is the selection process. The database selection process can be easily parallelized by dividing and storing the database in a plurality of secondary storage devices. In this case, the respective partial databases divided and stored in the respective secondary storage devices are read in parallel from the respective computer nodes, and are subjected to a search process to be parallelized. In the selection process, an access path such as an index may be used. However, by independently creating an index for each partial database, the selection process using the index can be easily parallelized. These parallelization techniques have also been adopted in commercial relational database systems.
For example, Oracle ("Application of partitioning technology to VLDB", IEICE IEICE Technical Report)
AI97-40, DE97-73, 1997-12).

[0004] Classifying records in a database by the value of a certain attribute (column) is called classification processing. For example, a process of classifying records in a bank account database by branch. Also, processing such as counting the number of records or calculating an average value or a maximum value of a certain column is referred to as totaling processing, and processing of sorting a record group by a certain column value is referred to as sorting processing.

A technique for parallelizing these processes is disclosed in
No. 10-97544. In this technique, a database is stored in a plurality of secondary storage devices under a plurality of input / output servers, and records in the database are read from the plurality of input / output servers in parallel. The read records are classified by the input / output server and distributed to a plurality of tallying servers. The tallying server performs tallying processing on records from each input / output server to obtain a partial tally result, and finally merges the partial tally results. The processing up to this point is performed in parallel by a plurality of input / output servers and a plurality of aggregation processing servers.

As described above, in parallel processing of queries to a database, a database is divided into a plurality of partial databases, distributed and stored in a plurality of secondary storage devices, and a plurality of computer nodes transmit data from each secondary storage device in parallel. It is based on reading and processing. However, in the case where the databases are distributed and stored in the secondary storage device and the search processing is performed on each partial database in parallel, even if the partial databases are stored equally in each secondary storage device, each of them is read and processed. The loads on the computer nodes are not equalized, which may degrade the processing efficiency of the entire system.

[0007] In the case of such load imbalance, if a shared memory parallel computer generates more processes than the number of CPUs, it is possible to prevent the processing efficiency of the entire system from deteriorating. Because even if the load of each process becomes unbalanced
The process is switched by OS scheduling,
This is because the CPU can proceed with the query without playing. However, in a parallel computer using a plurality of computer nodes, a process cannot be moved between computer nodes. In the processing of the entire system, finally, the completion of the processing of the computer node with the heaviest load is waited for, so that the computer node with the lightest load idles and the processing efficiency of the entire system deteriorates.

On the other hand, Japanese Patent Application Laid-Open No. Hei 10-301822 describes the following technique. An access path is generated by analyzing a query to the database, and the access path is further decomposed into a plurality of partial access paths. Then, for each group of partial access paths that can be executed simultaneously, some of the partial access paths belonging to the group are grouped into one group according to the processing cost. Here, the processing cost is calculated from statistical information such as the number of rows in a table, the distribution of columns, and the hit ratio of an index, and is a numerical value of the degree of load required to execute a partial access path. .
When some partial access paths belonging to the partial access path group are grouped into one group according to the processing cost, the total value of the processing costs of the partial access paths to which the group belongs is determined by the maximum processing cost of the partial access path group. Do not exceed cost.

[0009]

DISCLOSURE OF THE INVENTION Problems to be Solved by the Invention
According to the technique described in the 1822 publication, the processing cost of each group to which the partial access path group that can be executed simultaneously belongs can be made substantially equal. Therefore, by assigning processes to a plurality of computer nodes in a group unit, the load on each computer node can be balanced to some extent. However, JP-A-10-3018
Since the technology described in Japanese Patent Publication No. 22 uses static information called statistical information, it cannot deal with the dynamic load fluctuation of each computer node, and as a result, the processing efficiency of the entire system can be sufficiently improved. There was a problem that can not be high.

An object of the present invention is to dynamically balance the processing load on each computer node when a query is processed in parallel on a cluster system composed of a plurality of computer nodes, so that the processing of the entire system is performed. The goal is to increase efficiency.

[0011]

A query parallel processing system according to the present invention is a query processing computer node which analyzes a query, decomposes the query into partial queries, and sends the partial query to a plurality of partial query processing computer nodes. The distributed partial query is performed, and the result of the returned partial query is merged into a query result. This configuration is a basic configuration for processing queries in parallel.

In the query parallel processing system of the present invention, when distributing a partial query to computer nodes for partial query processing, the number of distributions to each computer node is determined and distributed as needed.
At this time, in order to keep the load on each computer node equal, for example, the distribution number is determined so that the number of partial queries being processed in each computer node is always equal. The undistributed partial query that exceeds the number of distributions will be processed by the computer node later when any computer node returns a result. That is, the computer node that has returned the result is processed for the remaining partial query, so that the computer node does not play. The above is the basic concept of load balancing in the present invention.

In the query parallel processing system of the present invention, a plurality of query processing computer nodes are provided to prevent the computer nodes from becoming a bottleneck. in this case,
A plurality of computer nodes receive separate queries and independently perform partial query distribution processing for each query. The distribution strategy at this time is exactly the same as the case where an inquiry is received by a single computer node.

When a computer node for partial query processing fails, the computer node is detected, and a partial query to be distributed to the computer node is distributed to another computer node for partial query processing and processed. Thereby, even if a failure occurs in the computer node for partial query processing, all partial queries can be processed and the query result can be completed.

In the non-shared type cluster system, each partial query processing computer node can access only the secondary storage device connected to itself. Therefore, only the partial database stored in the secondary storage device can be accessed. Therefore, the correspondence between each computer node and the partial database to which it can be accessed is managed. Thus, a partial query that can be processed by the node can be distributed to each computer node.

In the non-shared type cluster system, the original and the copy of the partial database are stored in the secondary storage devices under different computer nodes. As a result, the same partial database can be accessed from a plurality of computer nodes. That is, the same partial query can be processed by a plurality of computer nodes. By utilizing this fact, it becomes possible to cause the computer node on the light load side to process the partial query, thereby performing load balancing of the computer nodes. At the same time, even when a computer node fails, it is possible to cause another computer node to process a query to be distributed to that node.

[0017]

Embodiments of the present invention will now be described in detail with reference to the drawings.

An example of the first embodiment of the present invention will be described. The first embodiment is based on a shared cluster system in which a plurality of computer nodes share a secondary storage device,
FIG. 1 shows an example of the configuration of a query parallel processing system for processing queries to a database recorded in a secondary storage device in parallel.

Referring to FIG. 1, a plurality of computer nodes 1-1 to 1-m for query processing and a plurality of computer nodes 2-1 to 2-n for partial query processing are connected by a communication network 4, and Data can be transferred between nodes. As the communication network 4, a hub or switch of Ether or ATM, or a dedicated connection network can be used.

Computer nodes 1-1 to 1- for query processing
m is connected to the external communication network 5 and can receive an inquiry via the external communication network 5. As the external communication network 5, a LAN or WAN is used. As a query,
Focusing on requests to extract some data from a database that is a set of data under certain conditions, requests to aggregate the average or maximum value of data values, or focus on certain data values There are requests to group or sort. These requests are described and passed using a query language SQL or the like, or are passed by combining only dedicated keywords.

A computer node 2 for partial query processing
-1 to 2-n are connected to a storage device sharing network 6 that shares the secondary storage devices 3-1 to 3-j, and any one of these computer nodes 2-1 to 2-n is also a secondary storage device. Apparatus 3-1
3-j can be accessed. Storage device sharing network 6
These include SCSI and FiberChanel HUBs and switches, as well as dedicated networks for peripheral devices.
The database to be queried is divided into a plurality of partial databases 31, and a plurality of secondary storage devices 3-1 to 3-
j and distributed. How to split the database
Methods such as round robin type distributed arrangement, hash division, and key value division are employed.

Computer nodes 1-1 to 1- for query processing
m is a query analysis unit 11 and a failure detection unit 12
, A distribution number determination unit 13, an inquiry integration processing unit 14, and a recording medium K1.

The query analysis means 11 analyzes a query addressed to the own computer node transmitted via the external communication network 5, and is a processing procedure for processing the query. A function for deriving a processing procedure including a partial query (a search request for the partial database) for each partial database to be performed.

The failure detecting means 12 has a function of detecting that a failure has occurred in the computer nodes 2-1 to 2-n for partial query processing.

The number-of-distribution determining means 13 is a computer node 2-1.
2 to the number of partial queries distributed to each computer node 2
It has a function of dynamically determining according to loads of -1 to 2-n.

The query integration processing means 14 is a query analysis means.
In accordance with the processing procedure derived in step 11, the query processing is advanced, and when distributing the partial query to each of the computer nodes 2-1 to 2-n, the undistributed number of distributions determined by the distribution number determination unit 13 is used. Is distributed, and when the result of the partial query is returned, a function of deriving the result of the query while merging the related results is provided.

The recording medium K1 provided in the computer nodes 1-1 to 1-m is a disk, a semiconductor memory, or another recording medium. The computer nodes 1-1 to 1-m are used as computer nodes for query processing. A program for functioning is recorded. This program runs on computer node 1
-1 to 1-m and read by computer node 1-1
1-m, the computer node 1-1
A query analysis unit 11, a failure detection unit 12, a distribution number determination unit 13, and a query integration processing unit 14 are realized on .about.1-m.

Computer nodes 2-1 to 2-1 for partial query processing
2-n includes a partial inquiry processing means 21 and a recording medium K2.

The partial query processing means 21 receives the partial query distributed from the query processing computer nodes 1-1 to 1-m, processes the partial query, and returns the result to the distribution source computer node. Having.

The recording medium K2 provided in the computer nodes 2-1 to 2-n is a disk, a semiconductor memory, or another recording medium. The computer nodes 2-1 to 2-n are used as computer nodes for partial query processing. A program for functioning as a program is recorded. This program is read by the computer nodes 2-1 to 2-n,
By controlling the operations of −1 to 2-n, the computer node 2
The partial query processing means 21 is realized on -1 to 2-n.

Next, the operation of this embodiment will be described.

Computer node 1 connected to external communication network 5
When the query analysis means 11 operating at −k (1 ≦ k ≦ m) receives the query addressed to the own computer node input via the external communication network 5, as shown in the flowchart of FIG. An analysis is performed to derive a processing procedure for obtaining a query result, including a processing step for a database to be processed (A1). For example, for a query as shown in FIG. 3A, a tree-like processing procedure including a plurality of steps as shown in FIG. 3B is required. In this example, since there are two databases to be processed, two processing steps for the databases are included. FIG. 8B is merely an example, and the processing procedure derived in A1 is in response to an inquiry. For example, in a case where data extracted from the databases 1 and 2 are combined and grouped, a processing procedure for performing a grouping process instead of the sorting process illustrated in FIG.

Next, the query analysis means 11 divides the processing steps for each database into processing steps (partial queries) for each of the partial databases constituting the database, and divides the processing steps for each set of partial queries for the same database. A processing step for merging the results of the query is added (A2). Here, a set of partial queries for one database is called a set of partial queries. When there are a plurality of databases to be processed, one set of partial queries is created for each database. For example, a processing procedure as shown in FIG. 3C is created from a processing procedure as shown in FIG. In this example, two sets of partial queries are created because one query targets two databases. In addition, the number of partial queries included in one set of partial queries is the same as the number of partial databases configuring the corresponding database since the partial queries are created for each database configuring the database.

When the query analysis means 11 derives the processing procedure for the query, the query integration processing means 14
As shown in the flowchart of FIG. 4, the processing procedure is obtained from the query analysis means 11 and stored, and it is determined whether or not a set of partial queries exists in the processing procedure (B1, B2). At the time when the processing procedure is obtained from the query analysis means 11, a set of partial queries exists, and the determination result of B2 is y
Since the answer is es, the query integration processing means 14 performs the process of B3.

In B3, the query integration processing means 14
Analyze dependencies between sets of partial queries included in the processing procedure, and recognize distributable partial queries. For example, if there are two sets of partial queries and they are independent sets with no processing dependency, all partial queries included in the two sets are recognized as distributable partial queries. For example, if there are two sets of partial queries S1 and S2 and there is a dependency relationship that the processing of the set S2 requires the processing result of the set S1, only the partial queries included in the set S1 Is recognized as a distributable partial query.

Thereafter, the query integration processing means 14 assigns the computer node to which the node number "0" is assigned among the computer nodes 2-1 to 2-n for the partial query processing to the computer to which the partial query is to be distributed. A node (B4). In the present embodiment, it is assumed that the node numbers “0” to “n−1” are assigned to the computer nodes 2-1 to 2-n, respectively.

Next, the query integration processing means 14 uses the distribution number determination means 13 to execute the computer node 2 having the node number “0”.
The number of distributions of the partial query to 1 is determined (B6). At this time, the distribution number is determined so that the number of partial queries being processed in each of the computer nodes 2-1 to 2-n is equal among the computer nodes.

The processing of B6 will be described in detail. When determining the number of partial queries to be distributed to the computer node 2-1, the query integration processing means 14 sends the node number "0" of the computer node 2-1 to be distributed to the number-of-distribution determining means 13.
give.

The number-of-distribution determining means 13 is a query integration processing means.
When the node number “0” is passed from 14 (FIG. 5, C1),
Referring to the progress management table, the number of partial queries for which no result has been returned from the computer node 2-1 (the number of results not returned) is acquired (C2).

The progress management table contains the computer nodes 2-1 to 2
-N result unreturned number and each computer node 2-1 to 2-
It manages distributed partial queries for which no result has been returned from n, and has, for example, the configuration shown in FIG. The progress management table shown in FIG. 6 includes the node numbers “0” to “0” of the computer nodes 2-1 to 2-n for the partial query processing.
In correspondence with “n−1”, the computer nodes 2-1 to 2-n
Are registered, and the IDs of partial queries for which no results have been returned from the computer nodes 2-1 to 2-n are registered. The number of returned results and the partial query ID are changed by the query integration processing means 14.

In C2, when the number of unreturned results of the computer node 2-1 is acquired, the number-of-distribution determining means 13 determines whether or not the condition of “number of unreturned results <target number of distributions” is satisfied (see FIG. 2). C3). Here, the target number of distributions is the computer node 1-k being processed by each of the computer nodes 2-1 to 2-n.
Is a value set in the computer node 1-k so as to make the number of partial queries from a predetermined number, and when the number of partial queries is made to be, for example, two, the target distribution number is “2”. Become. When it is determined that the condition is satisfied in C3 (the determination result is yes),
The number of distributions is obtained by performing an operation of "target distribution number-result unreturned number", and the obtained distribution number is notified to the inquiry integration processing means 14 (C4). On the other hand, when it is determined that the condition is not satisfied (C3 is no), the number of distributions = “0” is notified to the inquiry integration processing means 14 (C5). When the number of distributions is notified from the number-of-distribution determining means 13, the query integration processing means 14 determines the value as the number of partial queries distributed to the computer node 2-1. The above is the process performed in B6.

Then, the computer node 2- determined in B6
The number of partial queries distributed to 1 is greater than "0",
In addition, when there is an undistributed partial query that can be distributed (B
7 and B8 are both YES), a general-purpose communication protocol such as TCP, UDP / IP, or various dedicated communication protocols is used, and one of the undistributable distributable partial queries is sent to the computer node 2-. 1, the distributed partial query is excluded from the partial queries recognized as being distributable (B10), the number of distributions is reduced by one, and the computer node 2-1 in the progress management table is managed. The information is updated (B11). For example, when the progress management table has the configuration shown in FIG. 6, the number of unreturned results of the computer node 2-1 registered in association with the node number “0” is incremented by one, and the current computer node 2-1 is also updated. I of the partial query distributed to
D is registered in association with the node number “0”. The query integration processing means 14 distributes the number of partial queries determined in B6 to the computer node 2-1, or until the undistributed partial queries no longer exist (B7 becomes no, or B8 becomes no). The above-described processes B7 to B11 are repeated.

When B7 or B8 becomes no,
The query integration processing means 14 increments the node number by 1, and sets the computer node 2-2 as the computer node to which the partial query is to be distributed (B9). Hereinafter, the inquiry target integration processing means 14
The above-described processes B5 to B11 are repeated until “node number <number of nodes” (until B5 becomes no).
When the result of the determination at B5 is no, the inquiry integration processing means 14 enters a waiting state. By the above processing, the computer nodes 1-k to each of the computer nodes 2-1 to 2-n
The number of partial queries indicated by the target number of distributions is distributed. However, if the number of partial queries that can be distributed is smaller than “target distribution number × the number of computer nodes 2-1 to 2-n”, some of the computer nodes 2-1 to 2-n The number of partial queries indicated by the target number of distributions is distributed only to the computer nodes.

On the other hand, when a partial query is sent from the computer node 1-k, the partial query processing means 21 in the computer nodes 2-1 to 2-n, based on the partial database 31 designated by the partial query, Data that satisfies the conditions specified by the partial query is extracted, and the extracted data is returned to the query source computer node 1-k. At this time, the ID of the processed partial query and the node number of the own computer node are also returned.

The query integration processing means 14 in the computer node 1-k returns the partial query ID and node number as a result of the partial query from the computer node 2-i (1 ≦ i ≦ n) (FIG. 4). , B13), based on the result returned this time, it is checked whether or not there is a partial query set for which all results have been obtained (B14).

If it is determined that such a partial query set does not exist (B14 is no), the node number of the computer node 2-i that has returned the result is obtained, and The number of partial queries distributed to the node 2-i is determined (B16, B17). The processing of B17 is the same processing as B6.

After determining the number of partial queries to be distributed to the computer node 2-i, the query integration processing means 14 distributes the partial queries with the number of distributions determined in B17 to the computer node 2-i, or Until there is no more partial query that can be distributed (until B18 becomes no or B19 becomes no), the partial query is distributed to the computer node 2-i, and the distributed partial query is recognized as being distributeable. The process (B20) of excluding from the partial queries that are present and the process (B21) of updating the progress management table by subtracting the number of distributions by one are repeated. Here, for example, when the progress management table has the configuration shown in FIG. 6, the process of subtracting -1 from the number of unreturned results registered in association with the node number "i-1" of the computer node 2-i. Then, of the partial query IDs registered in association with the node number “i−1”, the ID returned from the computer 2-i is deleted. If the result of the determination at B18 or B19 is no, the query integration processing means 14 enters a waiting state.

Also, in B14, when it is determined from the result returned this time that there is a set of partial queries for which all results have been obtained (determination result is yes)
Removes the set of partial queries for which all results have been obtained from the stored processing procedure (B15), and then performs the processing of B2.

If it is determined in B2 that the set of partial queries exists in the stored processing procedure (the determination result is yes), the process of B3 is performed, and it is determined that the set of partial queries does not exist. Case (judgment result is n
In o), processing on the root side is sequentially performed from the partial inquiry toward the root, and the result of the inquiry is returned to the inquiry source via the external communication network 5 (B12). For example, if the processing procedure obtained in B1 is as shown in FIG.
As shown in FIG. 7, when all the results are obtained for two sets of partial queries, a process for merging the results for each set of partial queries is performed, and a process for combining the merged results and a process for sorting the combined results are performed. Perform sequentially. still,
In the present embodiment, merging is performed after all the results of the partial query are obtained.
There is a way to merge them one by one. For example, when an average value, a total value, or the like is obtained as a result for a query, the result is added for each set each time the result of the partial query returns. As a result, each step of the processing procedure is processed in a pipeline manner.

As described above, the query integration processing means 14
When the processing procedure is obtained from the query analysis means 11 (FIG. 4,
B1), a partial query of a predetermined target distribution number is distributed to each of the computer nodes 2-1 to 2-n (B5 to B1).
1) The remaining partial queries are not distributed. The non-distributed partial query is later thrown to the computer node that has returned the result of the partial query early (B16 to B21). In other words, the partial query that has not been distributed is distributed to the computer node whose load has been reduced due to the completion of the processing of the partial query.

As described above, each of the computer nodes 2-1 to 2-
The number of partial queries being processed at n is kept below a certain number. Therefore, in a situation where a large number of partial queries occur, the number of partial queries being processed in each computer node is always kept at a certain fixed number. As a result, the load on each computer node is balanced.

Next, FIG. 8 shows an embodiment of the partial query processing means 21 which operates independently in each of the computer nodes 2-1 to 2-n. In the embodiment shown in FIG. 8A, there is a queue for storing the distributed partial queries, from which the partial queries are taken out one by one to the partial query processing unit and processed sequentially. The result of the processed partial query is returned from the partial query processing unit to the distribution source computer node.

On the other hand, FIG. 8B shows an embodiment in which a plurality of partial queries are processed in a multiplex manner. By using a multi-thread or a multi-process, a plurality of partial query processing units are simultaneously operated, and each distributed partial query is allocated to each partial query processing unit. Each partial query processing unit operates multiplexly according to the scheduling of the OS of the computer node. For example, the partial query processing unit
In the case where the partial database is composed of a process of reading a partial database from the secondary storage devices 3-1 to 3-j in blocks of a fixed size and a process of performing a search / operation on the data of the read block, a plurality of partial query processing units In this case, the input / output processing of the secondary storage devices 3-1 to 3-j and the CPU processing in the computer node are temporally overlapped and processed.

The partial inquiry processing means 21 may be a combination of the queue shown in FIG. 8A and the queue shown in FIG. 8B.

Next, computer node 2 for partial query processing
The operation when a failure occurs in -1 to 2-n will be described.

The failure detecting means 12 in the query processing computer node 1-k is provided with a partial query processing computer node 2-k.
In order to detect that a failure has occurred in each of the computer nodes 2-1 to 2-n, as shown in the flowchart of FIG.
The state of n is monitored (D1). This monitoring is performed, for example, by periodically exchanging small data with each of the computer nodes 2-1 to 2-n via the communication network 4. That is, when data exchange can be performed normally, it is determined that no failure has occurred, and when data exchange has been disabled, it is determined that a failure has occurred. In addition, a dedicated failure diagnosis mechanism configured by hardware is provided in each of the computer nodes 2-1 to 2-n, and when a failure occurs, the failure diagnosis mechanism of each of the computers 2-1 to 2-n is connected via the communication network 4. The occurrence of a failure may be notified to the computer nodes 1-1 to 1-n.

Failure detection means 12 in computer node 1-k
For example, upon detecting that a failure has occurred in the computer node 2-1, first obtains the node number "0" of the computer node 2-1 (D2). Then, the failure detection means 12
Is the computer node 2 with the node number “0” from the progress management table
In step D3, the ID of the partial query that has been distributed to -1 and for which the result has not been returned is acquired. Further, the row of the node number “0” of the failed computer node is obtained from the progress management table. Then, the computer node 2-1 is deleted from the distribution target of the partial query (D4). For example, when the contents of the progress management table are as shown in FIG.
Acquires “ID1, ID2” as the ID of the partial query (D3), and deletes the first line of the progress management table (D
4). After that, the failure detecting means 13
14 is notified of the ID of the partial query acquired in D3 (D5). In the flowchart of FIG. 9, the row of the node number of the failed computer node is deleted from the progress management table, thereby excluding the failed computer node from the distribution target of the partial query. It is also possible to take a method of attaching a mark indicating "."

The query integration processing means 14 is
Is notified of the ID of the partial query, the partial query is recognized as an undistributed partial query. In this way, by recognizing the partial queries that have been distributed to the failed computer node and have not returned the results as the undistributed partial queries, they are recognized as having failed in B20 in FIG. It is distributed to any of the computer nodes for partial query processing that have not been performed. Secondary storage device 3-
In a shared cluster system in which 1 to 3-j are connected by the storage device shared network 6, each of the computer nodes 2-1 to 2-n
Since any of the partial databases 31 can be accessed via the storage device shared network 6, a partial query to be distributed to a certain computer node can be processed by another computer node. Therefore, all partial queries can be processed by distributing the partial queries distributed to the failed computer node to other normal computer nodes.

Next, an example of the second embodiment of the present invention will be described. In the second embodiment, an inquiry to a database recorded in a secondary storage device is performed on a non-shared type cluster system in which a secondary storage device accessible only by the computer node is connected to each of a plurality of computer nodes. FIG. 10 shows an example of the configuration of a query parallel processing system for processing in parallel.

The difference between this embodiment and the first embodiment shown in FIG. 1 is that computer nodes 10-1 to 10-m are provided instead of computer nodes 1-1 to 1-m. And that the secondary storage devices 30-1 to 30-n are directly connected to the respective computer nodes 2-1 to 2-n without the storage device sharing network, and that the storage medium K3 is used instead of the storage medium K1. It is a point that has.

The database to be queried is divided into a plurality of partial databases 31 and a plurality of secondary storage devices.
30-1 to 30-n. Further, copies of the respective partial databases are distributed and stored in a secondary storage device different from the original.

Each of the computer nodes 10-1 to 10-m is provided with a query integration processing means 14a instead of the query integration processing means 14 and a storage status management means 15 in FIG.
Are different from the computer nodes 1-1 to 1-m shown in FIG.

The storage status management means 15 manages the IDs of the partial databases accessible by the computer nodes 2-1 to 2-n for the partial query processing. When the node number of n is notified, it has a function of returning a list of IDs of partial databases accessible by the computer node of the notified node number. The management of the ID of the partial database 31 accessible to each of the computer nodes 2-1 to 2-n is performed using, for example, a storage status management table as shown in FIG. In the storage status management table, the IDs of the partial databases that can be accessed by the computer node of the node number in association with the node numbers “0” to “n−1” of the computer nodes 2-1 to 2-n
Is registered. In the example of FIG. 11, the computer node 2-1 having the node number “0” has IDs “A-0, A-4, A-
8, B-1, B-5,..., And IDs “A-2, A-6, A-10, B-
0, B-4,... "Can be accessed. This copy is a copy of the original of the partial database that can be accessed by the computer node 2-2 having the node number “1”. The ID of the partial database is the ID of the database such as A or B,
It is a combination of the order of the partial database in that.

The query integration processing means 14a has a function of notifying the storage status management means 15 of the node number of the computer node for partial query processing, in addition to the functions of the query integration processing means 14 shown in FIG. It has a function of determining a partial inquiry to be distributed based on the list of partial database IDs returned from the storage status management means 15 in response to this notification.

The recording medium K3 is a disk, a semiconductor memory, or another recording medium, and includes the computer nodes 10-1 to 10-10.
A program for causing −m to function as a computer node for query processing is recorded. This program is
By being read by the computer nodes 10-1 to 10-m and controlling the operations of the computer nodes 10-1 to 10-m, the query analysis unit 11 and the failure detection unit are provided on the computer nodes 10-1 to 10-m. 12, distribution number determination means 13, query integration processing means 14
a, The storage status management means 15 is realized.

FIG. 12 is a flowchart showing only a part of the processing performed by the query integration processing means 14a which is different from the processing performed by the query integration processing means 14 in FIG. Steps E1 and E2 are performed between steps B18 and B19, respectively, and step B1
FIG. 4 shows that steps E3 and E4 are performed in place of 0 and B20.
Is different from the processing shown in FIG. FIG. 13 is a flowchart showing a processing example of the storage status management means 15. Hereinafter, the operation of the present embodiment will be described with reference to the drawings.

When the inquiry analyzing means 11 in the inquiry processing computer node 10-k receives the inquiry addressed to the own computer node via the external communication network 5, it performs the processing shown in A1 and A2 in FIG. A processing procedure as shown in FIG.

When the query analysis means 11 derives the processing procedure, the query integration processing means 14a performs the same processing as B1 to B7 in FIG. 4, and executes the computer node 2-1 having the node number "0".
Determine the number of partial queries to be distributed to Thereafter, as shown in the flowchart of FIG. 12, the query integration processing means 14a determines one partial query to be distributed to the computer node 2-1 using the storage status management means 15 (E1). This E1
The processing of the above will be described in detail below.

The query integration processing means 14a notifies the storage status management means 15 of the node number "0" of the computer node 2-1 to which the partial query is to be distributed.

When the node number “0” is notified, the storage status management means 15 refers to the storage status management table as shown in FIG. 13 and the computer node 2-1 with the node number “0” can access. A list of partial database IDs is acquired (F1), and the acquired list is returned to the query integration processing means 14a (F2).

When the list is returned, the query integration processing means 14a returns one of the partial queries to the partial database whose ID is registered in the list and which is recognized as being distributable. Is selected as a partial query to distribute it. If there is no such partial query, it is determined that there is no partial query that can be distributed. The above is the details of the processing performed in E1.

The partial query to be distributed is determined in E1, and B
If the result of the determination in step 8 is yes, the query integration processing means 14
a, the partial query determined in E1 is transferred to the computer node 2-1.
And the distributed partial query is excluded from the partial queries recognized as being distributable (E3). afterwards,
The same processing as B11 shown in FIG. 4 is performed.

Computer node 2 for partial query processing
FIG. 4 shows a case where the result of the partial query is sent from -i.
Of the computer node 2-i
Determine the number of partial queries to be distributed to Thereafter, the query integration processing means 14a uses the storage status management means 15 to determine one partial query to be distributed to the computer node 2-i (E2), distributes the partial query to the computer node 2-i, and Then, the distributed partial query is excluded from the partial queries recognized as being distributable (E4). Thereafter, the same processing as B21 shown in FIG. 4 is performed.

Next, the operation when the failure detecting means 12 detects that a failure has occurred in the computer node 2-i will be described.

When the failure detecting means 12 detects that a failure has occurred in the computer node 2-i as shown in the flow chart of FIG. 9, the failure has been distributed to the computer node 2-i and no result has been returned. The ID of the partial query is obtained from the progress management table, and is notified to the query integration processing means 14a. The query integration processing means 14a recognizes the partial query for which the ID has been notified as an undistributed partial query. As described above, by recognizing a partial query that has been distributed to the failed computer node and has not returned a result as an undistributed partial query, the partial query described above in E4 of FIG. Are distributed to computer nodes that are operating normally and are capable of accessing the partial database corresponding to the partial query described above.

[0076]

As described above, the query parallel processing system of the present invention reduces the load on each computer node when executing a query in parallel on a cluster system composed of a plurality of computer nodes for partial query processing. Since the number of partial queries to be distributed dynamically is determined accordingly,
The load on each computer node can be dynamically balanced. As a result, the processing efficiency of the entire query parallel processing system can be increased, and the performance can be improved.

Further, in the query parallel processing system of the present invention, when a failure occurs in a computer node for partial query processing, it should be dynamically distributed to computer nodes other than the failed computer node in accordance with the load. Since the number of partial queries is determined, the result of the query can be completed even if a failure occurs in the computer node for processing the partial query.

[Brief description of the drawings]

FIG. 1 is a block diagram illustrating a configuration example of a first embodiment of the present invention.

FIG. 2 is a flowchart showing a processing example of a query analysis unit 11;

FIG. 3 is a diagram for explaining the operation of a query analysis unit 11;

FIG. 4 is a flowchart showing a processing example of an inquiry integration processing means 14;

FIG. 5 is a flowchart showing a processing example of a distribution number determining unit 13;

FIG. 6 is a diagram showing a configuration example of a progress management table.

FIG. 7 is a flowchart showing a processing example of the inquiry integration processing means 14;

FIG. 8 is a block diagram showing an embodiment of a partial inquiry processing means 21.

FIG. 9 is a flowchart illustrating a processing example of a failure detection unit 12;

FIG. 10 is a block diagram illustrating a configuration example of a second embodiment of the present invention.

FIG. 11 is a diagram illustrating a configuration example of a storage status management table.

FIG. 12 is a flowchart showing a part of a processing example of an inquiry integration processing means 14a.

FIG. 13 is a flowchart showing a processing example of the storage status management means 15;

[Explanation of symbols]

1-1 to 1-m, 10-1 to 10-m ... computer nodes for query processing 11 ... query analysis means 12 ... fault detection means 13 ... number of distribution determination means 14, 14a ... query integration processing means 15 ... storage status Management means 2-1 to 2-n Computer node for partial query processing 21 Partial query processing means 3-1 to 3-j, 30-1 to 30-n Secondary storage device 31 Partial database 4 Communication Network 5: External communication network 6: Storage device sharing network K1 to K3: Recording medium

Claims (14)

[Claims] A shared storage system in which a plurality of secondary storage devices in which a plurality of partial databases created by dividing a database into a plurality of portions are dispersedly stored are shared by a plurality of computer nodes for partial query processing. A query parallel processing system for processing queries to the database in parallel on a type cluster system, comprising: a query processing computer node that receives a query to the database; and wherein the query processing computer node is A procedure for analyzing a query to the database and processing the query, the procedure including a partial query for each partial database constituting the database to be queried and a merging process of the result of the partial query are derived. Query analysis means, and a computer node for each partial query processing Distribution number determining means for dynamically determining the number of partial queries to be distributed according to the load; and processing of the query in accordance with the processing procedure derived by the query analysis means. When distributing a partial query to the user, distribute the undistributed partial queries for the number of distributions determined using the distribution number determining means, and when the result of the partial query is returned, Query integration processing means for deriving a query result while merging, wherein the plurality of partial query processing computer nodes respectively receive the distributed partial query, process the partial query, A query parallel processing system comprising a partial query processing means for returning a result to a computer node. 2. The query parallel processing system according to claim 1, wherein said distribution number determination means sets a predetermined number of partial queries being processed by each of said partial query processing computer nodes. A query parallel processing system having a configuration for determining the number of partial queries to be distributed. 3. The query parallel processing system according to claim 1, further comprising a plurality of query processing computer nodes. 4. A shared storage system in which a plurality of secondary storage devices in which a plurality of partial databases created by dividing a database into a plurality of portions are distributed and stored are shared by a plurality of computer nodes for partial query processing. A query parallel processing system for processing queries to the database in parallel on a type cluster system, comprising: a query processing computer node that receives a query to the database; and wherein the query processing computer node is A procedure for analyzing a query to the database and processing the query, the procedure including a partial query for each partial database constituting the database to be queried and a merging process of the result of the partial query are derived. The query analysis means and the computer node for the partial query processing fail. Failure detection means for detecting that a failure has occurred, and among the computer nodes for partial query processing, of the computer nodes other than the computer node for which the failure detection has been detected by the failure detection means, dynamically responding to the load thereof. A distribution number determining means for determining the number of partial queries to be distributed to the query processing section, and processing of the query in accordance with the processing procedure derived by the query analyzing means, and when distributing the partial query to each of the partial query processing computer nodes, Distributes undistributed partial queries for the number of distributions determined using the distribution number determining means, and derives query results while merging related results when the results of the partial queries are returned. Query integration processing means, wherein each of the plurality of computer nodes for partial query processing receives the distributed partial query, processes the partial query, Contact parallel processing system comprising the partial query processing unit that returns the result to the fabric source computer nodes. 5. The query parallel processing system according to claim 4, wherein the distribution number determining means is a computer node for partial query processing other than a computer node whose failure has been detected by the failure detecting means. A query parallel processing system having a configuration for determining the distribution number of partial queries so that the number of partial queries being processed by a computer node becomes a predetermined number, respectively. 6. The query parallel processing system according to claim 4, comprising a plurality of computer nodes for said query processing. 7. A plurality of partial query processing computer nodes are each connected to a secondary storage device accessible only by the computer node, and the plurality of partial databases created by dividing the database into a plurality of parts are provided. A query parallel processing system for processing queries to the database in parallel on a non-shared cluster system distributed and stored in a plurality of secondary storage devices, wherein one query process for receiving a query to the database A computer node for query processing, wherein the computer node for query processing analyzes the query to the database and processes the query, comprising: a partial query to each partial database constituting the database to be queried; Introduce a processing procedure including merge processing of partial query results. Query analysis means to be issued, a computer node for each of the partial query processing, a storage status management means for managing a correspondence relationship with a partial database accessible from the computer node, and a computer node for each of the partial query processing A distribution number determining unit that dynamically determines the number of partial queries to be distributed according to the load; and a query processing unit that proceeds with the query according to the processing procedure derived by the query analysis unit, and sends the query to a computer node for partial query processing. When distributing the partial query, all the partial databases accessible by the computer node are recognized by using the storage status management unit, and from among the partial queries to those partial databases, the distribution number determining unit is used. The undelivered partial queries for the determined number of distributions are distributed, and when the result of the partial query is returned, the related Query integration processing means for deriving a query result while merging results, wherein the computer nodes for each of the partial query processes receive the distributed partial query, perform the partial query processing, and And a partial query processing means for returning a result to the computer node. 8. The query parallel processing system according to claim 7, wherein said distribution number determining means determines the number of partial queries being processed in each of said partial query processing computer nodes by a predetermined number. A query parallel processing system having a configuration for determining the number of queries to be distributed. 9. The query parallel processing system according to claim 7, further comprising a plurality of query processing computer nodes. 10. A plurality of partial query processing computer nodes are each connected to a secondary storage device accessible only by the computer node, and the plurality of partial databases created by dividing the database into a plurality of parts are provided. On a non-shared cluster system in which a copy of each of the partial databases is stored in a secondary storage device different from the original, the queries for the databases are distributed and stored in a plurality of secondary storage devices, and queries to the databases are performed in parallel. A query parallel processing system for processing, comprising one query processing computer node for receiving a query to the database, wherein the query processing computer node analyzes the query to the database and processes the query. This is a procedure, and each partial data constituting the database to be queried is A query analysis unit that derives a processing procedure including a partial query to the database and a merge process of the result of the partial query; a failure detection unit that detects that a failure has occurred in the computer node for the partial query processing; For the computer nodes for query processing, of the computer nodes other than the computer node for which the failure detection is detected by the failure detecting means, the distribution number determination for dynamically determining the number of partial queries to be distributed according to the load. Means, a computer node for each of the partial query processing, a storage status managing means for managing a correspondence relationship between an original and a copy of a partial database accessible from the computer node, and a processing procedure derived by the query analyzing means. When processing the query and distributing the partial query to the computer node for partial query processing, Recognize all of the accessible partial databases using the storage status management means, and among the partial queries for those partial databases, extract undistributed partial queries for the number of distributions determined by the distribution number determination means. And query integration processing means for deriving the result of the query while merging the related results when the result of the partial query is returned. The computer nodes for the partial query processing are distributed, respectively. A query parallel processing system comprising: a partial query processing unit that receives a partial query, processes the partial query, and returns a result to a distribution source computer node. 11. The query parallel processing system according to claim 10, wherein said distribution number determining means is a computer node for which said failure detecting means detects a failure among said computer nodes for said partial query processing. A query parallel processing system characterized in that the number of partial queries distributed is determined so that the number of partial queries being processed by the computer nodes becomes a predetermined number. 12. The query parallel processing system according to claim 10, further comprising a plurality of query processing computer nodes. 13. A shared configuration in which a plurality of secondary storage devices in which a plurality of partial databases created by dividing a database into a plurality of portions are distributed and stored are shared by a plurality of computer nodes for partial query processing. A program for realizing a query parallel processing system for processing queries to the database in parallel on the type cluster system by a plurality of computer nodes for partial query processing and one computer node for query processing A computer-readable recording medium on which the computer node for query processing is analyzed, and a query for the database is processed, and the query is processed. Processing procedure including a partial query for, and a merge process of the result of the partial query Query analysis means to be derived, distribution number determination means for dynamically determining the number of partial queries to be distributed according to the load, for each of the partial query processing computer nodes, according to the processing procedure derived by the query analysis means When a query is advanced and a partial query is distributed to each partial query processing computer node, undistributed partial queries corresponding to the distribution number determined by using the distribution number determining means are distributed, and the partial query is distributed. When the result is returned, a program for functioning as a query integration processing means for deriving a query result while merging related results, and a computer node for each of the partial query processing, A program for functioning as a partial query processing means that receives a query, processes the partial query, and returns a result to the distribution source computer node. Was recorded bets, machine-readable recording medium recording a program. 14. A plurality of partial query processing computer nodes are each connected to a secondary storage device accessible only by the computer node, and the plurality of partial databases created by dividing the database into a plurality of parts are provided. On a non-shared cluster system distributed and stored in a plurality of secondary storage devices, a query parallel processing system for processing queries to the database in parallel is provided with a plurality of computer nodes for partial query processing and one computer A computer-readable recording medium that records a program to be realized by an inquiry processing computer node, wherein the inquiry processing computer node analyzes an inquiry to the database and processes the inquiry. Each database that composes the database to be queried Query analysis means for deriving a processing procedure including a partial query to be performed and a merge process of a result of the partial query, and manages a correspondence relationship between the computer nodes for each of the partial query processes and a partial database accessible from the computer node. Storage status management means, distribution number determination means for determining the number of partial queries to be distributed according to the load for each of the partial query processing computer nodes, processing of queries according to the processing procedure derived by the query analysis means When distributing the partial query to the computer nodes for processing the partial queries, all the partial databases accessible by the computer nodes are recognized using the storage status management means, and the partial From the queries, undistributed partial queries for the number of distributions determined using the distribution number determination means And, when the result of the partial query is returned, a program for functioning as query integration processing means for deriving the result of the query while merging the related results, and a computer node for the partial query processing A program for functioning as a partial query processing means for receiving the distributed partial query, processing the partial query, and returning a result to the distribution source computer node; recoding media.