JP2000322292A - Cluster type data server system and data storage method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To increase the processing speed of an entire cluster system by adaptively distributing the load of processing dependent upon specific data on a data storage device to plural disks and computer nodes at the time of execution to uniformly charge individual data storage devices and computer nodes with the load in a non-shaped cluster system. SOLUTION: Data to be stored in a disk device is roughly divided and is subdivided furthermore, and all subdivided data 100-i1 to 100-im derived from the same roughly divided data are stored in the same disk, and meanwhile, copies 110-i1 to 110-im,..., 1r0-i1 to 1r0-im of respective subdivided data are stored in disks different from the disk, where original data is stored, and different from one another, and processing requests requiring individual subdivided data are distributed to the computer node having the disk, where original data is stored, and computer nodes having disks, where copy data are stored, in consideration of respective load conditions.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a cluster system composed of a plurality of computer nodes, and more particularly to a request from a client outside a cluster in a non-shared type cluster system in which each computer node has a local data storage device. And a server system that executes a server application that processes data using data in a local data storage device of each node.

[0002]

2. Description of the Related Art A system constituted by using a plurality of independent computers to execute one application is called a cluster system. The difference between a cluster system and a distributed system is that, in a cluster system, the entire cluster is always involved in executing an application. The difference between the cluster system and the parallel computer is that the cluster system is configured as a set of independent computers that can be used alone.

[0003] The purpose of a cluster system is threefold: large, high performance, large capacity, and high reliability. That is,
By using more computers, achieving higher processing performance, handling more data, and using multiple computers, even if some of them are stopped, the remaining computers That is, processing can be continued.

[0004] If these objects are achieved, a cluster system is suitable for a system that executes a large-scale server application. This is because such an application generally needs to process a large amount of data on a disk device (including a magnetic disk device, an optical disk device, and the like) at a high speed, and always receives a request from a client without interruption. It is necessary to be able to respond to

[0005] Server applications that perform processing using a large amount of data on a disk device include an online transaction processing server on a database management system and multimedia that supplies such data to a client that plays back audio and video data. There are known a stream server, a full-text search server that finds a document including a keyword specified by a user from a huge number of documents, and returns a list of the documents.

[0006] In a cluster system, it is roughly classified into two types according to a mode of connecting a disk and a computer.
One of them is called "shared disk type", which connects a computer and a disk device via a network so that any computer can access any disk device in the same procedure. The other is
This is called a "shared nothing type", in which a disk device is directly connected locally to each computer, and a network for interconnecting the computers is provided.

[0007] Compared to the non-shared type, the shared disk type has an advantage that processing requiring specific data on a disk device can be performed by any computer node in the cluster.

Conventionally, a transaction processing server or the like built on a disk-sharing type cluster system constantly monitors the load and operating state of each computer node, and receives a new processing request from a client. Some include means for selecting a computer node with a lighter load from among them and assigning processing to the selected computer node. In the transaction processing server, it is called a transaction monitor.

With this configuration, if a computer node is stopped or the load is high, another arbitrary computer node in the cluster is selected as a substitute for the computer node, and the processing is substituted and continued. be able to.

In this manner, in the disk-sharing type cluster system, it is possible to relatively easily continue operation when a certain computer node is stopped and to equalize the load on each computer node.

On the other hand, in a non-shared cluster, in order for one computer node to access a disk device that is locally connected to another computer node, the disk device must be transmitted from one computer node to another computer node. It is necessary to request input / output processing to the disk device and to have another computer node take over the input / output processing to the disk device. This is called "remote disk access".

At this time, the contents of the input / output processing requested via the network and the accompanying data are exchanged between the two computer nodes, and one computer node is connected to the local disk directly connected to the one computer node. Access time increases compared to accessing the device,
Network traffic and the like increase.

Further, when one computer node in the cluster is stopped, another computer node accesses data stored in a local disk device directly connected to the stopped one computer node. Means are gone.

For this reason, the non-shared cluster system is
It is difficult for the computer node to continue operation when the computer node is stopped as it is.

In a non-shared type cluster system, as a technique for continuing operation when a computer node stops, a redundant configuration such as a disk mirroring or a redundant disk array is well known. In this method, the same data is stored on disks connected to a plurality of different computer nodes.

Thus, even if one computer node stops or a local disk device connected to the computer node stops, processing can be continued by using the same data copy in another disk device. .

The distributed file system “Ti” described below
"gerShark" is an example of such a system. "Tige
rShark ”is described in the literature (RL Haskin,“ Tiger Sh
ark-- A scalable file system for multimedia ”, IBM
J. RES. DEVELOP. VOL. 42 NO. 2, March 1998, pp187
-189).

One of the features of the “Tiger Shark” is a technique called striping, in which files are distributed and stored in disk devices connected to a number of different computer nodes. Briefly, when there are n computer nodes and their local disk devices, the i-th block in the file is stored in the local disk device of the (i mod n) -th (mod is a remainder operator) computer node. .

Another feature of this "Tiger Shark" is a distributed redundant storage technique in which a copy of a file is created in block units and placed on a disk device different from the original. This distributed file system is used to realize a video stream server.

The striping technique has the following effects. A server that supplies stream data such as a video (video signal) and an audio signal is required to have a capability of supplying a large amount of data per unit time. The data supply amount per unit time is called “throughput”.

The throughput required for a stream data server often exceeds the throughput that can be provided by a single disk device or computer node.

In view of the above, a plurality of computer nodes read data in parallel from different disk devices and send the data to the clients in parallel with a plurality of computer nodes, thereby achieving high data supply performance as a whole cluster system. be able to.

On the other hand, the functions and effects of the distributed redundant storage technique are as follows.

By providing data replication to a plurality of disk devices in the first place, data supply can be continued even when some disk devices or computer nodes fail. "Tiger Shark" also uses a technology called "Equal Random Striping" to evenly load the remaining normal disk units and computer nodes in order to replace the failed computer nodes and disk units when a failure occurs. It is going to take.

In uniform random striping,
When files are distributed and stored in k disk units, blocks forming one file are k blocks from the beginning.
The blocks are divided into individual groups, and the blocks are arranged in a randomly determined order different for each group. By duplicating each file equally in the same manner, if data on a certain disk device cannot be used due to a failure or the like, the duplication exists in an even ratio on a different disk for each block group. Therefore, as long as the file is sequentially read from the head, the file read load is equally applied to different disk devices.

[0026]

However, conventional systems such as the above-mentioned "Tiger Shark" have the following problems.

The first problem is that when used for purposes other than the supply of multimedia stream data, there is a high possibility that the processing load on the disk device and the computer node will be biased.

[0028] The reason is that in "Tiger Shark", the file to be handled is a multimedia such as a moving image or audio.
This is because media stream data is assumed, and a configuration suitable for an access method of reading relatively large files in order from the beginning to the end is specialized.

That is, in the case of "Tiger Shark", each file is distributed and stored in a uniformly striped disk device across all the computer nodes. Therefore, even if the access frequency is uneven for each file, the processing load is not uneven for each computer node.

However, in online transaction processing and search processing using an index, the access frequency differs depending on the location within one file,
The range of data to be accessed may be different each time.

In the search processing for reading and scanning the entire file in order from the beginning to the end, the access format of the disk device is the same as that of the stream data, but the same amount of data is determined depending on the number of data corresponding to the search condition. Even if it is read, the processing amount for each computer node may vary greatly.

The above-mentioned "Tiger Shark" does not have a countermeasure against such load imbalance.

The second problem is that, when used for applications other than the supply of multimedia stream data, the conventional striping or uniform random striping does not always improve the processing performance, and conversely, may decrease the processing performance. That is.

The reason is basically the same as that of the first problem. In stream data access, it can be predicted that it is necessary to read a relatively large file in order from the beginning to the end. Therefore, striping is performed, and all disk units are simultaneously instructed to start reading and parallel processing is performed. And thereby improve throughput.

However, in the transaction processing and the search processing using the index file, access to the file is performed not on the entire file but on a small part thereof, and not in the order from the beginning but in an unspecified unpredictable order.

For this reason, individual access to the striped file is not performed for all disk devices,
Access to several disk devices is required, and the effect of parallel processing is weak.

Since a series of accesses is likely to be made to different disk devices, the effect of parallel processing by a plurality of disk devices can be expected. However, unlike stream data, in any part of a file, Since it is not possible to predict in what order access will occur, access can only be started after each access request has arrived, and the effect of parallel processing is weak.

Furthermore, when compared with a configuration in which one file is stored in one disk device without striping,
It takes time and effort to perform remote disk access to a plurality of disk devices, and on the contrary, processing performance may be reduced.

A third problem is that when the striping is stopped, a large deviation occurs in the load distribution.

The reason is that if the striping is completely stopped and all the blocks of the file are stored in one disk device, accesses are concentrated on the disk device.

Even if the situation is not so extreme, another problem arises.

For example, as an intermediate technique, when there are k disks, a method of setting a large block size so that a file can be accommodated in k blocks and then applying uniform random striping can be considered. .

In this case, since only one group of blocks can be formed, a copy of a block arranged in one disk device is stored only in another disk device.

For this reason, when one disk device stops, the load taking over for that disk device is concentrated on the other disk device, and the effect of the uniform random striping is lost.

To avoid this problem, the file must be (m
× k) The block size can be set so that it fits within the number of blocks, and even random striping can be performed.However, as the value of m is increased, the load on the replacement disk device is more well-distributed. There remains a problem that the second problem becomes more remarkable.

Therefore, the present invention has been made in view of the above problems, and an object of the present invention is to reduce a processing load depending on specific data on a data storage device in a non-shared cluster system. To provide a system and method for adaptively spreading to a plurality of disks and computer nodes at the time of execution to equalize the load on each data storage device and computer node, and consequently speeding up the processing of the entire cluster system. It is in.

Another object of the present invention is to distribute the load of the data storage device or the computer node to a plurality of operating computer nodes evenly even when some of the data storage devices or the computer nodes are stopped. To provide a system and method that can take over.

[0048]

According to a first aspect of the present invention, there is provided a cluster server system comprising a plurality of computer nodes, wherein each of the computer nodes has a local data storage device. A server system for performing a process depending on data on a data storage device on a system, wherein coarsely divided data stored in the data storage device is defined as coarsely divided data, and the coarsely divided data is further divided. The divided data is regarded as subdivided data, all subdivided data derived from the same coarsely divided data are stored in the same data storage device, and a copy of all subdivided data derived from the same coarsely divided data is referred to as an original. The data is stored in different and different data storage devices.

According to a second aspect of the present invention, in the server system according to the first aspect, an input / output for certain subdivided data is performed by a data storage device that stores an original of the subdivided data and a data storage that stores a copy of the subdivided data. There is provided an input / output request allocating means for allocating the request to be executed using any of the devices.

According to a third aspect of the present invention, in the server system according to the first aspect of the present invention, a request for a process requiring certain subdivided data is sent to a computer node having a data storage device for storing an original of the subdivided data. And a processing request allocating means for allocating the processing to one of the computer nodes having a data storage device for storing a copy of the subdivided data.

According to a fourth aspect of the present invention, in the server system according to the second aspect of the present invention, a load on a data storage device for recording an original of certain subdivided data and a load on a data storage device for recording a copy of the subdivided data are estimated. Means for allocating an input / output request for the subdivided data so as to execute input / output to the subdivided data using a data storage device estimated to have a lighter load. .

Further, the fifth invention of the present application is the server system according to the third invention, wherein a load on a computer node having a data storage device for recording an original of certain subdivided data;
Means for estimating the load on a computer node having a data storage device for recording a copy of the subdivided data, and processing requiring the subdivided data using the computer node estimated to have a lighter load. Request allocation means for allocating a processing request that requires the subdivision data to be executed.

According to a sixth aspect of the present invention, in the server system according to the second aspect of the present invention, there is provided a computer node comprising means for monitoring the operating state of each computer node, and comprising a data storage device for recording an original of certain subdivided data. An input / output request for the subdivided data so as to execute input / output for the subdivided data by using an operating computer node among computer nodes having a data storage device for recording a copy of the subdivided data Request allocation means for distributing the request.

Further, the seventh invention of the present application is the server system according to the third invention, further comprising means for monitoring the operating state of each computer node, and a computer node having a data storage device for recording an original of certain subdivided data. Needing the subdivided data so as to execute a process that requires the subdivided data using an operating computer node among computer nodes having a data storage device that records a copy of the subdivided data. Request allocation means for allocating processing requests to be performed.

[0055]

Embodiments of the present invention will be described. First, the principle and operation of the present invention will be described.

As an embodiment in which the present invention is applied to a server system having a non-shared cluster system configuration, a data storage device (2_i) (where i is an integer of 1 or more and n or less), a request processing means (7_i), and a file system ( 5_i), and a network (3) connecting all of these computer nodes.

In the present invention, the data to be stored and arranged on the data storage device is roughly divided into coarse part data, and the coarse part data is further divided into fine part data to form the same coarse part data. Is stored in the data storage device of the same computer node, and one or more copies of the small data are stored in different data storage devices that are different from the original data.

A request for a process requiring data on the data storage device is distributed to a data storage device having original data and a data storage device having a copy in units of small data portions.

More specifically, at the time of execution, it is determined at the time of execution whether a processing request is to be assigned to an original or a copy of small-part data based on information on the operating state and load state of each data storage device or computer node. , Request allocating means (8_i).

In the present invention, all the fine data derived from one coarse data is stored in the same data storage device, while the duplicated data is distributed to a plurality of data storage devices. Is remembered.

Each time the small data is required to satisfy the processing request, the request allocating means determines whether to use the original or the duplicate of the small data.

For this reason, when all the data storage devices and the computer nodes are operating and the load is substantially equally applied, access to a relatively wide range of data that can be contained in one coarse partial data is performed. At most one access to one data storage device to store
It can be done with remote disk access times.

When the load on a certain data storage device or a computer node having the data storage device is high, the request allocating means decides to use the copy instead of the original stored on the data storage device. By increasing the ratio, the load of the data storage device or the computer node can be spread to a plurality of data storage devices or computer nodes, and the load imbalance between the data storage devices or the computer nodes can be equalized.

Further, when a certain data storage device or a computer node is stopped, the request allocating means uses the copy instead of the original on the data storage device or the data storage device of the computer node. By determining, the load of the data storage device or the computer node can be replaced by a plurality of data storage devices or computer nodes.

In the present invention, the data storage device is not limited to an HDD (hard disk device), but includes various auxiliary storage devices which can freely write and read data on and from a medium. Further, the data storage device provided in each computer node only needs to logically constitute one data storage device, and is of course not limited to a single drive control device and drive unit.

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

[0067]

Embodiment 1 FIG. 1 is a diagram showing a configuration of a first embodiment of the present invention. In the following, a description will be given assuming that the data storage device is a disk device. Referring to FIG. 1, a first embodiment of the present invention provides a plurality of computer nodes 1_1 to 1_n (n is an integer of 2 or more) and a disk device 2_1.
To 2_n and a SAN (System Area Network; network within the system) 3, and each of the computer nodes 1_1 to 1_n is a LAN (Local
Area Network (Local Area Network) 4 is also connected.

Each of the computer nodes 1_1 to 1_n includes a file system 5_1 to 5_n, a request receiving unit 6_1 to 6_n, and a request processing unit 7_1 to 7_n.

The file systems 5_1 to 5_n are respectively
It has request allocating means 8_1 to 8_n.

Each request receiving means of each of the computer nodes 1_1 to 1_n
6_i (where i is an integer of 1 or more and n or less) receives a processing request that requires data on the disk devices 2_1 to 2_n, divides the processing request into a plurality of partial processing requests, and The partial processing requests are sent to request processing means 7_1 to 7_n, respectively.
And integrates the partial processing results returned as a response to each request into one processing result.

The request processing means 7_1 to 7_n are connected to the request receiving means 6_
i creates a partial processing result in response to the partial processing request.

The file systems 5_1 to 5_n receive an input / output request for a file and perform input / output for data on a disk device.

The request allocating means 8_1 to 8_n determine which one of the original file and one or more copies of the file is used to process the input / output request for the file.

FIG. 2 is a diagram for explaining the processing according to the first embodiment of the present invention. Referring to FIG. 2, the data on the disk device is divided into n pieces of coarse data,
Each coarse part data is further divided into m pieces of fine part data 100_ij (where i is an integer of 1 or more and n or less, j is an integer of 1 or more and m or less) and r pieces (where r is 1 or more) Integer)
Along with 110_ij, 120_ij, ..., 1r0_ij, they are arranged as follows.

On the disk device 2_i (where i is an integer of 1 or more and n or less), there are original data 100_i1 to 100_im (where m is an integer of 1 or more and n-1 or less) of the thin portion data. j is 1
The replicas 110_ij to 1r0_ij of integers equal to or greater than m and equal to or less than disk devices 2_i (k is 1 or more and n
On the following integer).

Next, the operation of the first embodiment of the present invention will be described with reference to FIGS.

The request receiving means 6_1 to 6_n are connected to the LAN 4
A processing request is accepted via. Upon receiving the processing request, the request receiving means 6_k (where k is an integer of 1 or more and n or less) stores the original in the disk device 2_i (where i is an integer of 1 or more and n or less). It generates n partial processing requests 200_1 to 200_n requesting an executable part using only the stored small partial data 100_i1 to 100_im.

Then, each partial processing request 200_i is sent to the request processing means 7_i via the SAN 3.

The request processing means 7_i performs the sub-part data 100_i1
An input / output request for ~ 100_im is sent to the file system 5_i.

The file system 5_i stores the subpart data 100
When an input / output request for _ij is received, it is passed to the request allocating means 8_i.

The request allocating unit 8_i checks which computer node stores the original data 100_ij and the replicas 110_ij to 1r0_ij of the small-part data, and checks each disk device storing the original data and the replica and each disk. Estimate the operating status of the computer nodes directly connected to the device and the load applied to these disk devices and the computer nodes, and among the operating disk devices connected to the operating computer nodes, the lighter the load Select a disk unit or computer node with priority
Assign an input / output request for 100_ij.

The file system 5_i is a computer node 2_i
And executes the input / output request assigned to the request processing unit 7_i.

Also, a request is made to the file system 5_h of the computer node 2_h via the SAN 3 for processing of an input / output request assigned to another computer node 2_h (where h is an integer other than i and an integer from 1 to n). Upon receiving the result, the request processing means 7
Return to _i.

A request for an input / output request is received from another arbitrary file system 5 — l (where l is an integer other than i and an integer from 1 to n) via the SAN 3, the request is executed, and the result is sent to the file system 5 — 1. return.

The request processing means 7_i obtains all the results of the necessary file input / output, and when the partial processing is completed, returns the result to the request receiving means 6_k.

The request accepting means 6_k includes request processing means 7_1 to 7_
When the partial processing results are received from all of n, they are combined into one processing result and returned to the sender of the processing request via the LAN 4.

Next, the operation and effect of the first embodiment of the present invention will be described.

In the first embodiment of the present invention, when the load of the disk device 2_i is higher than that of the other disk devices, the request allocating means 8_1 to 8_n sets the disk device 2_i
Decide to reduce the percentage of use of all the sub-part data above, so that if an unexpected I / O pattern causes the load on a particular disk unit to increase significantly, that load can be reduced. By equalizing the load on each disk device, the processing performance of the entire system can be improved.

The second operation and effect of the first embodiment of the present invention is that, when the disk device 2_i is stopped, the request allocating means 8_1 to 8_n replaces all the thin data on the disk device 2_i. On multiple other disk drives,
Decide to use the duplicate or original. As a result, even if the disk device 2_i is stopped for an unexpected reason such as a failure or for a predetermined reason such as a maintenance purpose, the processing can be continued.

The third operation and effect of the first embodiment of the present invention is that a copy corresponding to a plurality of originals stored in a certain disk device is stored in different disk devices, respectively. By redirecting an I / O request to another disk device that stores a corresponding copy, the load on one disk device can be spread to a plurality of disk devices. As a result, the load can be spread to many disk devices in a small number of steps.

[0091]

Embodiment 1 Next, the first embodiment of the present invention will be described in more detail with reference to a specific embodiment. Hereinafter, an embodiment in which the present invention is applied to a full-text search server system will be described. The basic configuration of one embodiment of the present invention is the same as the configuration of the embodiment described with reference to FIG.

Here, the full-text search is a process of searching for a document including a given character string group from a given document group and obtaining a list of the documents. Here, full-text search is realized using an index file. An index file is a file in which a list of all documents in a document group that includes the word that appears in the document group can be extracted from any word that appears in the document group. A list is obtained, and a data structure in which a pair of a document list corresponding to each word is associated with a tree structure or a hash table is stored in a file.

In the present embodiment, a sub-index created by dividing an index file is provided as the data corresponding to the small part data.

It is determined that the sub-index of the index file of a certain document group is an index of a partial document group created by dividing the document group.

FIG. 3 is a diagram showing an example of the configuration of a full-text search server according to an embodiment of the present invention. Referring to FIG. 3, eight computer nodes are provided, and disk devices 2_i (where i is an integer of 1 or more and 8 or less) include four sub-index originals 500_i1 to 500_i4 and a sub-index copy 600_jk ( Here, k is an integer of 1 or more and 4 or less, j is an integer, and j = (ik) mod8) is stored.

The request allocating means 8_i includes an input / output request allocating table 9_i for recording the input / output requests allocated to each computer node, and a state monitoring means 10_i for monitoring the operating state of each computer node.

FIG. 4 is a schematic diagram for explaining the procedure for configuring the original sub-index 500_ij and the replica 600_ij in one embodiment of the present invention.

There are 80,000 documents to be searched, and document IDs 00001 to 80,000 are assigned to them. This is divided into eight pieces, each of which is 10,000 in order from the top, and is assigned to the disk devices 2_1 to 2_8 one by one.

Further, the disk device 2_i (where i is 1 or more and 8
The document group assigned to the following integer) is divided into four groups of 2500 documents, and sub-indexes are created for each
i1 to 500_i4.

Further, the sub index 500_ij (where j is 1
A copy of the above (4 or less integer) is created and set to 600_ij.

FIG. 5 is a schematic diagram showing the arrangement of the sub-index original 500_ij and the replica 600_ij on the disk devices 2_1 to 2_8. Referring to FIG. 5, originals 500_i1 to 500_i4
(Where i is an integer from 1 to 8)
Put on 2_i.

When h = (i + j) mod 8, the replica 600_ih
Is placed in the disk device 2_h.

FIG. 6 is a diagram showing an example of the configuration of the input / output request assignment table 9_1 in one embodiment of the present invention. The input / output request assignment table 9_i includes an input / output request queue 30_1 for the disk device 2_i having the original data 500_i1 to 500_i4, and the data 500_i1 to 500_i4.
Disk unit 2 with duplicates 600_i1 to 600_i4 of _i1 to 500_i4
_h ~ 2_k (where h = (i + 1) mod 8, k = (i + 4) mod
8) The input / output request queues 30_2 to 30_5 for each of the above are provided.

FIG. 7 is a sequence diagram for explaining the operation of one embodiment of the present invention.
FIG. 9 is a diagram for explaining the flow of processing when the request arrives at the first request receiving means 6_1.

Referring to FIG. 7, the request receiving means 6_1
The query is copied as it is and input to the request processing means 7_1 to 7_8 on all the computer nodes.

Each request processing means 7_i (1 ≦ i ≦ 8) issues a read request for the sub-indexes 500_i1 to 500_i4 to the file system 5_i in order to perform a search.

The file system 5_i requests the request allocating means 8_i to allocate a computer node that should process a read request to the subindex 500_ij.

The request allocating means 8_i connects each read request to one of the queues 30_i to 30_ (i + 4).

[0109] If there is a read request connected to the queue 30_i, the file system 5_i sequentially retrieves the read request from the head and executes reading from the disk device 2_i. Also, the queue 30_ (i + q) (where q is 1 or more and 4 If there is a read request connected to any of the following integers), the file system 5_
(i + q) via the SAN3.

[0110] The file system 5_ (i + q) reads the requested data from the disk 2_ (i + q) device and returns it to the file system 5_i via the SAN 3.

The file system 5_i is the disk device 2_i
Alternatively, it returns the read result received from the file system 5_ (i + q) to the request processing means 7_i.

The request processing means 7_i transmits the result of the read request to the disk device 2_i or the file system 5_
When obtained from (i + q), the read request is removed from the queue.

In this way, the request processing means 7_i creates a corresponding document ID list for the query using the originals or copies of the sub-indexes 500_i1 to 500_i4, and returns this to the request receiving means 6_1.

The request receiving means 6_1 is composed of request processing means 7_1 to 7_
The document ID list is received from 8, and these are combined into one and returned to the query issuer.

FIG. 8 is a flow chart for explaining the operation of the request allocating means 8_i in one embodiment of the present invention. FIG.
The operation of the request assignment means 8_i will be described with reference to FIG.

The request allocating means 8 — i stores the subpart data 500
When a request to allocate a read request to _ij is received, it is checked which disk device has a copy 600_ij of 500_ij (step S101). Here, it is assumed that it is in the disk device 2_h.

Next, referring to the operating state monitoring means 10_i,
It is checked whether the disk device 2_h is operating (step S102).

If stopped, then the disk device 2_i
It is checked whether or not is running (step S108).

If this is also stopped, the disk device
2_i notifies the failure of the read processing and ends (step S109).

On the other hand, in step S108, the disk device 2_i
Is running, a read request is connected to the queue 30_i, and the process ends.

If it is determined in step S102 that the disk 2_h is operating, it is checked whether the disk device 2_i is operating (step S103). If stopped, a read request is connected to the queue 30_h (step S
107), end.

If the disk device 2_i is operating in step S103, the queues 30_i and 30_i are referred to by referring to the allocation table 9_i.
The length of h is checked (step S104).

Even if the queue 30_i is connected to the queue 30_h, if the queue 30_i is still more than four times longer than the queue 30_h, the queue 30_i is connected to the queue 30_h (step S107). The connection is made (step S106), and the process ends.

Next, the operation of the request allocating means 8_i will be described in detail with reference to a specific example.

In FIG. 6, the queue 30_1 already has input / output requests a, b, and c, and the queue 30_4 has d. Here, when the input / output request 40_1 for 500_11 is input, the original of 500_11 is in the disk device 2_1 and its copy 600_11 is in the disk device 2_2, and comparing the queues for both disk devices, the queue 30_1 has Still three
Since there are only requests, the queue 30_2 is empty, but the input / output request 40_1 is decided to be put in the queue 30_1.

Next, when an input / output request 40_2 for 500_12 arrives, the original of 500_12 is copied to the disk device 2_1 and the copy 600
_12 is in disk unit 2_3, and comparing the queues for both disks, queue 30_1 contains four requests and queue 30_3 is empty, so I / O request 40_2 is queue
Decide to put in 30_3.

Next, when an input / output request 40_3 for the index 500_13 arrives, the original of 500_13 is stored in the disk device 2_1.
In addition, since the copy 600_13 is in the disk device 2_4, comparing the queues corresponding to both disks, the queue 30_1 contains four requests, and the queue 30_4 also contains one request. Request 40_3 decides to enter queue 30_1.

Next, it is assumed that before the input / output request 40_4 for the sub-index 500_14 is determined, the processing on the disk 2_1 proceeds, and the input / output requests a and b are removed from the queue 30_1.

The original of 500_14 is on disk 2_1 and the copy is on disk 2_5. When the queues for both disks are compared, there are only three requests for queue 30_1.
_5 is empty, but request 40_4 decides to queue 30_4.

In this example, whether to use the original or to use the copy is determined based on the ratio of the number of requests in the queue being four times or more, but the ratio used as a criterion is an arbitrary numerical value. Can be used. The optimal value depends on the cost of accessing the original and the cost of accessing the duplicate.

Also, the difference may be determined not by the ratio but by the difference between the numbers being equal to or more than a certain value.

Further, it can be determined based on the recent response time of each disk or the magnitude of the accumulated access time,
As soon as each disk has completed one I / O request, the oldest request for data on that disk may be assigned.

[0133]

[Embodiment 2] Next, a second embodiment of the present invention will be described. FIG. 9 is a diagram showing a configuration of the second exemplary embodiment of the present invention. Referring to FIG. 9, the second embodiment of the present invention is different from the first embodiment shown in FIG. 1 in that the request allocating units 8_1 to 8_n are replaced with request accepting units 6__.
1 to 6_n are provided with second request allocating means 21_1 to 21_n.

Upon receiving the processing request from the client via the LAN 4, the request receiving means 6_i receives the different sub-data 50
Create n × m partial processing requests that require 0_ij (where i is an integer between 1 and n, and j is an integer between 1 and m).

Then, it requests the second request allocating means 21_i to determine which request processing means to send each partial processing request to. The second request allocating means 21_i stores
The partial processing request that requires 500_ij is input to the request processing unit 7_i having the original on the disk device 2_i, or the request processing unit 7 having the duplicate 600_ij on the disk device 2_h.
_h (where h is an integer from 1 to n different from i) is determined.

The request receiving means 6_i sends a partial processing request to the request processing means according to the determination of the request allocating means 21_i.

The request processing means 7_k is the file system 5_i
To access the original or the copy of the thin part data on the disk 5_i to execute the partial processing request, and return the result to the request receiving means 6_i.

The request receiving means executes the processing request using the results of the n partial processing requests, and returns the result to the client.

Next, the function and effect of the second embodiment of the present invention will be described.

In the second embodiment of the present invention, the load of the computer node 1_i (where i is an integer of 1 or more and n or less) having the original 500_ij of the small part data in the disk device 2_i is smaller than that of the other computer nodes. The second request allocation means 21_i
Sends a partial processing request that requires the thin partial data 500_ij to another computer node 1_h that has a copy 600_ij in the disk device 2_h.
Assign to For this reason, even if the load on a specific computer node becomes prominently high, the load can be reduced.

The second operation and effect of the second embodiment of the present invention is that when the disk device 2_i or the computer node 1_i is stopped, the second request allocating means 21_1 to 21_n replaces the computer node 1_i. It is decided to use another plurality of computer nodes having another disk device having a copy of the thin portion data stored in the disk device 2_i.

Thus, even if the disk is stopped for unexpected reasons such as a failure or for a predetermined reason such as a maintenance purpose, the processing can be continued.

The third operation and effect of the second embodiment of the present invention is that a copy corresponding to a plurality of originals stored in a certain disk device is placed in different disk devices, respectively. By allocating a partial processing request requiring an original to a computer node having another disk device that stores a corresponding copy, the load of one computer node can be spread to a plurality of computer nodes. Thus, the load can be spread to many computer nodes in a small number of steps.

[0144]

Embodiment 2 Next, the operation of the second embodiment of the present invention will be described with reference to a specific embodiment.

FIG. 10 is a diagram showing the configuration of a full-text search server system according to a second embodiment of the present invention. The difference from the configuration of the embodiment shown in FIG. 3 is that the second request allocating units 21_1 to 21_8 are each provided with one request receiving unit 6_1.
6_8, and a ninth computer node 1_9, which is connected to the LAN 4.

The second request allocating means 21_1 to 21_8 include second state monitoring means 22_1 to 22_8, respectively.
The computer node 1_9 includes a load monitoring unit 23 and a load distribution recalculation unit 24.

Each of the second request allocating means 21_1 to 21_8 stores a copy use target ratio for each data 500_ij, and allocates a partial processing request to the original and the copy of each data so as to satisfy this ratio. .

The request processing means 7_1 to 7_8 each time one partial processing request is processed, measures the actual time required for the request, and sends the measured time to the load monitoring means 12 via the LAN 4.

The load monitoring means 12 records, as a search load, the accumulated value of the actual time required for processing for each request processing means,
At regular time intervals, the values of these variables are sent to the load distribution recalculating means 13 to return the accumulated value to zero and continue recording.

The load distribution recalculating means 13 is provided with the request processing means 7_
Every time a search load of 1 to 7_8 is received for a predetermined period of time, a new replication usage target ratio is determined from the replication usage target ratio of each data 500_ij, and all the second request allocation means 21_1 to
Send to 21_8.

The second request allocating means, upon receiving the new copy use target ratio, stores it and thereafter allocates the partial processing request using the value.

FIG. 11 is a flow chart for explaining the processing of the second request allocating means 21_i in the second embodiment of the present invention. Referring to FIG. 11, first, the small part data 500_ij
It is checked in which disk device the original and its copy 600_ij are located (step S201). Here, the original is a disc
It is assumed that a copy exists in the disk device 2_h at 2_i.

Next, it is determined whether or not the computer nodes 1_i and 1_h are operating with reference to the state monitoring means 22_i (steps S202 and S208). If neither of them is operating, a notification of allocation failure is given and the process ends (step S20).
9).

If only the computer node 1_i is operating,
Allocate a partial processing request to the computer node 1_i (step S
206), if only the computer node 1_h is operating, assign a partial processing request to the computer node 1_h (step S20)
7).

If both are operating, the random number is used (step S204), and the probability of selecting the computer node 1_h is 500_ij
Computer node 1_h so that the replication specification target ratio of
(Step S207) Alternatively, a partial processing request is assigned to the computer node 1_i.

Next, the procedure in which the load distribution recalculating means 13 determines a new replication specification target ratio will be described.

The new target ratio R_ij relating to the small-part data 500_ij is obtained by setting the search load L_i of the request processing means 7_i on the computer node 1_i having the disk 2_i for recording the original 500_ij to L_i and the disk 2_h for recording the replica 600_ij. L_h is the search load of the request processing means 7_h on the computer node having the latest fixed time, M_ih is the average of L_i and L_h, and R is the old target ratio.
Assuming _ij_old, the new target ratio R is determined as follows.

R_ij = R_ij_old + (1−R_ij_old) × (L
_i-M_ih) / L_i; if L_i> = M_ih

R_ij = R_ij_old * L_i / M_ih; i
f L_i <M_ih

As a result, if the load on the original side is high, the ratio of using the copy is set higher, and if the load on the copy side is high, the ratio of using the copy is set lower.

In this embodiment, a random number is generated and the ratio is stochastically satisfied in order to satisfy the copy use target ratio, but the method of the present invention is not limited to this.
For example, it is possible to record the number of times each of the original and the copy has been used, determine the copy usage ratio, and choose to use the copy if the ratio is below the target ratio, otherwise use the original. Good.

In this embodiment, the new target ratio is
Although the calculation is performed using a specific formula, in the present invention, a method of setting a new target ratio is not limited to this specific formula.

[0163]

As described above, according to the present invention, the following effects can be obtained.

A first effect of the present invention is that access to a relatively wide range of data that can be accommodated in the coarsely divided data can be performed in many cases with at most one remote disk access.

The reason is that, in the present invention, data is arranged so that all the subdivided data derived from the same coarsely divided data are stored in one disk device.
This is because the disk device is operating, and unless the load is particularly high, access to the disk device alone is sufficient.

A second effect of the present invention is that the load imbalance of a disk device or a computer node can be reduced.

The reason for this is that, in the present invention, by determining whether to access the original file or the copy of the file at the time of execution, the load on a disk or computer node with a high load is reduced to a plurality of other disk devices or computer nodes. This is because it can diffuse.

The third effect of the present invention is that the processing can be continued even when some of the disk devices or computer nodes constituting the system are stopped. In the above, when a certain disk device or a computer node is stopped, execution is performed such that a corresponding copy is used instead of the original of the fine part data recorded on the disk device or the disk device of the computer node. This is because the switching control is sometimes performed.

A fourth effect of the present invention is that the load of a certain disk device or computer node can be spread to and replaced by many other disk devices or computer nodes in a small number of steps.

The reason is that, in the present invention, a copy of data stored in one disk device is distributed and stored in a plurality of disk devices. A copy on a disk device can be used, and a plurality of computer nodes having a disk device storing the corresponding copy can be used instead of a computer node having a disk device storing one original. Because you can.

[Brief description of the drawings]

FIG. 1 is a diagram showing a configuration of a first exemplary embodiment of the present invention.

FIG. 2 is a schematic diagram showing a procedure for creating an original and a copy of fine part data according to the first embodiment of the present invention.

FIG. 3 is a diagram showing a configuration of a full-text search system according to a first embodiment of the present invention.

FIG. 4 is a diagram showing a procedure for creating an original and a copy of a sub-index according to the first embodiment of the present invention.

FIG. 5 is a diagram showing an arrangement of sub-index originals and duplicates on a disk in the first embodiment of the present invention.

FIG. 6 is a diagram showing a specific example of the operation of the request assignment unit in the first embodiment of the present invention.

FIG. 7 is a sequence diagram showing an operation of the first exemplary embodiment of the present invention.

FIG. 8 is a flowchart illustrating a process performed by a request allocating unit according to the first embodiment of this invention.

FIG. 9 is a diagram showing a configuration of a second exemplary embodiment of the present invention.

FIG. 10 is a diagram showing a configuration of a full-text search system according to a second embodiment of the present invention.

FIG. 11 is a flowchart illustrating a process of a second request allocating unit according to the second embodiment of this invention.

[Explanation of symbols]

1_1 to 1_n Computer node 2_1 to 2_n Disk 3 SAN (System Area Network, network in system) 4 LAN (Local Area Network, local network) 5_1 to 5_n File system 6_1 to 6_n Request receiving means 7_1 to 7_n Request processing means 8_1 to 8_n Request allocating means 9_1 ~ 9_n Request allocating table 10_1 ~ 10_n Operating status monitoring means 21_1 ~ 21_n Second request allocating means 22_1 ~ 22_n Second operating status related means 23 Load status monitoring means 24 Load distribution recalculating means 30_1 ~ 30_5 Output request queue 40_1 ~ 40_4 I / O request 100_11 ~ 100_nm Original data of thin partial data 1r0_11 ~ 1r0_nm Duplication of thin partial data 200_1 ~ 200_n Partial processing request 500_11 ~ 500_84 Original of sub index 600_11 ~ 600_84 Duplication of sub index

Claims (13)

[Claims] 1. A non-shared cluster system comprising a plurality of computer nodes, wherein each of the computer nodes has a local unique data storage device, and performs processing depending on data on the data storage device. Server
A coarsely divided data obtained by roughly dividing data to be stored in the data storage device, wherein the coarsely divided data is further divided to generate finely divided data; All of the subdivided data to be copied are stored in the same data storage device, and copies of all the subdivided data derived from the same coarsely divided data are copied to a data storage device different from the data storage device in which the original is stored. And a means for storing the data in a plurality of different data storage devices. 2. The server system according to claim 1, further comprising: a data storage device for storing an original of said subdivided data; and a data storage device for storing a copy of said subdivided data. A server system comprising an input / output request allocating means for allocating the input / output request to be executed using any one of the above. 3. The server system according to claim 1, wherein a request for a process requiring certain subdivided data is sent to a computer node having a data storage device for storing an original of the subdivided data; And a processing request allocating means for allocating the processing request to a computer node having a data storage device for storing a copy of the processing request. 4. The server system according to claim 2, wherein a load on a data storage device for recording an original of a certain subdivided data and a means for estimating a load on a data storage device for recording a copy of the subdivided data, Request allocating means for distributing an input / output request for the subdivided data so as to execute an input / output for the subdivided data using a data storage device estimated to have a lighter load by the load estimating unit; And a server system comprising: 5. The server system according to claim 3, further comprising: a load on a computer node having a data storage device for recording an original of certain subdivided data; and a data storage device for recording a copy of said subdivided data. Means for estimating the load on the computer node; and means for performing the processing requiring the subdivision data by using the computer node estimated to have a lighter load by the means for estimating the load. A server system, comprising: request allocation means for allocating a processing request requiring data. 6. The server system according to claim 2, further comprising: means for monitoring an operation state of each of the computer nodes; a computer node having a data storage device for recording an original of certain subdivided data; Among the computer nodes having the data storage device that records the data copy, the operating computer node is used to distribute the input / output request for the subdivided data so as to execute the input / output to the subdivided data. A server system comprising request assignment means. 7. The server system according to claim 3, further comprising: means for monitoring an operation state of each of the computer nodes; a computer node having a data storage device for recording an original of certain subdivided data; A processing request that requires the subdivided data so as to execute a process that requires the subdivided data using an operating computer node among computer nodes having a data storage device that records a copy of data. A server system comprising request allocation means for distribution. 8. A computer system comprising: a plurality of computer nodes connected to each other to form a cluster; each of the plurality of computer nodes having a local data storage device and receiving an input / output request for a file to receive the data; A file system for inputting / outputting data on a storage device, a request receiving unit for receiving a request from a client, and a request processing unit, wherein the request receiving unit of each of the computer nodes includes: Receiving a processing request requiring data, dividing the processing request into a plurality of partial processing requests, sending these partial processing requests to any of the request processing means of a plurality of computer nodes, and Is integrated into one processing result returned as a response to The request processing means creates a partial processing result in response to the partial processing request from the request receiving means, and the file system of each of the computer nodes transmits an input / output request for a file to an original file and one or more files. A request allocating means for determining which of the duplicates is to be processed; first, the data to be stored in the data storage device is roughly divided, and this is further subdivided into the same coarsely divided data; A cluster system, wherein all the subdivided data derived from the above are stored in the same data storage device, and a copy of each of the subdivided data is stored in a different data storage device from the original. 9. An input / output request assignment table for recording the input / output requests assigned to each of the computer nodes, and a status monitoring means for monitoring the operating status of each of the computer nodes. Distribute the processing request that requires each of the subdivided data to a computer node equipped with a data storage device that stores the original and a computer node equipped with a data storage device that stores a copy based on the respective load conditions, 9. The server system according to claim 8, wherein: 10. A computer system comprising: a plurality of computer nodes connected to each other to form a cluster; each of the plurality of computer nodes includes a local data storage device, and receives a data input / output request for a file; A file system for inputting / outputting data on the storage device, a request receiving unit for receiving a request from a client, and a request processing unit, wherein the data to be stored in the data storage device is first roughly divided, This is subdivided, and all the subdivided data derived from the same coarsely divided data are stored in the same data storage device, and a copy of each of the subdivided data is separated from the original by a different data storage device. The request receiving means comprises a request allocating means, wherein the request receiving means receives a processing request from a client. Upon receiving the request, the request allocating unit requests the request allocating unit to create a plurality of partial processing requests that require different pieces of partial data, and determine which computer node to send the respective partial processing requests to the request processing unit. The request allocating means inputs a partial processing request requiring the fine partial data to one request processing means having the original data on the data storage device, or another request having a duplicate thereof on the data storage device. The request accepting means sends a partial processing request to the request processing means in accordance with the determination of the request allocating means, and the request processing means stores the data through the file system. A partial processing request is executed by accessing an original or a copy of the fine partial data on the device, and the result is returned to the request receiving means. Server system of using the result to the execution of the processing request, and returns the results to the client, and wherein the. 11. One of the plurality of computer nodes includes a load monitoring unit and a load distribution recalculation unit, and the request allocating units of the other computer nodes respectively determine an operation state of the own node. The request allocating means includes a duplication use target ratio for each data, and allocates a partial processing request to an original and a copy of each data so as to satisfy this ratio. The request processing means, each time one partial processing request is processed, measures the actual time required for the partial processing, sends the measured time to the load monitoring means, and the load monitoring means The cumulative value of the real time required for processing is recorded as a search load, and at regular time intervals, the values of these variables are sent to the load distribution recalculating means, and the cumulative value is reset. The load distribution recalculating means determines a new replication usage target ratio from the replication usage target ratio of each data every time receiving the latest search load of the request processing means, and 11. The method according to claim 10, wherein the request allocating means receives the new replication use target ratio, stores the received target usage ratio, and allocates a partial processing request using the value. Server system. 12. A data storage method for a non-shared cluster system including a plurality of computer nodes, wherein each of said computer nodes has a locally unique data storage device. First, data to be stored in a data storage device is roughly stored. And further subdivides the coarsely divided data, stores all the subdivided data derived from the same coarsely divided data in the same data storage device, and copies each of the subdivided data separately from the original. And storing the data in different data storage devices. 13. A method of processing data stored in a data storage device by the data storage method of a non-shared cluster system according to claim 12, wherein a processing request requiring each of the subdivided data is stored as an original. A load sharing method for a non-shared cluster system, wherein the load is distributed to a computer node having a data storage device and a computer node having a data storage device that stores a copy of the data storage device, based on respective load conditions.