JP2001167098A - Distributed parallel analysis of large amounts of data - Google Patents

Abstract

(57) [Summary] [Problem] Regarding a parallel distributed processing method of data mining that discovers knowledge from a large amount of data, conventionally, data is divided and distributed to a plurality of processing devices, so that data to be analyzed among the processing devices is analyzed. In order to execute the parallel distributed processing method, it is necessary to distribute data from a database to each processing device as preprocessing. Another problem is that the processing device requires a large-capacity main memory. One or a plurality of data storage means, an analysis result totaling means, and a plurality of data analysis means are used. The data storage means transmits the data to be analyzed once, and the plurality of data analysis means receives the data. Each data analysis unit analyzes the received data, and thereafter, the analysis result obtained by each data analysis unit is totaled by the analysis result totaling unit, and the total analysis result is obtained.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

[0001] 1. Field of the Invention [0002] The present invention relates to a data analysis technique for a large amount of data.

[0002]

2. Description of the Related Art A technique for finding knowledge from a large amount of data is called data mining. As a specific example of the discovered knowledge, an association rule (Association Rule) is well known. The basic concept of association rules is described in the document "Mining A
ssociation Rules between Setof Items in Large Data
bases "(procededing of ACM SIGMOD, 1993). According to it, m binary attributes from I_1 to I_m (called an item and having one of 0 or 1) and a binary vector consisting of m elements corresponding to the item (called a transaction) ) And this set of transactions T, the association rule is "X
→ I_j ”. Here, X is a set of items (called an item set) consisting of some of the m items I_1 to I_m, and I_j is a single item not included in X. When one transaction t and one item i are considered, if the value of the element corresponding to i among the elements of t is 1, it is said that the transaction t satisfies the item i, and t is included in the item set X. A transaction t satisfies item set X when all items satisfied are satisfied. Let K be the number of transactions that satisfy itemset X in transaction set T.
Assuming that the number of transactions that satisfy item set X and also satisfy item I_j is J, the ratio J / K
Is referred to as “confidence” of the association rule “X → I_j”. In addition, the ratio of the above J to the entire transaction set T is determined by “support” of the correlation rule “X → I_j”.
Call. Also, the ratio of the above K to the entire transaction set T is called “support” of the item set X.

There can be many combinations of the item set X and the item I_j, and a basic method for finding an association rule having a given minimum confidence c and a minimum confidence s and higher than s is provided. Refer to the "Fast Algorithms for Mining
Association Rules "(Proceedings of VLDB, 1994)
It is described in. In this document, among itemsets consisting of n items, a set of itemsets with support greater than or equal to the minimum support s is defined as a large item set L_n
The process of obtaining L_k based on L_ (k-1) is considered as one pass, and the value of k is incremented by 1 while repeating the pass until a new large item set cannot be obtained. Find a set of item sets that satisfy.

[0004] To obtain L_k based on L_ (k-1), first,
Create all possible itemsets of k items included in L_ (k-1), set these itemsets as the candidate itemset C_k, then scan the transaction set T, and For each itemset, count the number of transactions that satisfy the itemset, and calculate the support value accordingly. A set of item sets having support of s or more in the candidate item set C_k is defined as a new large item set L_k. This process is repeated until a new large item set cannot be obtained. After the large item set is obtained, in each of the item sets included in the large item set, the confidence is calculated for the association rule that can be created by using the included item, and the correlation rule that satisfies the minimum confidence c is selected. The association rule thus obtained is the final result.

When the above basic algorithm is executed by a computer, a set of transactions is held in a secondary storage, and a counter for counting the number of transactions satisfying the candidate item set is held in a main storage. At this time, there are two problems. One is to scan the entire transaction set each time one pass is performed,
The point is that a lot of processing time is spent reading data from the secondary storage. The other is that the candidate item set becomes very large depending on the number of items appearing in a given transaction set, and a large-capacity main memory is required to hold a counter.
A parallel distributed processing method for solving these problems is described in the document "Parallel Mining of Association Rules" (IEEE Tra
nsactions on Knowledge and Data Engineering, 199
6). This document describes three parallel algorithms. Each of these three algorithms has a plurality of processing units, and it is premised that a transaction set is divided and distributed and held in a secondary storage local to each processing unit. The first parallel algorithm is called “Count Distribution”, and it is possible to reduce the time required for reading data by scanning a transaction set in parallel with a plurality of processing devices when calculating support for a candidate item set. It is possible. (However, since the counters for calculating the support of the candidate item set are all held in each processing device, the point that a large-capacity main memory is required cannot be solved.)
The second parallel algorithm is called “Data Distribution”. By distributing and holding counters to a plurality of processing devices, it is possible to reduce the amount of main memory required in each processing device. (However, another problem arises in that transactions distributed to each processing unit need to be transferred between the processing units.)
For both ibution and Data Distribution, the support of the candidate item set obtained by each processing device is totaled for each pass, so if there is a difference in the processing time for each processing device, the slowest for each pass It is necessary to wait for the end of the processing device, and there is a problem that processing time is wasted. The third parallel algorithm is "Candidate Distributi
It is called "on" and aims to solve this problem. A set of transactions that satisfy a certain item set is called a support transaction set.

In the Candidate Distribution algorithm, an intermediate path m (m is determined heuristically)
In, the item sets included in the candidate item set C_m are grouped. At this time, the support transaction sets should not overlap as much as possible between the item sets belonging to different groups. According to the grouping, the candidate item set and its support transaction set are redistributed to a plurality of processing devices. Thus, a large item set in a subsequent pass can be obtained by scanning only the transactions distributed to the respective processing devices, and it is not necessary for all the processing devices to synchronize for each pass.

As a data mining method other than the association rule, a characteristic rule (CharacteristicRule) and a method for finding the characteristic rule are described in the document "Characteristic Rule Induction Algorithm"
forData Mining "(proceedings of PAKDD, 1998). In this document, a set of records including a plurality of fields is used as analysis target data. The feature rule is described as “if A then B”. A is a combination of one or more conditions, and B is a single condition. The "condition" here is a pair of a field and its value. For example,
If “X_i” is a field and “v_ij” is a value, a combination of these is a condition, and is described as “X_i = v_ij”. In the record to be analyzed, field X
If _i has the value v_ij, then that record is the condition "X_i =
v_ij ”. The feature rule has an evaluation value, and the evaluation value of the feature rule “if A thenB” is

[0008]

## EQU1 ## It is defined as P (A) ^ a log {P (B | A) / P (B)}. Here, P (A) and P (B) are the proportions of records that satisfy conditions A and B in the entire record to be analyzed, respectively, and P (B | A) is the percentage of records that satisfy condition A. Of these, the percentage of records that satisfy both conditions A and B. The index a is a heuristically determined positive constant.

[0009] The feature rule discovery method described in this document is based on the condition that the data to be analyzed and the then part, the upper limit of the number of conditions included in the if part, and the number of rules M are given.
This is a method of repeatedly generating feature rules and calculating evaluation values to find the M feature rules with the highest evaluation values.

In the feature rule finding method described in the above-mentioned document, it is necessary to generate one feature rule and scan the whole or a part of the data to be analyzed every time an evaluation value is calculated. As in the case of the rule, there is a problem that it takes time to read data from the secondary storage. To solve this problem, Japanese Patent Laid-Open No. Hei 11-3360 discloses a method of performing a scan of data to be analyzed only once. This method prepares a counter for counting the number of records satisfying the condition for all possible conditions, and counts the records for all conditions in one data scan. According to this, the time required for reading data can be reduced. However, on the other hand, there is a problem that a large-capacity main memory is required to hold the counter.

[0011]

In both cases of the association rule and the feature rule, the basic discovery method has problems of processing time required for scanning the data to be analyzed and capacity of main memory required for holding the counter. There is a problem that if one is to be made smaller, the other becomes larger. There are several parallel distributed processing methods for solving this. However, regarding the parallel distributed processing method for finding association rules, since data is divided and distributed to a plurality of processing devices, the processing time required for scanning the analysis target data is reduced. Although it is possible to reduce the data size, it is necessary to transfer the data to be analyzed between the processing devices, and each processing device still needs a large-capacity main memory to hold the counter. In addition, when data to be analyzed is stored in a normal database system without a data mining function, it is necessary to distribute data from the database to each processing device as preprocessing in order to execute the parallel distributed processing method. However, there is a problem that processing time for this is extra.

An object of the present invention is to reduce the transfer amount and the number of transfers of data to be analyzed in parallel distributed processing using a plurality of processing devices, and to reduce the amount of main memory required in each processing device. An object of the present invention is to provide a data mining method that can be executed while being suppressed and connected to an ordinary database system.

[0013]

The parallel variance analysis method of the present invention uses one or more data storage means and a plurality of data analysis means. When a single data storage unit is used, all the data to be analyzed is stored in the single data storage unit. When a plurality of data storage units are used, the data to be analyzed is divided and distributed and stored in the plurality of data storage units.

The data storage means transmits the data to be analyzed to a plurality of data analysis means. The plurality of data analysis units receive the same data, and each of the data analysis units analyzes the received data. After that, the analysis results obtained by the respective data analysis means are put together to make an overall analysis result.

The data storage means collectively transmits the data to be analyzed to a plurality of data analysis means once, and the plurality of data analysis means share the data. That is, the data storage unit does not individually transmit the analysis target data to each of the plurality of data analysis units.

In the parallel variance analysis method of the present invention, a common storage means shared by a plurality of data analysis means may be used. The plurality of data analysis units hold respective analysis results on the shared storage unit. Therefore, the entire analysis result can be obtained by reading the analysis result held in the shared storage means.

[0017]

DESCRIPTION OF THE PREFERRED EMBODIMENTS A first embodiment of the present invention will be described. FIG. 1 shows the configuration of the present embodiment. In this embodiment,
A data storage device 101, a plurality of data analysis devices 102,
The analysis result totaling device 103 is connected by a bus-type communication path 104. The data to be analyzed is stored in the data storage device.

FIG. 2 shows the procedure of data analysis. In the preparation process 201 of the data analysis device, each of the data analysis devices prepares for data analysis, and when preparation for receiving data to be analyzed is completed, a preparation completion signal is transmitted to the data storage device. The data storage device waits for reception of a preparation completion signal from all data analyzers (process 202). After receiving the ready signals from all the data analyzers, in step 203, the data storage device scans the data to be analyzed, and if there is any data that has not been transmitted, the process proceeds to step 204; If all the data to be analyzed have been transmitted), the process proceeds to step 206. In the record transmission process 204, the data storage device transmits one record of data that has not been transmitted yet.

A record transmitted once is treated as transmitted. The transmitted record is transmitted to a plurality of data analyzers via a bus-type communication path. Since the communication path is a bus type, the record is transmitted to all the data analyzers connected to the communication path in one transmission from the data storage device. In the data reception / analysis processing 205, the records transmitted from the data storage devices are received by the respective data analysis devices. After receiving the record, the data analyzer prepares to receive the next record, and when the preparation for reception is completed, transmits a signal indicating completion of preparation to the data storage device. In addition, data analysis is performed on records received together with records already received. After the data receiving and analyzing process 205, the process returns to the process 202. In the analysis result tallying process 206, the analysis result tallying device receives the analysis results from the data analysis device and totals them to obtain the entire analysis result. The above is the outline of the analysis method. In this way, a plurality of devices perform data analysis while cooperating with each other. Hereinafter, the processing performed in each device will be described in detail, taking discovery of a feature rule as an example.

The data to be analyzed is a set of records composed of a plurality of fields, and all records have the same number of fields. A record corresponds to an object to be analyzed, and a field corresponds to an attribute of the object. Taking shop customer information as an example, one record corresponds to one customer, and each field corresponds to a customer attribute such as gender or age. In feature rule discovery, each attribute value is converted into a small number of categories as preprocessing. For example, “age” is usually 10 to 100
The value can be in the range of about
It is converted into a small number of categories such as "36 to 55" and "56 and over". In addition, since "sex" can originally take only two values of "male" and "female", it is often used as it is as two categories. FIG. 3 shows an example of the analysis target data which has been converted into the category as described above.

The feature rules are written as follows.

"If gender = male and age = 56 or more then purchase amount = large", that is, an if-then rule composed of a set of a target attribute and a categorized value. If of feature rule
The attribute appearing in the part is called "condition item", and the attribute appearing in the then part is called "conclusion item". An attribute cannot be both a condition item and a conclusion item at the same time. The feature rule has an evaluation value. In general, set the feature rule to "if A then B"
, The evaluation value is defined by the following equation.

[0023]

P (A) ^ a log {P (B | A) / P (B)} where P (A) and P (B) are the conditions A and the conditions in the whole data to be analyzed, respectively. P (B | A) is the ratio of records that satisfy both conditions A and B among records that satisfy condition A. The index a is a heuristically determined positive constant. Also, the following expression may be used as another definition of the evaluation value.

[0024]

[Equation 3] P (A) ^ a P (B | A) log {P (B | A) / P (B)} In each of Equations (1) and (2), a record satisfying a condition appearing in a rule, and analysis The evaluation value can be calculated by knowing the number of records of the entire target data.

The feature rule discovery is a process for finding a feature rule having a large evaluation value based on the rule evaluation value defined above. At this time, the upper limit of the number of feature rules to be discovered, the fields and their values as conclusion items, the plurality of fields as condition item candidates, and the upper limit of the number of condition items appearing in one feature rule are given by the analyst. It is assumed that FIGS. 4, 5, and 6 show the procedure for finding rules. 4 shows a processing procedure in the data storage device, FIG. 5 shows a processing procedure in the data analysis device, and FIG. 6 shows a processing procedure in the analysis result totaling device. Taking the data shown in FIG. 3 as an example of the analysis target data, the conclusion item is “purchase amount”, the value is “large”, and the candidate condition items are “sex”, “age”, “occupation”, and the condition of the condition item. The upper limit of the number is 2. In addition, "sex" is a binary value of "male" and "female", "age" is a ternary value of "35 and under", "36 to 55", "56 and over",
“Occupation” can take two values, “Yes” and “No”.

First, a data storage device, a data analysis device,
Assignment setting processing 4 in all of the analysis result totaling devices
01, 501 and 601 are performed. In the assignment setting process, counters corresponding to all possible feature rules are prepared according to the designated condition item candidates and the upper limit of the number of conditions. When the above condition item candidates and the upper limit of the number of conditions are used, there are 23 possible combinations of condition items, and the 23 feature rules shown in FIG. 7 are possible. The above-described P (A), P (B | A), and P (B) are required for calculating the evaluation value of the feature rule. Here, since the field and the value which are the conclusion items are specified as one, P (B) is the same for all the feature rules. Therefore, two counters for knowing P (A) and P (B | A) are prepared for each feature rule. These counters are distributed and assigned to a plurality of data analyzers.

Which feature analysis counter is assigned to which data analyzer is one of the data analyzers.
Alternatively, it is determined by either the analysis result totaling device or the data storage device. Then, each identifier of the data analyzer to which the counter is assigned, or
The number of data analyzers to which the counter has been assigned is notified to the data storage device. Each data analyzer prepares a counter according to the assignment.

In the data storage device, the assignment setting processing 40
After 1, the process waits for reception of a preparation completion signal from all data analyzers and analysis result totalization devices to which the counters are assigned (processing 402). In the process 403, it is confirmed whether or not there is untransmitted data. If there is untransmitted data, the process proceeds to the record transmission process 404. If all the data has been transmitted, the process proceeds to the transmission end process 405. In the record transmission process 404, one record of the data not yet transmitted is transmitted to the communication path, and the process returns to the process 402.
In the transmission end process 405, a signal indicating that all data has been transmitted is transmitted to the communication path.

With the above, the processing in the data storage device ends.

In the data analyzer, an assignment setting process 50 is performed.
After step 1, a signal indicating that preparation for receiving the data to be analyzed is completed is transmitted to the data storage device (step 50).
2). In the data reception waiting process 503, the process waits for data to be received from the data storage device.
Proceed to process 504. In the process 504, it is determined whether or not the received data is a signal indicating the end of the data.
Otherwise, the process proceeds to the counter update processing 505. In the counter update processing 505, it is assumed that the received data is a record of the data to be analyzed, and the value of the field is evaluated. If the value of the field matches the condition of the feature rule assigned to the data analyzer, the value of the corresponding counter is updated. In general, one record may match the conditions of multiple feature rules. That is,
In processing one record, the counters of a plurality of feature rules are updated. In the data reception preparation process 506, the record received in the process 503 is discarded, preparations are made for receiving the next record, and the process returns to the process 502. In the rule evaluation process 507, the evaluation values of the rules assigned to the data analyzer are calculated based on the value of the counter, and the number of feature rules specified as the upper limit of the number of feature rules to be discovered is determined in descending order of the evaluation value. Take out. However, feature rules with an evaluation value smaller than 0 are not extracted. Therefore, only a small number of feature rules may be extracted from the specified number. Thereafter, the process proceeds to a process 508 of waiting for an instruction from the analysis result totaling device. If the instruction from the analysis result totaling device is a termination instruction (process 509), the process in the data analysis device is terminated. When the instruction from the analysis result totaling device is an analysis result transmission instruction (process 510)
Transmits the extracted feature rule and its evaluation value to the analysis result totaling device (process 511), and waits for an instruction 5
Return to 08.

In the analysis result totaling device, assignment setting processing 6
01, a signal indicating that preparation for receiving the analysis target data is completed is transmitted to the data storage device (processing 60).
2). In the data reception waiting process 603, the process waits for data to be received from the data storage device.
Proceed to process 604. In the process 604, it is determined whether or not the received data is a signal indicating the end of the data.
Otherwise, the process proceeds to the reception preparation processing 605. In the reception preparation processing 605, the data received from the data storage device is discarded, and preparations for receiving the next data are made.
Return to In the analysis result collection processing 606, an analysis result transmission instruction is sequentially sent to all the data analyzers, and the feature rules extracted by the respective data analyzers and their evaluation values are collected. In the analysis result totaling process 607, the number of feature rules specified as the upper limit of the number of feature rules to be found are taken out of the collected feature rules in descending order of the evaluation value, and this is set as a feature rule discovery result. . In the end instruction processing 608, an end instruction is transmitted to all data analyzers. Thus, the processing in the analysis result totaling device is completed.

FIG. 8 shows the configuration of the second embodiment of the present invention. In the first embodiment, a bus-type communication path is used, whereas in the second embodiment, a ring-type communication path is used.
In a ring communication channel, all devices have a transmission terminal and a reception terminal, and the transmission terminal of the device is connected to the reception terminal of another device via a unidirectional communication channel. When transmitting data, any of the data storage device, the data analysis device, and the analysis result totaling device adds an identifier of the transmitting device to the data and transmits the data from the transmission terminal. When receiving data, the data is received from the receiving terminal. If the identifier added to the data is your own, the data is discarded. If the identifier is not your own, the data is imported into the device as needed, and the data and the identifier are From the transmission terminal. As described above, data transmitted from a certain device is sequentially transferred between all devices and returns to the transmission source device, where the transfer ends.

Therefore, as in the case of the bus-type communication path, data is transmitted to all the other devices by one transmission from the transmission source device.

FIG. 9 shows the configuration of the third embodiment of the present invention. In the second embodiment, a star-type communication path is used. In the star communication path, all devices are connected to the line concentrator 901 via a bidirectional communication path. The concentrator has a plurality of connection terminals and has a function of transmitting a signal received at one connection terminal from all other connection terminals.
Therefore, as in the case of the bus-type communication path, data is transmitted to all other devices by one transmission from the transmission source device.

FIG. 10 shows the configuration of the fourth embodiment of the present invention. A data management device 1004, a plurality of data analysis devices 1002, and an analysis result totaling device 1003 are connected to one shared memory 1005, and the data management device 1004 is connected to a data storage device 1001 via a communication path. The data to be analyzed is stored in the data storage device.

The shared memory is divided into a data section 1006, a counter section 1007, and an analysis result section 1008.

FIGS. 11, 12, and 13 show the procedure for finding a feature rule in this embodiment. 11 is a procedure in the data management device, FIG. 12 is a procedure in the data analysis device,
FIG. 13 shows a procedure in the analysis result totaling device. First,
In all of the data management device, the data analysis device, and the analysis result totaling device, the assignment setting processes 1101, 1201,
Step 1301 is performed. In the assignment setting process, counters corresponding to all possible feature rules are prepared in the counter section of the shared memory in accordance with the designated condition item candidates and the upper limit of the number of conditions. In addition, for each of the feature rules, a data analyzer that is in charge of the counter update processing is determined. Which data analyzer is assigned the counter update process of which feature rule is determined by one of the data analyzers or the analysis result totaling device. Also, an area for writing each analysis result of the data analyzer is prepared in the analysis result section. In the data management device, a plurality of record buffers 1009 are prepared in the data section of the shared memory. One record buffer includes a record area in which one record can be stored, and a flag area corresponding to each of the plurality of data analyzers to which the counter update processing is assigned. Each of the flags takes a state of "data invalid" or "data valid". If all the flags of one record buffer are “data invalid”, the record buffer is said to be “empty”. In the initial state, all record buffers are empty. In addition, separately from the record buffer, the data end flag 10
Prepare 10 The data end flag can take one of two states of "true" and "false", and the initial state is "false".
It is.

In the data management device, one record of the data to be analyzed is input from the data storage device (step 110).
2). At this time, if a signal indicating the end of the data has been received from the data storage device, the process proceeds to process 1106, otherwise proceeds to process 1104 (process 1103). In step 1104, the record buffer of the shared memory is scanned to find one empty record buffer. If this is not found, the scan is repeated until found. If an empty record buffer is found, the process proceeds to processing 1105. In the process 1105, the record input from the data storage device is stored in the record area of the record buffer found in the process 1104, and all flags of the record buffer are set to values indicating "data valid". Then, the process returns to the process 1102. In process 1106, the data end flag of the data section is set to “true”. Thus, the processing in the data management device ends.

The data analyzer scans the record buffer of the shared memory and searches for one record buffer whose flag corresponding to its own data analyzer is "data valid" (process 1202). If this is found, process 120
If no, the process proceeds to step 1205. In process 1203, a record is read from the record buffer found in process 1202, and if the value of the field matches the condition of the feature rule assigned to the data analyzer, the value of the counter of the feature rule is updated. I do. In the process 1204, the flag corresponding to the own data analyzer in the record buffer found in the process 1202 is updated to “data invalid”. Then, processing 1
Return to 202. In processing 1205, the data end flag in the shared memory is checked, and if the state is “true”, processing 1205 is executed.
Then, if the status is “false”, the process returns to step 1202.
In processing 1206, the evaluation values of the rules assigned to the data analysis device are calculated based on the values of the counters, and the number of feature rules specified as the upper limit of the number of feature rules to be discovered are extracted in descending order of the evaluation values. . However, feature rules with an evaluation value smaller than 0 are not extracted. In processing 1207, the extracted feature rule and its evaluation value are written in the analysis result area of the shared memory. Thus, the processing in the data analyzer ends.

The analysis result totaling apparatus checks the analysis result area of the shared memory, and waits until the analysis results of all the data analyzers are written in the analysis result area (step 1302).
When the analysis results of all data analyzers are completed, the number of feature rules specified as the upper limit of the number of feature rules to be discovered are extracted from the feature rules written in the analysis result area in descending order of evaluation value. This is set as the result of feature rule discovery (process 1303). Thus, the processing in the analysis result totaling device is completed.

In the fourth embodiment described above, the counter is stored in the shared memory. However, when each of the data analyzers has the local memory 1401, the counter is stored in the local memory. (FIG. 14).

[0042]

According to the present invention, when a plurality of processing devices are operated in parallel to execute data analysis processing such as data mining for analyzing a large amount of data from various aspects, the amount of data to be analyzed is reduced. The number of transfers can be reduced, and the amount of main memory required in each processing device can be reduced. Further, since a data storage method designed so as to be suitable for a mass data analysis method is not particularly required, data stored in a general database system can be analyzed.

[Brief description of the drawings]

FIG. 1 is a configuration diagram of a first embodiment.

FIG. 2 is a flowchart showing an outline of an analysis method.

FIG. 3 is a diagram showing an example of analysis target data.

FIG. 4 is a flowchart showing processing in the data storage device of the first embodiment.

FIG. 5 is a flowchart showing processing in the data analyzer of the first embodiment.

FIG. 6 is a flowchart showing processing in the analysis result totaling device of the first embodiment.

FIG. 7 is a diagram listing all possible feature rules.

FIG. 8 is a configuration diagram of a second embodiment.

FIG. 9 is a configuration diagram of a third embodiment.

FIG. 10 is a configuration diagram of a fourth embodiment.

FIG. 11 is a flowchart showing processing in the data management device of the fourth embodiment.

FIG. 12 is a flowchart showing processing in the data analyzer of the fourth embodiment.

FIG. 13 is a flowchart showing processing in the analysis result totaling apparatus of the fourth embodiment.

FIG. 14 is a configuration diagram in which a counter is placed in a local memory in the fourth embodiment.

[Explanation of symbols]

101: data storage device, 102: data analysis device, 1
03: Analysis result totaling device, 104: Bus type communication path, 10
04: data management device, 1005: shared memory, 10
06: data section in shared memory, 1007: counter section in shared memory, 1008: analysis result section in shared memory, 1009: record buffer, 1010: end flag, 1401: counter in local memory.

Continued on the front page (72) Inventor Yukiyasu Ito 5030 Totsuka-cho, Totsuka-ku, Yokohama-shi, Kanagawa Prefecture F-term in the Software Division, Hitachi, Ltd. F-term (reference) 5B075 KK02 PQ05 QS05

Claims (4)

[Claims] In data analysis using a plurality of data analysis means for data stored in one or a plurality of data storage means, the plurality of data analysis means may store the same data from the data storage means. And performing analysis in each of the plurality of data analysis means, and combining the analysis results in each of the plurality of data analysis means to obtain an overall analysis result. 2. The data analysis method according to claim 1, wherein the plurality of data analysis units share the analysis target data output once by the data storage unit. 3. The method according to claim 1, wherein a plurality of data analysis units use a shared storage unit, and the plurality of data analysis units output respective analysis results to the shared storage unit. 3. The data analysis method according to 2. 4. A storage medium storing a computer program for executing the parallel analysis of variance method according to claim 1, 2 or 3 on a computer.