JP2001101149A - Distributed parallel data processor, recording medium recording distributed parallel data processing program and distributed parallel data processing system - Google Patents

Abstract

PROBLEM TO BE SOLVED: To maximize the throughput of the whole system by uniforming the loads of respective computers in the parallel data processing for mainly executing reading processing from a secondary storage device and to prevent the sharp decline of performance by uniforming a subrogated load by remaining computers even when a part of computers are stopped. SOLUTION: Large-scale data are divided, distributed and stored in respective computers, redundantly storing divided data so as not to have the copies of the divided data overlapped to execute processing for the same divided data by different computers. Processing time including reading time from the secondary storage device is measured, a distribution ration is calculated so that the number of requests for the processing of a computer having long processing time is reduced and the number of requests for the processing of a computer having short processing time is increased to subrogate the processing. When a part of the computers is stopped, the distribution ratio of the stopped computer is fixed to 0 and the distribution ratio is recalculated, so that subrogated processing is prevented from being concentrated to a part of the computers by distributing the loads.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a distributed parallel data processing device and a distributed parallel data processing system, in which a plurality of computers independently process a processing request from a user, and process results of the respective computers are integrated and returned to the user. Storage medium storing a distributed data processing program and a distributed parallel data processing system (hereinafter, simply referred to as a distributed parallel data processing device), particularly based on the time of input / output processing of a secondary storage device of each computer. The present invention also relates to a distributed parallel data processing device for equalizing the load on each computer.

[0002]

2. Description of the Related Art Conventionally, when a computer performs processing such as search, aggregation, and processing of large-scale data, input / output processing such as reading and writing of data from a secondary storage device becomes a bottleneck. For this reason, a method of dividing data, distributing and storing the divided data in a plurality of secondary storage devices, and performing input / output processing in parallel is generally performed.

[0003] In a distributed parallel data processing device composed of a plurality of independent computers, processing requests from a user, such as retrieval, aggregation, and processing, are stored in a secondary storage device of each computer. The processing can be speeded up by reading the divided data in parallel and independently.

[0004] Examples of an application that performs parallel processing on large-scale data include a full-text search system that returns a list of documents including a keyword specified by a user from a huge document group, and an online transaction on a database management system. There are a processing server and a multimedia stream server that supplies audio / video data.

For example, in a search process for extracting a record satisfying specified conditions from a large-scale database,
Processing can be easily parallelized by dividing and storing the database in a distributed manner. The processing flow in this case is as follows. The user sends the conditional expression to the computer of the receiving role. The computer as the receiving unit creates a partial process specifying the divided data to be processed from the conditional expression, and transmits the partial process to all the computers having the corresponding divided data. Each computer reads its own divided data independently, and transmits a record that satisfies the condition to the accepting computer as a partial result.
The accepting computer collects, integrates, edits all partial results, and replies to the user.

As described above, even if the original document group is large-scale, each computer can perform processing independently and in parallel, so that high-speed response is possible. Furthermore, by creating an auxiliary file in which records corresponding to each condition are indexed in advance, unnecessary data reading can be omitted, and the search processing can be further speeded up.

In such a distributed parallel type data processing apparatus, if the load among the computers is not uniform, there are always busy computers during the processing, and free computers that spend idle time waiting for the processing appear. Therefore, there is a problem that the overall processing efficiency is reduced.

For this reason, many proposals have been made since ancient times to equalize the load on each computer.

For example, as a first prior art, Patent No. 2
As in the patent described in Japanese Patent Publication No. 638065, there is a load distribution method that focuses on CPU utilization. According to this method, first, for all the tasks that can be executed by the CPU, the ratio between the CPU operation and the disk I / O operation is obtained in advance. Next, during the operation of the system, measure the utilization of each CPU, request a task with a high percentage of disk I / O operation to a CPU with a high utilization, and a high percentage of CPU operation to a CPU with a low utilization By requesting tasks, the load is equalized.

As another second conventional technique, there is a method of distributing load by equalizing the amount of divided data to be distributed and stored, as in the invention disclosed in Japanese Patent Application Laid-Open No. 6-271227, for example. According to this method, when the database is distributed and stored in each computer, the database is divided so that the number of records for each item is equal. Therefore, when performing a search process, each computer reads all the divided data owned by itself and extracts records that meet the conditions, so that the load on each computer can be equalized.

In the distributed parallel data processing apparatus, a copy of each divided data is created and redundantly stored, so that the duplicated data is owned even when some computers are stopped due to a failure or the like. Another major feature is that services can be continued without interruption by other computers taking over the processing.

For example, as a third prior art, a method of realizing a non-stop service in a distributed parallel data processing device is described in an example related to a video stream server in March 1998.
Moon, IBM Journal of Research and Development, Vol. 42, No. 2, pp. 187-189 (IB
M J. Res.Develop., Vol.42, No.2, p187-189, March, 1
998). According to this document, the divided data stored in advance in each computer is further finely divided, and a copy of the subdivided data is redundantly stored in a different computer. As a result, even if a failure occurs in a certain computer or a secondary storage device, the computer having the duplicate subdivided data can take over the processing, so that the service can be continued without interruption.

At this time, since the original divided data stored in one computer is subdivided and stored in a plurality of computers, the increase in the load due to the shoulder replacement processing increases the load of the plurality of computers. It is possible to distribute among them.
In particular, in a video stream server, the process of reading a file in order from the beginning to the end is the center. Therefore, if the file is subdivided with an equal capacity, the load due to the shoulder replacement process is also uniform.

[0014]

However, in the above-mentioned conventional distributed parallel data processing apparatus, the following problems exist when processing such as retrieval or totalization of large-scale data is performed.

The first problem is that, in the processing of large-scale data, the processing time of each computer is not necessarily equalized even if the load distribution control is performed according to the CPU utilization rate as in the first related art. That is.

The reason is that most of the processing time for large-scale data depends on the reading time from the secondary storage device rather than the CPU processing time.

The second problem is that when data is indexed in order to speed up the search processing, the second problem occurs.
Even if the amount of data to be distributed and stored is equalized as in the related art, the processing time of each computer is not necessarily equalized.

The reason is that when the data is indexed, the amount of data to be read differs depending on the number of data that satisfies the condition, so that the processing time of each computer varies.

If the number of records corresponding to each condition in the index is distributed and stored so as to be equal in each computer, the same record may be registered in a different index. Therefore, when performing complex conditions, it is necessary to exchange a large amount of partial results between computers, and even if the amount of data to be read is made uniform,
The communication processing increases, and conversely, the efficiency of parallelization greatly decreases.

A third problem is that it is difficult to cope with a change in the distribution of access frequencies to each data only by distributing and storing the data as in the third related art. The reason is that if the distribution of access frequency changes,
This is because in the third conventional technique, it is necessary to change the method of dividing and distributing data.

The fourth problem is that, when each computer reads out a different amount of data and performs processing, if a failure occurs in the computer or the secondary storage device, the reading amount is made constant as in the third related art. In the load distribution method, the performance is degraded due to the imbalance in the substitute load. The reason is the same as the second problem.

A main object of the present invention is to provide a distributed parallel type that can always perform load distribution both in normal times and when a failure occurs, even when performing processing such as searching and totalizing indexed large-scale data. A data processing device is provided. For that purpose, the following four objectives are realized.

A first object is to equalize the load between the computers and maximize the throughput of the entire system even when the processing time of each computer depends on the input / output processing from the secondary storage device. It is.

A second object is to equalize the overall processing time of each computer and maximize the throughput of the entire system, even if the processing involves different amounts of data to be read by each computer.

A third object is to automatically equalize the entire processing time of each computer even if the distribution of the access frequency to each data changes.

A fourth object of the present invention is to prevent a decrease in performance by equalizing the load even when some computers or secondary storage devices are stopped.

[0027]

A first distributed parallel data processing apparatus according to the present invention divides large-scale data, stores the divided data in each computer, and creates a copy of the divided data to prevent duplication. In this way, the data is redundantly stored, and partial processing on the same divided data can be performed by a plurality of different computers. Therefore, the load on a certain computer increases,
When a computer stops, another computer that owns the same divided data as that of the computer can take over the processing.

A load monitoring means for measuring the processing time of each divided data including the input / output processing of the secondary storage device, and a processing time based on the processing time for each divided data measured by the load monitoring means. Distribution ratio calculation means for calculating a distribution ratio so as to reduce the number of requests for partial processing to a computer with a long time and increase the number of requests for partial processing to a computer with a short processing time, and distribution obtained by the distribution ratio calculation means It has a load balancing means for determining the requested computer according to the ratio.

By adopting such a configuration, the processing time including the input / output processing of the secondary storage device is used as an index of load distribution, and the distribution ratio for equalizing the processing time in all the computers is obtained. The first and second objects of the invention can be achieved.

Further, the distribution ratio calculating means of the first distributed parallel data processing apparatus of the present invention may be configured to start at regular intervals and calculate a distribution ratio for equalizing the load. Further, the second distributed parallel data processing device of the present invention, in addition to the configuration of the first distributed parallel data processing device of the present invention, detects a bias in the load of each computer and activates the distribution ratio calculating means. It has calculation time control means.

The third object of the present invention is achieved by adopting such a configuration and recalculating the distribution ratio for equalizing the load periodically or when the load on each computer becomes uneven. Can be. Further, the first distributed parallel data processing apparatus of the present invention has a failure detecting means for detecting a stopped computer, and a request processing managing means for holding processing requested to each computer.

By adopting such a configuration, when some computers stop, the distribution ratio to the stopped computers is fixed to 0, the distribution ratio is recalculated, and the load is distributed by the configuration excluding the stopped computers. By doing so, the fourth object of the present invention can be achieved.

[0033]

Next, embodiments of the present invention will be described in detail with reference to the drawings. [First Embodiment] Referring to FIG. 1, in a first embodiment of the present invention, n independent computers 1_1 to 1_n connected by an external communication network 3 (n is an integer of 1 or more. n is the number of computers.), and the secondary storage devices 2_1 to 2_n independently owned by the computers 1_1 to 1_n.

Each secondary storage device stores m divided data obtained by dividing large-scale data into m × n pieces (m is an integer of 1 or more; m is referred to as a granularity). . Also,
R copies of each divided data (r is an integer of 1 or more; r is referred to as redundancy) are created and redundantly stored in the same secondary storage device so as not to be duplicated. Here, for convenience, data created by dividing large-scale data into m × n pieces is referred to as original divided data, and divided data created by copying the original divided data is referred to as duplicate divided data.

A method of distributing and redundantly storing divided data will be described with reference to FIG. First, large-scale data
Then, the original data D11 to Dnm are created. Next, the original divided data D11 to D1m are stored in the secondary storage device 2_1, and the original divided data D21 to D2m are stored in the secondary storage device 2_2. Similarly, the original divided data Di1 to Dim (i is an integer of 1 or more and n or less) are distributed and stored in the secondary storage device 2_i.

Next, a duplicate of the divided data D11 to Dnm is created, the duplicate divided data D11 is stored in the secondary storage device 2_2, and the duplicate divided data D12 is stored in the secondary storage device 2_3. D1j (j is an integer from 1 to m) is stored in the secondary storage device 2_j + 1. Further, the copy division data D21 is stored in the secondary storage device 2_3, the copy division data D22 is stored in the secondary storage device 2_4, and similarly, the copy division data D2j is stored in the secondary storage device 2_j + 2. Similarly, all the duplicate divided data Dij are redundantly stored in the secondary storage device 2_ (i + j).
However, when the subscript (i + j) exceeds n, the remainder obtained by dividing (i + j) by n is used as the subscript, and the storage destination secondary storage device is circulated.

When the redundancy r is 2, the duplicate divided data Dij is further shifted one by one and stored in the secondary storage device 2_ (i + j + 1). Also in this case, if (i + j + 1) exceeds n, (i + j + 1)
The remainder obtained by dividing +1) by n is used as a subscript and stored cyclically.

In general, in the case of the redundancy r, the copied r pieces of divided data Dij are stored in the secondary storage device 2_ (i + j + s) (s is 0 to r−
By storing them in the same secondary storage device, respectively, the divided data can be redundantly stored so as not to be duplicated. However, if (i + j + s) exceeds n, (i + j + s)
The remainder obtained by dividing by n is used as a subscript and stored in a cyclic manner. The greater the redundancy, the easier the load becomes, and the more computers that can be stopped at the same time.

With the arrangement of the divided data as described above,
Even if one computer stops, another computer can take over the processing. Note that the distribution and redundant storage method is not limited to the method described here as long as the same divided data can be arranged in the same secondary storage device without overlapping.

Referring to FIG. 1, a computer 1_i (i is 1 or more and n
The following integers) are roughly the reception processing unit 11, the load management unit 12,
It includes three control units of the data processing unit 13. Reception processing unit 11
Is a control unit that receives a processing request from a user and responds with a result. The request receiving unit 111, the request distribution unit 112, the divided data management unit 113, the partial result integration unit 114, the failure detection unit 115, the request processing management Means 116 are provided. The load management unit 12 is a control unit that measures the processing time of each computer and equalizes the load, and includes a load monitoring unit 121, a distribution ratio calculation unit 12
2. It has a load balancing means 123. The data processing unit 13
It is a control unit that executes a partial processing request requested by the reception processing unit, and includes a request processing unit 131.

The request receiving means 111 receives a processing request from a user via the external communication network 3 and creates a partial process for m × n divided data.

The request distribution unit 112 requests the request processing unit 131 in the computer designated by the load distribution unit 121 to perform a partial process via the external communication network 3. The divided data management unit 113 manages which secondary storage device stores the m × n original divided data and the r × m × n duplicate divided data by using a divided data management table. For example,
When divided data is arranged as shown in FIG. 2, the divided data management table created by the divided data management unit 113 is as shown in FIG. FIG. 3 shows that the computers that own the divided data D11 are 1_1, 1_2, and 1_3. The partial result integrating unit 114 collects the results of the partial processing on the m × n divided data from the request processing unit 131 of each computer, integrates and edits the results, and responds to the user via the external communication network 3. The failure detection means 115 detects a computer that has stopped due to a failure or the like. The request processing management means 116 stores the unprocessed partial processing among the partial processing requests requested by the computers as a request processing management table as shown in FIG. Referring to FIG. 4, for the different processing requests Q1 and Q2, the computer 1_1 performs partial processing (Q1, D14), (Q1, D1) on the divided data D14, D1m.
m), it can be seen that the partial processing (Q2, D11) for the divided data D11 has been requested and the result has not been replied.

The load monitoring means 121 measures the average processing time of the partial processing for each divided data, including the input / output processing from the secondary storage device. Distribution ratio calculation means 122
Calculates the distribution ratio of the partial processing that makes the processing time of each computer uniform based on the average processing time of the partial processing for each divided data measured by the load distribution unit 121, and calculates the distribution ratio as shown in FIG. Create a ratio table. Referring to FIG. 5, it can be seen that the partial processing for the divided data D11 is requested to the computers 1_1, 1_2, and 1_3 at a ratio of 40%, 20%, and 30%, respectively. The load distribution unit 123 determines the distribution destination of the partial process based on the distribution ratio calculated by the distribution ratio calculation unit 122, and notifies the request distribution unit 112.

The request processing means 131 executes the partial processing requested by the request distribution means 112 and returns a result of the partial processing to the partial result integrating means 114.

Next, the operation of this embodiment will be described in detail with reference to FIGS. The user sends a processing request such as a search or tally to any computer 1 through the external communication network 3.
_a (a is an integer of 1 or more and n or less) is transmitted to the request receiving means 111 (step S11 in FIG. 6). The request accepting unit 111 creates m × n partial processes that specify the divided data to be processed with respect to the accepted processing request, and passes them to the load distribution unit 123 (Step S12 in FIG. 6). The load distribution unit 123 refers to the distribution ratio table created in advance by the distribution ratio calculation unit 122 to determine the requested computer for all of the partial processes, and passes it to the request distribution unit 112. Request distribution means 1
12 requests the request processing means 131 of the designated request destination computer for partial processing via the external communication network 3 (steps S13 and S14 in FIG. 6).

At the same time, the request processing management means 116 registers the requested partial processing in the request processing management table.

The request processing means 131 of the requested computer executes the requested partial processing. If the request processing unit 131 of the requested computer is processing another partial process, the request is connected to a queue and the process is performed sequentially. Alternatively, the request processing unit 131 may simultaneously process a plurality of partial processes in parallel.

Next, the computer requested to perform the partial processing returns the result of the partial processing obtained by the request processing means 131 to the partial result integrating means 114 in the requesting computer 1_a through the external communication network 3 ( Step S15 in FIG. 6). In addition, the processing time required for the partial processing is transmitted to the load monitoring unit 121 in the computer 1_a of the request source. At the same time, the request processing management means 116 in the computer 1_a deletes the replied partial processing from the request processing management table. Partial result integration means 114 of computer 1_a
Accepts and holds the results of the partial processing, integrates and edits all partial results for the m × n divided data, and responds to the user via the external communication network 3 (FIG. 6). Steps S16 and S17).

Here, the flow up to the response of the processing request from the user has been described in a sequential manner, but the steps in FIG. 6 are performed asynchronously and in parallel to reduce the idle time of each computer. It is possible to further improve the efficiency of parallelization.

Next, a method in which the distribution ratio calculating means 122 calculates the distribution ratio of the partial processing in order to equalize the load will be described with reference to FIG.

The average processing time of partial processing for arbitrary divided data Dij (i is an integer of 1 or more and n or less; j is an integer of 1 or more and m or less) is described as T (Dij). Further, for the divided data Dij, a ratio of the number of times of requesting the computer 1_k (k is an integer of 1 or more and n or less) to perform a partial process is described as d (Dij, 1_k). Here, if the divided data stored in a certain computer 1_k is only D11, D12, and D13,
The average processing time E (1_k) of the computer 1_k is approximately E (1_k) = T (D11) d (D11,1_k) + T (D12) d d (D12,1_k) + T (D13)
It can be calculated as d (D13,1_k).

The distribution ratio calculating means 122 includes a load monitoring means 121
Based on the processing time of the partial processing for each divided data measured by the above and the divided data management table held by the divided data management means 113, the above-described calculation method is performed at regular intervals,
Calculate the average processing time of each computer (Step S2 in FIG. 7)
1). Next, the computer 1_e with the largest average processing time (e is 1 or more
(an integer less than or equal to n) (Step S23 in FIG. 7), and Calculator 1
Among the computers having the same divided data stored in _e, the computer 1_f having the shortest average processing time is specified (step S24 in FIG. 7). Further, regarding the distribution ratio of the divided data shared by the computers 1_e and 1_f, the computer
The ratio of the number of requests to 1_e is reduced, and the ratio of requests to the computer 1_f is increased (step S25 in FIG. 7). The average processing time of each computer is calculated again (step S21 in FIG. 7). If the average processing time is uniform, the process ends. If not, the change of the distribution ratio is repeated (step S22 in FIG. 7). As a result, it is possible to calculate a distribution ratio that can make the average processing time of each computer uniform.

Here, in step S22 in FIG. 7, the processing is repeated until the average processing time becomes uniform. However, the calculation may be terminated when the average value of the processing time of each computer changes to a steady state. . A method of determining the number of repetitions in advance and terminating the calculation in the middle is also conceivable.

Next, with reference to FIG. 8, an operation when a failure such as a failure occurs in some computers or secondary storage devices will be described.

At a certain moment, it is assumed that the computer 1_h (h is an integer of 1 or more and n or less) is stopped for an unexpected reason such as a failure. The fault detecting means 115 detects that the computer 1_h has stopped, notifies the distribution ratio calculating means 122, and starts an interrupt process (step S31 in FIG. 8). As a method of failure detection, there is a method in which small data exchange is periodically performed with each computer via the external communication network 3, and when the data exchange becomes impossible, it is determined that the computer has failed. Alternatively, there is a method of incorporating a dedicated failure diagnosis mechanism in hardware.

The distribution ratio calculating means 122, which has started the interruption processing at the time of failure, fixes the distribution ratio to the computer 1_h to 0 and recalculates the distribution ratio according to the method shown in FIG. 7 (FIG. 8).
Steps S32 and S33). Next, the request processing management means 11
6 referring to the request processing management table held by computer 1_h
Then, the partial process that has not returned the result after the request is passed to the load distribution unit 123, and is redistributed to computers that have not stopped (step S34 in FIG. 8).

Thereafter, the partial processing for the input processing request is not requested to the computer 1_h because the distribution ratio to the computer 1_h is 0. When the computer 1_h recovers, by appropriately giving the distribution ratio to the computer 1_h and re-calculating the distribution ratio by the distribution ratio calculating means 122, it is possible to equalize the load in the configuration including the computer 1_h. it can.

Next, the effect of the present embodiment will be described.

In the present embodiment, the same divided data is owned by a plurality of different computers and the distribution ratio of the partial processing is changed, so that the number of requests for the partial processing to the computer having a long average processing time is reduced. The number of requests for partial processing to a computer with a short average processing time is increasing. for that reason,
Even if the amount of data read by each computer is different,
The entire processing time of each computer can be made uniform, and the throughput of the entire system can be maximized.

The second effect of the present embodiment is that the calculation of the distribution ratio uses the average processing time of the partial processing for each divided data, so that the time of the input / output processing of the secondary storage device is reflected. That is, the load can be equalized.

The third effect of the present embodiment is that the distribution ratio calculation means calculates the distribution ratio so as to equalize the processing time of each computer at regular time intervals. It is possible to flexibly cope with a change in the frequency distribution.

The fourth effect of the present embodiment is that the distribution ratio is recalculated in a configuration other than the computer stopped when a failure occurs, and the unprocessed partial processes are redistributed from the stopped computer to other computers. Therefore, even if an unexpected failure occurs, the service can be continued without interruption,
Another object of the present invention is to make the load on the shoulder of the stopped computer uniform, thereby preventing performance degradation when a failure occurs.

[Second Embodiment] Next, a second embodiment of the present invention will be described.
An embodiment will be described in detail with reference to the drawings.

Referring to FIG. 17, according to the second embodiment of the present invention, the load management unit 12 of each of the computers 1_1 to 1_n (n is an integer of 1 or more and n or less) includes a load monitoring unit 121 and a distribution ratio calculation. The difference is that a calculation time control unit 124 is provided in addition to the configuration of the unit 122 and the load distribution unit 123.

The calculation time control means 124 counts the number of unprocessed partial processes after requesting each computer, and activates the distribution ratio calculation means 122 when the deviation becomes large. The operation of the present embodiment will be described in detail with reference to the drawings.

The method of storing the divided data in the secondary storage devices 2_1 to 2_n, and the request receiving unit 111, the request distribution unit 112, and the divided data management unit included in the reception processing unit 11 of each of the computers 1_1 to 1_n
113, the partial result integration means 114, the failure detection means 115, the request processing management means 116, the request processing means 131 included in the data processing unit 13, the load monitoring means 121 in the load management unit 12, and the operations of the load distribution means 123 Since the operation is the same as that of each means of the first embodiment, the description is omitted.

In the first embodiment, the distribution ratio calculating means 1
22 is started at a predetermined time interval specified in advance, and calculates a distribution ratio for equalizing the load of each computer. In the present embodiment, when the deviation in the number of partial processes that have not been replied to each computer after being requested increases, the distribution ratio calculation means 122
And recalculate the distribution ratio.

Referring to FIG. 18, calculation time control means 124
Is started each time the load distribution unit 123 determines the request destination of the partial process, and refers to the request process management table held by the request process management unit 116 to execute The number is counted (S41 in FIG. 18). Next, the calculation time control means 124 calculates the bias V of the unprocessed partial process of each computer (S42 in FIG. 18). Here, the bias V is the maximum value Umax of the number of unprocessed partial processes of all computers.
And the ratio Umax / Umin of the minimum value Umin. This bias V = Uma
When x / Umin exceeds a predetermined threshold value, the distribution ratio calculation means 122 is activated, and the distribution ratio is recalculated and corrected.

Here, the bias V is defined as the ratio between the maximum value and the minimum value of the number of unprocessed partial processes, but other than that, the ratio between the average value and the maximum value, the variance in each computer, etc. may be used. Good.

Next, the effect of the present embodiment will be described.

In the present embodiment, the calculation time control means 124
If the deviation of the number of unprocessed partial processes after requesting each computer increases, the distribution ratio calculation unit 122 is activated to recalculate the distribution ratio. Thus, even if the average processing time for each divided data suddenly changes and the load balance is broken, it is possible to flexibly cope with the load.

[Third Embodiment] Next, a third embodiment of the present invention will be described.
An embodiment will be described in detail with reference to the drawings.
Referring to FIG. 20, a third embodiment of the present invention is similar to the first and second embodiments of the present invention, in which computers 1_1 to 1
_n, secondary storage devices 2_1 to 2_n, an external communication network 3, and a recording medium 4 on which a control program is recorded. The recording medium 4 may be a magnetic disk, a semiconductor memory, a CD-ROM, or another recording medium.

The control program is transmitted from the recording medium 4 to the computer 1_1.
To the computer 1_2 through the external communication network 3.
Read in 1_n. The control program is computer 1_1
To 1_n, and the divided data is distributed and redundantly stored in the secondary storage devices 2_1 to 2_n. The computers 1_1 to 1_n execute the same processing as the processing by the computers 1_1 to 1_n in the first and second embodiments under the control of the control program.

[0074]

Next, the operation of the first embodiment of the present invention will be described with reference to specific examples. Here, as a configuration example using the first embodiment, a full-text search system that responds to a document group that matches a condition such as a keyword from a large-scale document group will be described.

FIG. 9 shows a full-text search system configured using this embodiment. Referring to FIG. 9, in the present embodiment, general-purpose personal computers P1_1, P1_2, P1_3 are used as computers each having a local secondary storage device.
And Ethernet as an external communication network connecting them. Here, a case where the number of computers is n = 3 will be described.

Each of the personal computers 1_1 to 1_3 includes a central processing unit functioning as three control units of a reception processing unit 11, a load management unit 12, and a data processing unit 13, and a hard disk as a secondary storage device. The reception processing unit 11 includes a request reception unit 111, a request distribution unit 112, a divided data management unit.
113, a partial result integrating unit 114, a failure detecting unit 115, and a request processing managing unit 116, and the load managing unit 12 has functions of a load monitoring unit 121, a distribution ratio calculating unit 122, and a load distributing unit 123. The unit 13 has the function of the request processing unit 131.

Referring to FIG. 10, a description will be given of a method of distributing and redundantly storing divided data on a hard disk in a case where the granularity is m = 2 and the degree of redundancy is r = 1. Also, in this example,
It is assumed that there are 60,000 large document groups as documents to be searched. Documents to be searched are assigned document numbers of ID00001 to ID60000, respectively.

First, a large document group of 60,000 documents is divided into m × n = 2 × 3 = 6 by 10,000 documents. Here, the divided document group D11 includes documents of ID00001 to ID10000, and the divided document group D12 includes documents of ID10001 to ID20000. Similarly, the divided document group Dij (i is an integer of 1 or more and 3 or less; j is an integer of 1 or more and 2 or less) includes ID ((k−1) × 10000 + 1) to ID (k × 10000) (where k = (i-1) * 2 + j). Here, the document is divided based on the number of documents, but there are other methods such as dividing the document based on the document capacity.

The divided document groups D11 to D32 are distributed and stored.
A search by a linear search may be performed on each divided document group. However, it is general to create an index from each document group in advance in order to improve the search speed. Here, the index is a file that creates a list of document numbers in which all words appearing in the document group appear. By using this index file, a list of documents in which the specified word appears can be obtained without reading out all the documents in the search.

In FIG. 10, a division index X11 is created from the division document group D11, and a division index X12 is created from the division document group D12. Hereinafter, similarly, the divided document group Dij
(I is an integer of 1 or more and 3 or less; j is an integer of 1 or more and 2 or less.), A division index Xij is created. The division index created here is an original division index.

Here, the original division indexes Xi1 to Xim
(I is an integer of 1 or more and 3 or less) is stored in the hard disk 2_i. Specifically, the original division indexes X11 and X12 are stored in the hard disk P2_1, and the original division indexes X21 and X21 are stored.
X22 is stored in the hard disk P2_2, and the original document division indexes X31 and X32 are stored in the hard disk P2_3. In addition, r = 1 copies of each split index are created,
Replication division index Dij (i is an integer from 1 to 3; j is 1
An integer between 2 and 2. Hard disk P while circulating)
Store redundantly in 2_ (i + j). Specifically, the replication division index X11 is stored in the hard disk P2_2, the replication division index X12 is stored in the hard disk P2_3, and the replication division index X21 is stored in the hard disk P2_3. Furthermore, the replication division indexes X22 and X31 are stored in the hard disk P2
Instead of _4, it is stored as a remainder of the subscript divided by 3 in the hard disk P2_1 in a circular manner, and the replication division index X32
Is also stored in a circular manner on the hard disk P2_2.

At this time, the personal computers 1_1 to 1_1
The divided data management means included in the central processing unit of _3 creates a divided data management table as shown in FIG.

Also, the personal computers 1_1 to 1_3
It is assumed that the distribution ratio of the partial process to the divided index is determined as shown in FIG. 12A by the distribution ratio calculating means of the central processing unit.

Next, a normal operation will be described. At a certain point, it is assumed that a search processing request is transmitted from the user to the personal computer P1_3 via Ethernet. The request accepting means of the central processing unit of the personal computer P1_3 receives this search request and creates six partial processes for which the divided indexes X11 to X32 are to be searched.

Next, the load distribution means of the central processing unit of the personal computer P1_3 determines the request destinations of the six partial processes. An example of a method in which the load distribution means determines the request destination of the partial processing will be described with reference to FIG. Referring to FIG. 12 (b), the load distribution unit prepares an array having a length of 100 for each of the division indices X11 to X32 in advance, and paints them in accordance with the distribution ratio calculated by the distribution ratio calculation unit. For example, the split index X11 is assigned to the personal computers P1_1 and P1_2 by 40% and 60%, respectively.
In order to request partial processing at a ratio of%, the number indicating the personal computer P1_1 is substituted into the first 40 elements of the array of length 100, and the number indicating the personal computer P1_2 is substituted into the remaining 60 elements. Keep it. When determining the request destination of the partial processing, first, a random number x from 1 to 100 is generated, and by referring to the x-th element of the array, the request destination personal computer can be determined according to the distribution ratio.

As a method for the load distribution means to determine the request destination of the partial processing according to the distribution ratio, there is a method in which the order of requests to the personal computers is determined in advance so as to satisfy the distribution ratio, and the order is followed. However, the method is not limited to the method described here.

Suppose that the load distribution means of the central processing unit of the personal computer 1_3 performs partial processing (Q1, X
22) to personal computer 1_2. Partial processing (Q1, X2
2) is requested by the request processing means of the central processing unit of the personal computer P1_2 designated via Ethernet by the request distribution means of the central processing unit of the personal computer P1_3. At the same time, personal
The request processing management means of the central processing unit of the computer P1_3 records in the request processing management table which personal computer requested the partial processing (Q1, X22). FIG. 13 shows an example of the request processing management table at a certain time.

The personal computer P1_2 requested for the partial processing (Q1, X22) searches the search target divided index X22 by the request processing means of the central processing unit, and obtains a list of the document group satisfying the search condition Q1. obtain. This is used as a partial result and transmitted to the partial result integrating means of the central processing unit of the personal computer P1_3, which is the request source, via Ethernet. Further, the processing time required to execute the partial processing (Q1, X22) is transmitted to the load monitoring means of the central processing unit of the personal computer P1_3, which is the request source. At the same time, the request processing management means of the central processing unit of the personal computer P1_3 deletes the partial processing (Q1, X22) from the request processing management table of FIG.

On the other hand, the partial result integrating means of the central processing unit of the personal computer P1_3 integrates the results when all the partial processing results for the search indexes Q11 to X32 have been completed for the search condition Q1. , The order of the documents is rearranged based on the frequency of appearance of the keywords, and the response is made to the user.

Here, a method of calculating the distribution ratio of the partial processing in order to equalize the load by the distribution ratio calculating means of the central processing unit of the personal computer P1_3 will be described with a specific example.

During a certain period, the average processing time for each divided index measured by the load monitoring means of the central processing unit of the personal computer P1_3 is shown in FIG.
Let's say it looks like 14. If the distribution ratio at this time has the value shown in FIG. 12A, the average processing time of the personal computers P1_1 to P1_3 can be obtained as shown in FIG. For example, the personal computer P1
_1, the average processing time E (P1_1) is E (P1_1) = T (X11) × d (X11, P1_1) + T (X12) × d (X12, P1_1) + T (X2
2) La (d22, P1_1) + T (X31) Lad (X31, P1_1) = 100 × 0.4 + 200 × 0.8 + 100 × 0.5 + 400 × 0.5 = 450

Referring to FIG. 15, a personal computer
It can be seen that the average processing time of 1_3 is the longest. Figure 1
Referring to the split data management table shown in Fig. 1,
Owns the same split index as computer P1_3,
Personal computers have the shortest average processing time
It turns out that it is P1_2.

Therefore, the personal computers P1_3 and
For the divided index X21 commonly owned by P1_2, the ratio of distribution to the personal computer P1_3 is reduced, and the ratio of distribution to the personal computer P1_2 is increased. Here, it is assumed that P1_2: P1_3 = 50: 50 is newly set.

[0094] Using the new distribution ratio, each personal
When the average processing time of the computers P1_1 to P1_3 is recalculated, they are 450, 350, and 400, respectively. This time, the average processing time of the personal computer P1_1 is the longest, and the personal computer P1_2 having the minimum average processing time and the division index X14 are commonly owned.

Therefore, the distribution ratio of the divided index X14 to the personal computer P1_1 is lowered and the distribution ratio to the personal computer P1_2 is newly raised. For example, when P1_2: P1_1 = 100: 0 is newly set, the average processing time of each of the personal computers 1_1 to 1_3 becomes 400, and the average processing time thereafter can be made uniform.

Here, the method of heuristically changing the distribution ratio has been described. However, a unit for changing the distribution ratio is determined, and the distribution ratio is gradually changed for each unit. Automatically calculate the distribution ratio by repeatedly changing the distribution ratio so that the maximum processing time matches the target value with the average processing time when the load is equalized as the target value, etc. There is also a method.

Next, the operation when some personal computers are stopped will be described with reference to a specific example.

Normally, personal computers P1_1 to P1
The fault detection means included in the central processing unit of # 3 periodically exchanges minute data with each other to check whether the other party is operating normally. At some point, if the personal computer P1_1 is stopped due to an unexpected failure, the response from the personal computer P1_1 is lost, so the failure detection means of the central processing unit of the personal computers P1_2, P1_3 is Can be detected.
At this time, both the personal computers P1_2 and P1_3 perform interrupt processing at the time of failure, but the following description focuses on the personal computer P1_3.

When the failure detecting means of the central processing unit of the personal computer P1_3 determines that the personal computer P1_1 has stopped, it notifies the distribution ratio calculating means of the central processing unit of the personal computer P1_3. The distribution ratio calculation means sets the distribution ratio of the partial processing to the personal computer P1_1 to zero. Then, the distribution ratio becomes as shown in FIG.

As it is, a personal computer
Since the processing time between P1_2 and the personal computer P1_3 is biased, the distribution ratio calculating means recalculates the distribution ratio. The recalculation method is the same as the normal calculation method described above, except that the distribution ratio of the partial processing to the personal computer 1_1 is fixed to 0.

The average processing time of each divided index is shown in FIG.
In the distribution ratio of FIG. 16, the average processing time E (P1_2) of the personal computers P1_2 and P1_3 is
E (P1_3) can be calculated as follows. E (P1_2) = T (X21) La d (X21, P1_2) + T (X22) La d (X22, P1_2) + T (X
11) LA d (X11, P1_2) + T (X32) LA d (X32, P1_2) = 300 × 0.5 + 100 × 1 + 100 × 1 + 100 × 0.1 = 360 E (P1_3) = T (X31) LA d (X31, P1_2) + T (X32) d (X32, P1_2) + T (X
12) LA d (X12, P1_2) + T (X21) LA d (X21, P1_2) = 400 × 1 + 100 × 0.9 + 200 × 1 + 300 × 0.5 = 840

Therefore, the personal computer P1_2,
For the divided index X21 commonly owned by P1_3, the ratio of distribution to the personal computer P1_2 is increased and the ratio of distribution to the personal computer P1_3 is decreased. When P1_2: P1_3 = 100: 0 newly, E (P1_2) = 5
10, E (P1_3) = 690. Although the bias still remains, even if the distribution ratio is newly changed to P1_3: P1_2 = 0: 100 for the split index X32 commonly owned by the personal computers P1_2 and P1_3, E (P1_
2) = E (P1_3) = 600, and the load on the shoulder when a failure occurs can be equalized.

Next, after requesting the personal computer P1_1, the central processing unit of the personal computer P1_3 distributes the partial process for which the result has not been returned with reference to the request processing management table created by the request processing management means. Hand over the means,
Redistribute according to the new distribution ratio.

Next, a second embodiment will be described with reference to the drawings. This example corresponds to the second embodiment.

The present embodiment has the same configuration as that of the first embodiment, except that the personal computers P1_1 to P1_3 also function as start-up time control means for determining the start timing of the distribution ratio calculation means. This is different from the first embodiment.

The method of storing the divided data, the operation at the time of normal operation, the method of calculating the distribution ratio, and the operation in a state where some personal computers are stopped are the same as those in the first embodiment, and therefore will not be described here.

A method of determining the timing at which the start-up time control means included in the central processing unit of the personal computer P1_1 starts the distribution ratio calculation means will be described.
Here, the bias V of the unprocessed partial process is defined as the ratio Umax / Umin of the maximum value Umax and the minimum value Umin, and the threshold value of the bias V is set to 3 as a condition for activating the distribution ratio calculation means.

At a certain point in time, the calculation time control means included in the central processing unit of the personal computer P1_1 counts the number of unprocessed partial processes after requesting each computer, and as shown in FIG. 19 (a). I do. At this time, since the maximum value Umax of the number of unanswered partial processes is 7 and the minimum value Umin is 4, the bias V = Umax / Umin = 1.75. Since this is smaller than the threshold value 3, it is determined that the load is equalized,
The distribution ratio calculation means continues without being activated.

Next, at another time, the calculation time control means included in the central processing unit of the personal computer P1_1 counts the number of unprocessed partial processes after requesting each computer. Let's say that. At this time, since the maximum value Umax of the number of unanswered partial processes is 10 and the minimum value Umin is 2, the bias V = Umax / Umin = 5. Since this is larger than the threshold value 3, it is determined that the load balance has been lost due to a change in the average processing time for each of the divided data, and the distribution ratio calculation means included in the central processing unit of the personal computer P1_1 is started. , The distribution ratio is recalculated and corrected.

[0110]

The first effect is that it is possible to equalize the load reflecting the input / output processing from the secondary storage device. The reason is that the distribution ratio for making the processing time uniform is obtained by using the average processing time of the partial processing including the input / output processing from the secondary storage device.

The second effect is that, even if the amount of data read and the amount of processing read by each computer are different, the processing time of each computer can be made uniform and the throughput of the entire system can be maximized. The reason is that a plurality of identical divided data are distributed and redundantly stored between different computers, the number of requests for partial processing to a computer with a long average processing time is reduced, and the partial processing to a computer with a short average processing time is reduced. By increasing the number of requests, the processing can be replaced, and the overall processing time of each computer is made uniform.

A third effect is that the entire processing time of each computer can be flexibly equalized even if the distribution of the access frequency to each data changes. The reason is that the distribution ratio for equalizing the load is recalculated periodically or when the load of each computer becomes uneven.

The fourth effect is that the service can be continued even when some computers are stopped due to a failure or maintenance, and a significant performance drop is prevented. The reason is that the distribution ratio is recalculated in a configuration other than the computer stopped when a failure occurs, and the unprocessed partial process is redistributed from the stopped computer to another computer.

[Brief description of the drawings]

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

FIG. 2 is a schematic diagram showing a method of creating divided data and a method of distributing and storing redundant data.

FIG. 3 is an example of a divided data management table created by a divided data management unit.

FIG. 4 is an example of a request processing management table created by a request processing management unit.

FIG. 5 is an example of a distribution ratio calculated by a distribution ratio calculation unit.

FIG. 6 is a flowchart showing the operation of the first embodiment of the present invention.

FIG. 7 is a flowchart showing a procedure in which a distribution ratio calculation unit calculates a distribution ratio.

FIG. 8 is a flowchart showing an operation when a failure occurs in the first embodiment of the present invention;

FIG. 9 is a schematic configuration diagram of the first embodiment of the present invention.

FIG. 10 is a schematic diagram showing a method for storing divided data according to the first embodiment of the present invention;

FIG. 11 is a divided data management table according to the first embodiment of the present invention.

FIG. 12 shows a distribution ratio according to the first embodiment of the present invention;
Examples of distribution methods

FIG. 13 is a request processing management table according to the first embodiment of the present invention.

FIG. 14 shows the processing time of each divided index measured by the load monitoring means in the first embodiment of the present invention.

FIG. 15 is an explanatory diagram of a calculation method of a processing time of each computer in the first embodiment of the present invention.

FIG. 16 is a diagram for explaining recalculation of a distribution ratio when a failure occurs in the first embodiment of the present invention.

FIG. 17 is a schematic configuration diagram of a second embodiment of the present invention.

FIG. 18 is a flowchart showing the operation of the second embodiment of the present invention.

FIG. 19 is a request processing management table according to the second embodiment of the present invention.

FIG. 20 is a schematic configuration diagram of a third embodiment of the present invention.

[Explanation of symbols]

 1_1 ~ 1_n Computer 2_1 ~ 2_n Secondary storage device 3 External communication network 4 Recording medium 11 Reception processing unit 12 Load management unit 13 Data processing unit 111 Request reception unit 112 Request distribution unit 113 Divided data management unit 114 Partial result integration unit 115 Failure Detection means 116 Request processing management means 121 Load monitoring means 122 Distribution ratio calculation means 123 Load distribution means 124 Calculation time control means 131 Request processing means D11-Dnm Divided data X11-X32 Divided index

Claims (14)

[Claims] A divided data obtained by dividing data to be searched is distributed and stored in a secondary storage device provided in a plurality of computers, and a duplicate of the divided data is created so as not to be duplicated. A distributed parallel data processing apparatus, wherein a computer that performs partial processing on each divided data is determined in consideration of one or both of a load state of a computer and a failure state of the computer. . 2. A distributed parallel data processing apparatus for performing a search process on a secondary storage device in which divided data obtained by dividing data to be searched is stored in a distributed manner, wherein a processing time for the divided data is measured. A distribution ratio capable of equalizing the load of the computer that manages the secondary storage device is determined based on the distribution ratio, and a computer that performs partial processing on each divided data is determined based on the distribution ratio thus determined. A distributed / parallel data processing apparatus for transmitting the partial processing request to a corresponding computer. 3. A distributed parallel data processing device for performing a search process on a secondary storage device in which divided data obtained by dividing data to be searched is stored in a distributed manner, wherein a processing request from a user is received. Request accepting means for creating a plurality of partial processes specifying designated divided data, request distribution means for transmitting the partial processes to designated computers, divided data managing means for holding storage locations of the respective divided data, Collects, integrates,
A partial result integrating means for editing and responding to the user; a load monitoring means for measuring a processing time for each divided data; a load on each computer in advance based on the processing time for each divided data measured by the load monitoring means. Distribution ratio calculation means for obtaining a distribution ratio capable of equalizing, load distribution means for determining a computer for distributing a partial process for each divided data based on the calculated distribution ratio and designating the request distribution means, Request processing means for performing a partial process on the divided data stored in the secondary storage device. 4. A distributed parallel data processing device for performing a search process on a secondary storage device in which divided data obtained by dividing data to be searched is stored in a distributed manner. Request accepting means for creating a plurality of partial processes specifying designated divided data, request distribution means for transmitting the partial processes to designated computers, divided data managing means for holding storage locations of the respective divided data, Collects, integrates,
The partial result integration means that edits and responds to the user, the failure detection means that automatically detects a stopped computer, and the partial processing that has been requested to other computers are retained. Request processing management means for redistributing the partial processing of the reply, load monitoring means for measuring the processing time for each divided data, and stopping when a failure occurs, based on the processing time for each divided data measured by the load monitoring means. A distribution ratio calculating means for performing recalculation with the distribution ratio to the computer set to 0, and a computer for distributing the partial processing for each divided data based on the distribution ratio calculated in advance, and designating the request distribution means A distributed parallel data processing device comprising: a load distribution unit; and a request processing unit that performs a partial process on the divided data stored in the secondary storage device. Place. 5. A distributed parallel data processing device for performing a search process on a secondary storage device in which divided data obtained by dividing data to be searched is stored in a distributed manner, wherein a processing request from a user is received. Request accepting means for creating a plurality of partial processes specifying designated divided data, request distribution means for transmitting the partial processes to designated computers, divided data managing means for holding storage locations of the respective divided data, Collects, integrates,
A partial result integrating means for editing and responding to the user; a load monitoring means for measuring a processing time for each divided data; a load on each computer in advance based on the processing time for each divided data measured by the load monitoring means. Distribution ratio calculation means for obtaining a distribution ratio capable of equalizing, load distribution means for determining a computer for distributing a partial process for each divided data based on a distribution ratio calculated in advance, and designating the computer as a request distribution means, When the deviation in the number of unprocessed partial processes after requesting each computer increases, a calculation time control unit that activates the distribution ratio calculation unit, and a request to perform a partial process on the divided data stored in the secondary storage device And a processing means. 6. A distributed parallel data processing device for performing a search process on a secondary storage device in which divided data obtained by dividing data to be searched is stored in a distributed manner, wherein a processing request from a user is received. Request accepting means for creating a plurality of partial processes specifying designated divided data, request distribution means for transmitting the partial processes to designated computers, divided data managing means for holding storage locations of the respective divided data, Collects, integrates,
The partial result integration means that edits and responds to the user, the failure detection means that automatically detects a stopped computer, and the partial processing that has been requested to other computers are retained. Request processing management means for redistributing the partial processing of the reply, load monitoring means for measuring the processing time for each divided data, and stopping when a failure occurs, based on the processing time for each divided data measured by the load monitoring means. A distribution ratio calculation unit that performs recalculation with the distribution ratio to the computer set to 0, and a computer that determines the computer that distributes the partial processing for each divided data based on the distribution ratio calculated in advance and specifies the load to the request distribution unit Distribution means, calculation time control means for activating the distribution ratio calculation means when the deviation of the number of partial processes not responded to each computer after the request becomes large, and the secondary storage device Distributed Parallel data processing apparatus characterized by comprising: a request processing means for performing partial processing on the pay has been divided data. 7. A recording medium which records a distributed parallel data processing program for performing a search process using a computer on a secondary storage device in which divided data obtained by dividing data to be searched is distributed and stored. A request receiving function for receiving a processing request from a user to a computer and creating a plurality of partial processes specifying a target divided data; a request distribution function for transmitting a partial process to a designated computer; A divided data management function for holding a data storage destination; collecting and integrating results of partial processing on the divided data;
A partial result integration function for editing and responding to the user, a load monitoring function for measuring a processing time for each divided data, and a processing time for each divided data measured by the load monitoring function for each computer in advance. A distribution ratio calculation function for obtaining a distribution ratio capable of equalizing the load, and a load distribution function for determining a computer for distributing a partial process for each divided data based on the calculated distribution ratio and specifying the computer for the request distribution function. And a request processing function of performing a partial process on the divided data stored in the secondary storage device, wherein the recording medium stores a distributed parallel data processing program. 8. A recording medium recording a distributed parallel data processing program for performing a search process using a computer on a secondary storage device in which divided data obtained by dividing data to be searched is distributed and stored. A request receiving function for receiving a processing request from a user to a computer and creating a plurality of partial processes specifying a target divided data; a request distribution function for transmitting a partial process to a designated computer; A divided data management function for holding a data storage destination; collecting and integrating results of partial processing on the divided data;
The partial result integration function that edits and responds to the user, the failure detection function that automatically detects a stopped computer, and the partial processing that has been requested to other computers are retained and the A request processing management function for redistributing the partial processing of the response, a load monitoring function for measuring the processing time for each divided data, and a stop when a failure occurs based on the processing time for each divided data measured by the load monitoring function. A distribution ratio calculation function that performs recalculation by setting the distribution ratio to the calculated computer to 0, and a computer that distributes partial processing for each divided data based on the distribution ratio calculated in advance, and designates it to the request distribution function And a load processing function for performing a partial processing on the divided data stored in the secondary storage device. A recording medium on which a type data processing program is recorded. 9. A recording medium storing a distributed parallel data processing program for performing a search process using a computer on a secondary storage device in which divided data obtained by dividing data to be searched is distributed and stored. A request receiving function for receiving a processing request from a user to a computer and creating a plurality of partial processes specifying a target divided data; a request distribution function for transmitting a partial process to a designated computer; A divided data management function for holding a data storage destination; collecting and integrating results of partial processing on the divided data;
A partial result integration function for editing and responding to the user, a load monitoring function for measuring a processing time for each divided data, and a processing time for each divided data measured by the load monitoring function for each computer in advance. A distribution ratio calculation function that calculates the distribution ratio that can equalize the load, and a load distribution function that determines the computer that distributes the partial processing for each divided data based on the distribution ratio calculated in advance and specifies it as the request distribution function When the deviation of the number of unprocessed partial processes after requesting each computer increases, a calculation time control function for activating a distribution ratio calculation function and a partial process for the divided data stored in the secondary storage device are performed. And a request processing function, wherein the recording medium records a distributed parallel data processing program. 10. A recording medium recording a distributed parallel data processing program for performing a search process using a computer on a secondary storage device in which divided data obtained by dividing data to be searched is distributed and stored. A request receiving function for receiving a processing request from a user to a computer and creating a plurality of partial processes specifying a target divided data; a request distribution function for transmitting a partial process to a designated computer; A divided data management function for holding a data storage destination; collecting and integrating results of partial processing on the divided data;
The partial result integration function that edits and responds to the user, the failure detection function that automatically detects a stopped computer, and the partial processing that has been requested to other computers are retained and the A request processing management function for redistributing the partial processing of the response, a load monitoring function for measuring the processing time for each divided data, and a stop when a failure occurs based on the processing time for each divided data measured by the load monitoring function. A distribution ratio calculation function that performs recalculation with the distribution ratio to the calculated computer set to 0, and a computer that distributes partial processing for each divided data based on the distribution ratio calculated in advance and specifies it to the request distribution means A load distribution function, a calculation time control function for activating the distribution ratio calculation means when the deviation of the number of partial processes that have not been replied to each computer after being requested increases, Distributed Parallel data processing program recorded recording medium to a request processing function, characterized in that to realize the performing partial processing on has been divided data stored in the device. 11. A computer comprising a plurality of computers connected by an external communication network, each of which has a local secondary storage device, divides large-scale data to be searched and stores the divided data in the secondary storage device in a distributed manner. A distributed parallel data system for performing processing on the divided data stored in the secondary storage device in parallel on a non-shared type cluster system in which a copy of the divided data is redundantly stored so as not to be duplicated in the same secondary storage device A request receiving means for receiving a processing request from a user and creating a plurality of partial processes specifying a target divided data; a request distribution means for transmitting the partial process to a designated computer; A divided data management unit that holds a storage location of each divided data, and collects, integrates, and collects results of partial processing on the divided data.
A partial result integrating means for editing and responding to the user; a load monitoring means for measuring a processing time for each divided data; and a processing time for each computer in advance based on the processing time for each divided data measured by the load monitoring means. Distribution ratio calculation means for obtaining a distribution ratio capable of equalizing the load; load distribution means for determining a computer for distributing a partial process for each divided data based on a distribution ratio calculated in advance, and designating the computer as a request distribution means; Request processing means for performing a partial process on the divided data stored in the secondary storage device. 12. A computer comprising a plurality of computers connected by an external communication network, each computer having a local secondary storage device, dividing large-scale data of interest and storing the divided data in the secondary storage device in a distributed manner. A distributed parallel data processing system that performs processing on divided data stored in a secondary storage device in parallel on a non-shared type cluster system in which copies of the divided data are redundantly stored so as not to be duplicated in the same secondary storage device Request receiving means for receiving a processing request from a user and creating a plurality of partial processes specifying target divided data; request distributing means for transmitting the partial process to a designated computer; Division data management means for holding a storage location of data; collecting and integrating results of partial processing on the division data;
The partial result integration means that edits and responds to the user, the failure detection means that automatically detects a stopped computer, and the partial processing that has been requested to other computers are retained. Request processing management means for redistributing the partial processing of the reply, load monitoring means for measuring the processing time for each divided data, and stopping when a failure occurs based on the processing time for each divided data measured by the load monitoring means. Distribution ratio calculating means for performing recalculation with the distribution ratio to the calculated computer set to 0, and a computer for distributing the partial processing for each divided data based on the distribution ratio calculated in advance, and designating the request distribution means A distributed parallel data processing system comprising: a load distribution unit; and a request processing unit for performing a partial process on the divided data stored in the secondary storage device. Stem. 13. A computer comprising a plurality of computers connected by an external communication network, each of which has a local secondary storage device, divides large-scale data of interest and stores the divided data in the secondary storage device in a distributed manner. A distributed parallel data processing system that performs processing on divided data stored in a secondary storage device in parallel on a non-shared type cluster system in which copies of the divided data are redundantly stored so as not to be duplicated in the same secondary storage device Request receiving means for receiving a processing request from a user and creating a plurality of partial processes specifying target divided data; request distributing means for transmitting the partial process to a designated computer; Division data management means for holding a storage location of data; collecting and integrating results of partial processing on the division data;
A partial result integrating means for editing and responding to the user; a load monitoring means for measuring a processing time for each divided data; and a processing time for each computer based on the processing time for each divided data measured by the load monitoring means. Distribution ratio calculation means for obtaining a distribution ratio capable of equalizing the load; load distribution means for determining a computer for distributing a partial process for each divided data based on a distribution ratio calculated in advance, and designating the computer as a request distribution means; When the deviation in the number of unprocessed partial processes after requesting each computer increases, a calculation time control unit that activates the distribution ratio calculation unit and a request to perform a partial process on the divided data stored in the secondary storage device A distributed parallel data processing system comprising: processing means. 14. A computer comprising a plurality of computers connected by an external communication network, each of which has a local secondary storage device, divides large-scale data of interest, and stores the divided data in the secondary storage device in a distributed manner. A distributed parallel data processing system that performs processing on divided data stored in a secondary storage device in parallel on a non-shared type cluster system in which copies of the divided data are redundantly stored so as not to be duplicated in the same secondary storage device Request receiving means for receiving a processing request from a user and creating a plurality of partial processes specifying target divided data; request distributing means for transmitting the partial process to a designated computer; Division data management means for holding a storage location of data; collecting and integrating results of partial processing on the division data;
The partial result integration means that edits and responds to the user, the failure detection means that automatically detects a stopped computer, and the partial processing that has been requested to other computers are retained. Request processing management means for redistributing partial processing of a reply, load monitoring means for measuring the processing time for each divided data, and stopping when a failure occurs based on the processing time for each divided data measured by the load monitoring means Distribution ratio calculating means for performing recalculation with the distribution ratio to the calculated computer set to 0, and a computer for distributing the partial processing for each divided data based on the distribution ratio calculated in advance, and designating the request distribution means Load distribution means, calculation time control means for activating the distribution ratio calculation means when the number of unprocessed partial processes after requesting each computer becomes large, and storage in the secondary storage device. Distributed Parallel data processing system characterized in that it and a request processing means for performing partial processing on the divided data.