JP6775022B2 - Learning data processing device and method - Google Patents

Description

The present invention relates to a method of processing data acquired in IT system operation management.

With the spread of virtualization mechanisms and the emergence of new system provision forms such as the cloud, the operation management of IT systems has become complicated. In addition, with the explosive increase in the amount of data handled by IT systems, the scale of IT systems is expanding year by year, and the number of objects handled by management software that manages IT systems (for example, the number of volumes provided by storage devices) is also increasing. It is increasing. It is required to manage an IT system having a complicated and large amount of data at a low management cost.

In order to reduce the management cost of the IT system, it is conceivable to automate the management. Machine learning technology is one of the technologies that can be used in the automation of IT system management. By collecting various information in each object of the IT system and learning it as learning data, it is possible to identify the function that best applies to the learning data regarding the relationship between any element in the IT system and other elements. It becomes.

For example, Non-Patent Document 1 states that the response performance of processing executed in an IT system is predicted from parameters related to settings for processing such as the number of parallel processes and parameters related to processing targets such as the size of data to be processed. It describes the technique for finding the functions that enable it by learning. By using this function, for example, the response time of the process can be estimated from the number of parallel processes and the data size, so that the execution schedule of the process can be determined. In addition, since the number of parallel processes required to obtain the required response performance can be estimated, the amount of resources required to obtain the required response performance can be estimated.

In addition, a cloud service in which an IT system is provided by a business operator via the Internet and the user is charged according to the usage is widespread. As a form of cloud service, there are forms such as IaaS (Infrastructure as a Service), PasaS (Platform as a Service), and SaaS (Software as a Service).

In addition, cloud services are not suitable for use as data storage locations for highly confidential data and applications that require real-time performance, but management operations can be separated from the original functions of the IT system. It is also suitable for cloud services because of the need to pay for usage.

Against this background, there are movements such as providing SaaS (Software as a Service), a management software that has been operating on-premises, and undertaking part of the operation work as a service. Patent Document 1 discloses a method in which a business operator monitors on-premises storage via a network and performs maintenance work such as configuration change and disk replacement when an event occurs.

Japanese Unexamined Patent Publication No. 2006-107080

Statistics-driven workload modeling for the cloud, Archana Ganapathi, University of California at Berkeley, ICDE 2010

The techniques disclosed in Patent Document 1 and Non-Patent Document 1 are intended to be used in an environment with relatively little change, and are not intended to be frequently changed in the system configuration. On the other hand, with the spread of virtualization mechanisms and the emergence of new system provision forms such as the cloud, it is becoming relatively easy to change the IT system configuration, and it is thought that the frequency of system configuration changes will increase. Be done.

A large amount of learning data is required to improve the accuracy of machine learning. In IT system management, it is necessary to acquire various historical information such as performance information and capacity information from each object of the IT system over a long period of time. However, when a configuration change occurs in the IT system, it becomes necessary to perform long-term learning again after the configuration change occurs. Then, it is conceivable that the accuracy of machine learning does not improve for a while after the configuration change, and efficient management work cannot be performed.

An object of the present invention is to provide a technique for improving the accuracy of machine learning for a system whose configuration is changed.

According to one aspect of the present invention, the learning data processing device is between an information collecting unit that acquires learning data from the components of the target system while the target system is operated and the components of the target system based on the learning data. Change the configuration of the target system, the predictive formula generator that generates the predictive formula that expresses the relationship between the objective information and the explanatory information, and the setting change content determination unit that determines the setting contents that change the configuration of the target system. A state in which training data is not sufficiently acquired within the range of the state of the components that can be taken when the setting change content determination unit changes the configuration of the target system, which has a configuration change unit, is regarded as a data shortage state. Extract and determine the setting contents to change the configuration of the target system so that the components are in the data shortage state, the configuration change unit changes the configuration of the target system according to the determined setting content, and the information collection unit Acquire training data when data is insufficient from the components of the target system.

Since the learning data that will be insufficient when the configuration of the target system is changed in the future can be acquired by temporarily changing the configuration of the target system in advance, the lack of learning data that occurs when the configuration is actually changed Is reduced, and the accuracy of machine learning is improved at an early stage. When machine learning in a target system in which the configuration is changed relatively frequently is applied to this, it is possible to suppress a decrease in the accuracy of machine learning that occurs when the configuration is changed.

It is a figure for demonstrating the outline of the computer system according to Example 1. FIG. It is a block diagram of an example of the computer system which concerns on Example 1. FIG. It is a figure which shows an example of the related information table which concerns on Example 1. FIG. It is a figure which shows an example of the performance history information table 1120 which concerns on Example 1. FIG. It is a figure which shows an example of the configuration information table 1130 which concerns on Example 1. FIG. It is a figure which shows an example of the configuration information table 1130 which concerns on Example 1. FIG. It is a figure which shows an example of the configuration information table 1130 which concerns on Example 1. FIG. It is a figure which shows an example of the prediction formula original information table 1140 which concerns on Example 1. FIG. It is a figure which shows an example of the prediction formula table 1150 which concerns on Example 1. FIG. It is a flowchart of the process which generates the prediction formula which concerns on Example 1. FIG. It is a flowchart of the process of determining the setting change content for acquiring the learning data which concerns on Example 1. It is a flowchart of the process which executes the setting change for learning data acquisition. It is a figure which shows an example of the business characteristic management table 1800 which concerns on Example 2. FIG. It is a flowchart of the process which executes learning data sharing which concerns on Example 2.

Some embodiments will be described with reference to the drawings. It should be noted that the examples described below do not limit the inventions claimed in the claims, and all of the elements and combinations thereof described in the examples are essential for the means for solving the invention. Is not always. In these drawings, the same reference numerals indicate the same components throughout the drawings. In the following description, the information of the present invention will be described by expressions such as "aaa table", but these information may be expressed by other than the data structure such as a table. Therefore, the "aaa table" and the like may be referred to as "aaa information" to show that it does not depend on the data structure. Further, when explaining the contents of each information, the expressions "identification information", "identifier", "name", and "ID" are used, but these can be replaced with each other.

In the following explanation, "program" may be used as the subject, but the program performs the processing defined by being executed by the processor in the memory and communication port (communication device, management I / F, data I / F). Since it is performed while using, the explanation may be based on the processor. Further, the process disclosed with the program as the subject may be a process performed by a computer such as a management server (management computer) or an information processing device. In addition, part or all of the program may be realized by dedicated hardware. In addition, various programs may be installed in each computer by a program distribution server or a storage medium that can be read by the computer.

Hereinafter, a set of one or more computers that manages a computer system and displays display information of the present invention may be referred to as a management system. When the management server displays display information, the management server is the management system. The combination of the management server and the display computer is also a management system. Further, in order to speed up and improve the reliability of the management process, a plurality of computers may realize the same processing as the management server. In this case, the plurality of computers (for display when the display is performed by the display computer). (Including the computer) is the management system.

The computer system according to this embodiment will be described.

FIG. 1 is a diagram for explaining the outline of the computer system according to the first embodiment. The operation described here is mainly executed by the setting change content determination program 1180.

(1) The setting change content determination program 1180 first refers to the prediction formula table 1150 and the configuration information table 1130, and extracts the parameters corresponding to the configuration information indicating the configuration of the computer system from the parameters of the prediction formula. (2) Subsequently, the setting change content determination program 1180 refers to the prediction formula source information table 1140, and identifies a range in which the training data information is insufficient within a range that can be taken in terms of the configuration of the extracted parameters. (3) Subsequently, the setting change content determination program 1180 determines the value of the parameter so as to set the extracted parameter in the specified range, and changes the setting of the parameter. (4) By operating the computer system with the settings changed, the missing learning data can be acquired.

FIG. 2 is a configuration diagram of an example of a computer system according to the first embodiment. The computer system according to this embodiment includes one or more management servers 1000, one or more storage devices 2000, and one or more servers 3000. The server 3000 and the storage device 2000 are connected to each other via a SAN (Storage Area Network) 4000. A specific example of SAN is Fiber Channel. The management server 1000, the storage device 2000, and the server 3000 are connected to each other via the management network 5000.

The management server 1000 includes a memory 1100, a communication device 1200, a processor 1300, an output device 1400, an input device 1500, and a storage device 1600. These are connected to each other via the internal bus 1700 in the management server 1000.

The memory 1100 includes a related information table 1110, a performance history information table 1120, a configuration information table 1130, a prediction formula source information table 1140, a prediction formula table 1150, an information collection program 1160, a prediction formula generation program 1170, and a setting change content determination program 1180. , The configuration change program 1190 is stored.

The communication device 1200 is a device for connecting the management server 1000 to the management network 5000. The management server 1000 can communicate with a program running on the server 3000 through the management network 5000. The processor 1300 executes various programs expanded on the memory 1100. The output device 1400 is a device that outputs a processing result executed by the management server 1000, such as a display. The input device 1500 is a device for the administrator to input an instruction to the management server 1000, such as a keyboard. The storage device 1600 is an HDD (Hard Disk Drive), an SSD (Solid State Drive), or the like that stores information.

In the example shown in FIG. 2, various programs and tables are stored in the memory 1100, but may be stored in the storage device 1600 or another storage medium (not shown). In this case, the processor 1300 reads the target program on the memory 1100 when the program is executed, and executes the read program.

Further, the above-mentioned programs and tables may be stored in the memory 2100 of the storage device 2000, and the storage device 2000 or the physical server 3000 may execute the stored program. In addition, another server 3000 or another device such as a switch (not shown) may store and execute the above-mentioned programs and tables.

The storage device 2000 includes a memory 2100, a logical volume providing unit 2200, a disk I / F controller 2300, a management I / F 2400, a processor 2500, and a data I / F 2600. These are connected via a communication path 2700 such as an internal bus in the storage device 2000. The memory 2100 has a disk cache 2110. In addition, the memory 2100 stores the configuration performance information collection program 2120. The disk cache 2110 is a storage area for temporarily storing information. The configuration performance information collection program 2120 is a program for transmitting and receiving management information and performance information of the storage device 2000 to and from the management server 1000. The configuration change program 2130 is a program called from the configuration change program 1190 of the management server 1000 to change the configuration of the storage device 2000.

The logical volume providing unit 2200 includes a disk pool 2220 composed of a physical area 2230, logically divides the storage area of the disk pool 2220, and provides the logically divided storage area as the volume 2210. Here, the physical area 2230 is a physical disk, a parity group composed of a plurality of physical disks, or the like. It is possible to access the physical area from a device other than the storage device 2000 via the volume 2210.

The physical area 2230 is assigned a physical area number, the disk pool 2220 is assigned a disk pool number, and the volume 2210 is assigned a volume number. As a result, the storage device 2000 can uniquely identify the physical area 2230, the disk pool 2220, and the volume 2210, respectively.

In the example shown in FIG. 2, the disk pool 2220 (POOL1) composed of one physical area (parity group PG1) is logically divided, and one volume 2210 (Vol1) is a device other than the storage device 2000, for example, a server. Provided to 3000.

The disk I / F controller 2300 is an interface device for connecting to the logical volume providing unit 2200. The management I / F 2400 is an interface device for connecting to the management network 5000. The processor 2500 executes a program expanded on the memory 2100.

The data I / F 2600 is an interface device for connecting to the SAN 4000. In the example shown in FIG. 2, the configuration performance information collection program 2120 and the configuration change program 2130 are stored in the memory 2100, but are stored in another storage device (not shown) or another storage medium (not shown). You may. In this case, the processor 2500 reads the configuration performance information collection program 2120 and the configuration change program 2130 on the memory 2100 at the time of processing execution, and executes the read programs.

Further, the logical volume providing unit 2200 may create the entire storage area of one disk pool 2220 as one volume 2210. Further, the logical volume providing unit 2200 may be a physical area 2230 other than the parity group, for example, a physical disk itself or a storage medium such as a flash memory.

Server 3000 is a physical server with memory 3100, data I / F 3200, processor 3300, and management I / F 3400. These are connected to each other via a communication path 3500 such as an internal bus of the server 3000.

The memory 3100 stores the configuration information collection program 3110, the business program 3120, and the configuration change program 3130. The configuration information collection program 3110 is a program for transmitting and receiving management information, performance information, and the like of the server 3000 to and from the management server 1000. The business program 3120 is a program for realizing the business executed by 3000, and is, for example, a DBMS (Data Base Management System), a file system, or the like. The configuration change program 3130 is a program reserved from the configuration change program 1190 of the management server 1000 and for changing the configuration of the server 3000.

The server 3000 uses the volume 2210 provided by the storage device 2000 to execute various tasks. In the example shown in FIG. 2, various programs are stored in the memory 3100, but may be stored in another storage device (not shown). In this case, the processor 3300 reads the target program on the memory 3100 at the time of processing execution, and executes the read program. In the example shown in FIG. 2, the server A and the storage device A are connected to each other via the SAN 4000. The connection between the storage device 2000 and the physical server 3000 is not limited to those directly connected via Fiber Channel, and may be connected via network devices such as one or more Fiber Channel switches. Further, the connection between the storage device 2000 and the physical server 3000 may be any network for data communication, and may be an IP (Internet Protocol) network.

FIG. 3 is a diagram showing an example of a related information table according to the first embodiment. The related information table stores related information indicating the managed object whose purpose information is its performance and the managed object logically associated with the managed object. Objects are components of a computer system. The component includes a physically existing component and a logically defined component. As an example, the related information table 1110 is a physical existing on an I / O (input / output) path from the business program 3120 running on the server 3000 to the physical area constituting the volume used by the server 3000. / Manages information indicating virtual devices and devices, that is, information indicating logical relationships between devices and devices based on I / O routes. Here, the logical relationship is "volume" and "pool constituting the volume", "volume" and "processor in charge of I / O processing to the volume", "volume" and "I / O to the volume". It is stored based on the settings such as "cache that temporarily stores".

Related information The table 1110 has fields for device ID 1111, volume ID 1112, processor ID 1113, cache ID 1114, pool ID 1115, and physical area ID 1116.

An identifier for uniquely identifying the storage 2000 is stored in the device ID 1111. The volume ID 1112 stores an identifier for uniquely identifying the volume 2210. The processor ID 1113 stores the identifier of the processor 2500 in charge of processing the volume indicated by the volume ID 1112. The cache ID 1114 stores an identifier uniquely indicating the disk cache 2110 in which the processing for the volume indicated by the volume ID 1112 is cached. The pool ID 1115 stores an identifier for uniquely identifying the disk pool 2220 in which the volume 2210 is created. The physical area ID 1116 stores an identifier for uniquely identifying the physical area 2230 constituting the disk pool, for example, a parity group or a disk. Information collected from the computer system is stored in the fields of each of the above columns. The method of collecting and storing information is not particularly limited.

Here, the related information table 1110 according to the present embodiment includes information on the processor 2500, the disk cache 2110, the disk pool 2220, and the physical area 2230 as managed objects related to the device ID 1111 and the volume ID 1112. However, the present invention is not limited to this. Any managed object in the IT system can be handled in the same way.

As another example, a drive for uniquely identifying the mount point of the server 3000, which is a managed object used for business access, the server 3000 accesses the volume 2210 indicated by the volume ID 1136. An identifier for identifying other managed objects including physical and virtual objects such as server data I / F for uniquely identifying the data I / F 3200 of the server 3000 used at the time may be stored. ..

Further, information such as switch data I / F may be included, and information of a business program (DBMS or the like) on the server 3000, which is a business server, may be associated and stored. Further, information of a processing unit executed by the business program may be stored in association with each other. For example, processing A in the business program may be associated with a server used for executing the processing, a CPU of the server, memory, and the like. You may store it.

FIG. 4 is a diagram showing an example of the performance history information table 1120 according to the first embodiment. Performance history information table 1120 The performance history acquired from each managed object by the operation of the computer system is stored. The performance history information table 1120 manages performance information regarding the performance of the managed object, for example, the volume 2210, the disk pool 2220, and the like in the storage device 2000. An entry can be added to the performance history information table 1120.

The performance history information table 1120 includes fields for time 1121, device ID 1122, device ID 1123, metric 1124, and performance value 1125.

At time 1121, data on the date and time when the information was collected from the managed object is stored. An identifier (device ID) that uniquely identifies the device is stored in the device ID 1122. The device ID 1123 stores an identifier (device ID) for uniquely identifying the device for which performance information is to be acquired.

The metric 1124 stores information indicating the type of performance information, such as the CPU usage rate, the number of I / Os (IOPS) per unit time (for example, 1 second) for the storage device, and the response time to the request. In the performance value 1125, the value of the performance information of the type indicated by the metric 1124 of the device indicated by the device ID 1123 is acquired from the device including the device and stored.

Here, in the performance history information table 1120 shown in FIG. 4, the storage volume 2210, the processor 2500, and the disk cache 2110 are listed as the devices for which the performance information is acquired, which are indicated by the device ID 1122 and the device ID 1123. Not limited to these. It may be a VM (not shown), a storage data I / F 2600, a server data I / F 3200, a switch or a switch port (not shown), or the like.

Further, FIG. 4 shows, as an example of the metric, the response performance to the request, the CPU usage rate, the cache usage rate, the IOPS, the response time to the request, and the like, but the present invention is not limited thereto. I / O busy rate, transfer rate, throughput, buffer hit rate of database management software, number of records inserted, updated or deleted, response time of Web server, free space or utilization rate of file system or disk, input / output Other performance indicators such as data volume, number of network interface errors, buffer overflows, and frame errors may be used as metrics.

5A, 5B, and 5C are diagrams showing an example of the configuration information table 1130 according to the first embodiment. 5A and 5B show a state before the operation is executed by the configuration change program 1190 in step 301 of FIG. 10, which will be described later. FIG. 5C shows the state after the operation is executed by the configuration change program 1190 in step 301 of FIG.

The configuration information table 1130 stores the configuration information of the managed object. For example, the cache size of the disk cache 2110, which is the configuration information for the storage device 2000, which is a managed object, is stored. In addition, the disk configuration of the physical area (parity group) 2230 is stored. Entries are added to the configuration information table 1130 by common means.

The configuration information table 1130 includes fields for device ID 1131, device ID 1132, metric 1133, and value 1134. The device ID 1131 stores an identifier for uniquely identifying the device. The device ID 1132 stores an identifier for uniquely identifying the device for which the configuration information is to be acquired. Information indicating the type of configuration information such as storage capacity and processing capacity is stored in the metric 1133. The value 1134 stores values for the type of configuration information indicated by the metric 1133 for the device indicated by device ID 1132. This value is obtained from the device including the device.

Here, in the configuration information table 1130 shown in FIGS. 5A to 5C, the devices indicated by the device ID 1131 and the device ID 1132 are the targets for acquiring the configuration information. Here, the disk cache 2110 (Cache1) and the physical area 2230 (PG1, PG5) of the storage 2000 are mentioned as the devices for which the configuration information is acquired, but the devices are not limited thereto. The configuration information of other managed objects may be retained. Further, as an example of the metric, the cache size, the RAID level of the parity group, and the disk type are mentioned here, but the metric is not limited thereto.

FIG. 6 is a diagram showing an example of the prediction formula source information table 1140 according to the first embodiment. The prediction formula source information table 1140 is a table for managing the information that is the basis for generating the prediction formula. In the prediction formula source information table 1140, the managed object to be predicted and its parameters, and other managed objects and their parameters related to the managed object to be predicted on the I / O path are managed. The managed object to be predicted and its parameters are the objective information of the prediction formula, and the related managed objects and their parameters are explanatory information.

The prediction formula source information table 1140 includes fields for time information 1141, purpose information 11411, and related information 11412. The time information 1141 stores data on the date and time when the information was collected from the managed object. The management target object identification information to be predicted and the parameter values of the management target object are stored in the target information 11411. The related information 11412 stores information on the parameter values of other managed objects that are related to the managed object to be predicted on the I / O path. In this embodiment, the device ID 1142, the volume ID 1143, and the volume response performance 1144 are stored as the target information 11411. Related information 11412 includes fields of Processor Busy 1145, Cache Use 1146, Cache Size 1147, Pool Busy 1148, and PG number 1149.

An identifier (device ID) that uniquely identifies the device is stored in the device ID 1142. The volume ID 1143 stores an identifier for uniquely identifying the managed object. The volume response performance 1144 stores information on the time required from the reception of the I / O request on the volume to the completion of processing. Here, the response performance of the volume is given as an example of the target information 11411, and the Processor Busy1145, the Cache Use 1146, and the like are given as an example of the related information, but the present invention is not limited thereto.

Among the values and information stored in the table shown in FIG. 6, the information in which the time information 1141 is 10:01 and the information in 10:02 are before the operation execution by the configuration change program 1190 in step 301 of FIG. The information in which the time information 1141 is 15:10 and the information in 15:11 indicate the state after the operation is executed by the configuration change program 1190 in step 301 of FIG.

FIG. 7 is a diagram showing an example of the prediction formula table 1150 according to the first embodiment. The prediction formula table 1150 is a table for managing information representing the prediction formula. The prediction formula table 1150 stores the metrics used in the prediction formula, the coefficients related to each metric, and the like. The prediction formula h, specifically, objective information = explanatory information 1 + explanatory information 2 + explanatory information 3 + explanatory information 4 ... Can be expressed. More specifically, it is the information of the function obtained by learning that the response performance of the volume 1 of the storage A = coefficient 1 × Processor Busy + coefficient 2 × Cache Size + coefficient 3 × Pool Busy + coefficient 4 × number of PGs.

The prediction table 1150 includes fields for objective information 11511 and explanatory information 11512. The objective information 11511 stores the identification information of the managed object to be predicted and the value of the parameter of the managed object. The explanatory information 11512 stores information on the parameters of other managed objects and their values that can explain the values of the parameters of the managed object to be predicted. In this embodiment, the device ID 1151, the device ID 1152, and the metric 1153 are managed as the target information 11511, and the processor Busy 1154, the Cache Size 1155, the Pool Busy 1156, the PG number 1157, and the fields representing the coefficients for each metric are used as the explanatory information 11512. Including. Here, the response performance of the volume is given as an example of the target information 11411, and the Processor Busy1154, Cache Size1155, and the like are given as an example of the related information, but the present invention is not limited thereto.

Further, here, the prediction formula is assumed to be a formula expressing a linear relationship, and the prediction formula table 1150 is assumed to represent a regression analysis formula for identifying the linear relation that best fits the data, but the present invention is not limited to this. .. As another example, the prediction formula may be a polynomial, and the prediction formula table 1150 may manage information representing the polynomial.

Next, each process executed by the management server 1000 will be described.

FIG. 8 is a flowchart of a process for generating a prediction formula according to the first embodiment. Prediction formula generation is to collect various information in each object as learning data and learn it to identify the function that best applies to the training data regarding the relationship between the target element and other elements. is there. This prediction formula generation process is performed by the processor 1300 of the management server 1000 executing the prediction formula generation program 1170 expanded on the memory 1100. A specific example of this flowchart is shown below.

First, the prediction formula generation program 1170 refers to the related information table 1110 illustrated in FIG. 3 and identifies the component to be generated for the prediction formula and the component related thereto (step 101). Here, the component of the prediction formula generation target is selected by any method, such as being selected by the user or automatically selected by the prediction formula generation program (for example, executing for all volume response performances). And may be specified. Further, the timing at which the prediction formula generation program 1170 is started is arbitrary, such as periodic execution or execution at an arbitrary timing specified by the user.

Here, as a specific example, it is assumed that the volume represented by the volume ID “Vol1” is selected by the user as the target for generating the prediction formula. In this case, the prediction formula generation program 1170 identifies Processor1, Cache1, Pool1, and PG1 as the components related to Vol1 (Volume1) from the information stored in the related information table 1110 of FIG.

Returning to FIG. 8, the prediction formula generation program 1170 refers to the performance history information table 1120 illustrated in FIG. 4, and sets the target component for generating the prediction formula, and as related to the component, in step 101. Acquire the performance history information of the specified component (step 102). For example, the response time of Volume 1 acquired at 10:01 is 10.2 msec, the usage rate of Processor 1 (Busy%) is 40%, the usage rate of Cake 1 (Usage%) is 80%, and I per unit time of Pool 1. Performance information that the number of times / O was 700 IOPS and the usage rate (Busy%) was 35% is acquired.

Next, the prediction formula generation program 1170 refers to the configuration information table 1130 illustrated in FIGS. 5A and 5B, and acquires the component of the prediction formula generation target and the component information of the component specified in step 101 (step). 103). For example, from FIG. 5A, configuration information that the size of Cache1 of the storage A is 8 GB is acquired. Further, from FIG. 5B, configuration information such as, for example, the RAID level of the physical area PG1 of the storage A is RAID5 (3D + 1P) is acquired.

Next, the prediction formula generation program 1170 stores the information related to the prediction formula generation acquired in steps 102 and 103 in the prediction formula source information table 1140 illustrated in FIG. 6 (step 104). With reference to FIG. 6, the performance information acquired at, for example, 10:01 is stored in the prediction formula source information table 1140 of Volume 1.

Finally, the prediction formula generation program 1170 generates a prediction formula from the information in the prediction formula source information table 1140 generated in step 104, and stores it in the prediction formula table 1150 illustrated in FIG. 7 (step 105). For example, in the prediction formula table 1150 of FIG. 7, (response performance of Volume 1 of storage A) = 33.76 (coefficient 1) × processor usage rate +7.27 (coefficient 2) × cache size +5.1 (coefficient 3). A prediction formula of × Pool usage rate + 0.80 (coefficient 4) × number of physical region PGs is stored.

The method for generating the prediction formula in step 105 is not particularly limited, and any method may be used, including a general method such as regression analysis. In the case of regression analysis, for example, after setting all the related information 11412 shown in the prediction formula source information table 1140 as explanatory variables, variables having low relevance to the target information are excluded from the explanatory variables. You can generate a prediction formula with. In this embodiment, among the related information stored in the prediction formula source information table 1140 shown in FIG. 6, Cache Use 1146 is excluded from the explanatory variables and is included in the information stored in the prediction formula table 1150 shown in FIG. Not done.

FIG. 9 is a flowchart of a process for determining the setting change contents for acquiring the learning data according to the first embodiment. The setting change content determination process 200 is performed, for example, after the process of generating the prediction formula shown in FIG. This process is performed by the processor 1300 of the management server 1000 executing the setting change content determination program 1180 expanded on the memory 1100.

A specific example of this flowchart is shown below.

First, the setting change content determination program 1180 extracts the metric of the explanatory information 11512 in the prediction formula table 1150 illustrated in FIG. 7, and performs the following processing for each metric.

First, the setting change content determination program 1180 checks whether or not the metric is included in the configuration information table 1130 (step 201). If the metric is not included in the configuration information table 1130, the setting change content determination program 1180 proceeds to process for the next metric in the prediction expression table 1150. When the metric is included in the configuration information table 1130, the setting change content determination program 1180 acquires information in a range that the metric can take (step 202). When the metric is the cache size of the storage, for example, information on the range of values that can be taken as the cache size is acquired in terms of hardware specifications. For example, information that the cache size is in the range of 1 GB to 72 GB is acquired. If the metric is a storage parity group, information on the RAID level range is acquired. For example, information that the possible RAID levels are RAID0 (2D), RAID1 (1D + 1P), and RAID5 (3D + 1P) is acquired. The method for obtaining the possible range of these metrics is not particularly limited. For example, information in a range that can be taken by each metric may be stored in a table (not shown) in advance, and the setting change content determination program 1180 may appropriately acquire necessary information from the table. Alternatively, the setting change content determination program 1180 may issue a request to hardware such as storage and acquire it.

Next, the setting change content determination program 1180 searches for a domain lacking data in the range acquired in step 202 (step 203). Next, it is determined whether or not there is a domain lacking data (step 204), and if it does not exist, the process proceeds to the next metric in the prediction expression table 1150. If there is a missing domain in step 204, the setting change content determination program 1180 generates a parameter for a setting change operation that enables acquisition of data in the missing domain (step 205). ..

For example, focusing on the cache size, which is a metric, it is assumed that there is no data other than when 8 GB is set as the Cache Size 1147 of the prediction formula source information table 1140 shown in FIG. In that case, the setting change content determination program 1180 tries to acquire the data when the setting is set to other than 8GB. For example, the setting change content determination program 1180 generates a parameter for changing the setting of the cache size to 16 GB.

Here, the setting change content determination program 1180 checks whether or not the SLA (Service Level Agreement) is satisfied when the setting of the parameter generated in step 205 is changed, and if the changed parameter does not satisfy the SLA. It may be excluded from the setting range of the parameter. For example, when the cache size of 8GB is changed to 4GB, predetermined requirements (response time within 1 second, etc.) are set as the performance of the volume and the performance of the business application running on the server 3000 that uses the volume. ) Satisfaction is not satisfied, the parameter setting may not be changed to 4GB.

Next, the setting change content determination program 1180 executes the learning data acquisition setting change process (step 206). Step 206 will be described in detail with reference to FIG.

FIG. 10 is a flowchart of a process for executing the setting change for learning data acquisition. The learning data acquisition required setting change process 300 (learning data acquisition required setting change process 206 in FIG. 9) is performed by the processor 1300 of the management server 1000 executing the setting change content determination program 1180 expanded on the memory 1100. It is carried out. A specific example of this flowchart is shown below.

First, the setting change content determination program 1180 requests the configuration change program 1190 to execute the setting change operation, and acquires the execution result (step 301). Next, the setting change content determination program 1180 confirms whether or not a new time entry has been added to the prediction formula source information table 1140 (step 302).

When the entry of the new time is added, the setting change content determination program 1180 acquires the number of acquired data in the target domain in the prediction formula source information table 1140 (step 303), and determines whether or not the data can be sufficiently acquired. Check (step 304).

Here, in order to determine whether or not the training data has been sufficiently acquired, a threshold value for the number of data is set in advance, or the number of explanatory information shown in the prediction formula table is set as a threshold value. You can keep it. When sufficient training data has been acquired, the setting change content determination program 1180 proceeds to the next step 305. If sufficient training data has not been acquired, the setting change content determination program 1180 executes the process again from step 302.

In step 305, the setting change content determination program 1180 requests the configuration change program 1190 to execute the setting change operation returned to before the execution of step 301, and acquires the execution result. If the setting change operation requested in steps 301 and 305 is not successful, this process is interrupted.

By executing the prediction formula generation process 100 shown in FIG. 8 after executing FIG. 10 to acquire sufficient training data, from the prediction formula source information table 1140 in a state where there is no shortage of training data in the new configuration. , It is possible to generate a prediction formula table 1150 showing a prediction formula with high accuracy.

In this embodiment, in steps 201 to 204 of the setting change content determination process 200 shown in FIG. 9, all the domain areas in which data is insufficient are extracted from the range that can be configured, and then shown in FIG. The training data acquisition setting change process acquires the data in the domain where the data is insufficient, and then the prediction formula generation process 100 shown in FIG. 8 generates the prediction formula. However, it is not limited to this. As another example, every time a domain with insufficient data is extracted, the data in that domain is acquired, and if the data can be acquired, a prediction formula is generated at that stage. It may be repeated for the number of domains in which it exists.

A specific example is shown. Here, it is assumed that the possible range of the cache size is 1 GB to 72 GB. For example, a domain in which data is insufficient in the range of possible cache size may be extracted in units of, for example, 1 GB, and a prediction formula may be generated after all the extracted domain data have been acquired. Alternatively, as another example, the process of acquiring data, generating a prediction formula, and proceeding to the next domain may be repeated for each domain in which the data extracted in 1 GB units is insufficient.

Further, in this embodiment, in the setting change content determination process 200 shown in FIG. 9, data is insufficient for all the items included in the configuration information table 1130 among the items included in the prediction formula table 1150. After extracting the defined domain, the training data acquisition setting change process 206 is executed. Therefore, the prediction formula is generated after collecting data for all the items included in the prediction formula table 1150 and the configuration information table 1130. However, it is not limited to this. As another example, it is decided to repeat the process of executing the setting change process for learning data acquisition and the predictive expression generation process for one item included in the predictive expression table 1150 and the configuration information table 1130 and proceeding to the next item. You may.

As described above, according to this embodiment, by actively collecting the learning data that is insufficient in the range that the computer system can take in the configuration, a highly accurate prediction formula is performed at an early stage when the configuration is changed. It is possible to shorten the learning time and implement efficient management based on machine learning technology immediately after the configuration change.

For example, when using a function for predictive monitoring, if the measured value obtained from the IT system is far from the relationship indicated by the function even immediately after the configuration change or even if the configuration is newly constructed, there is a problem with the IT system. It is possible to determine that it has occurred.

In addition, when the function is used to isolate the cause of failure, if the measured value acquired from the IT system is far from the relationship indicated by the function even immediately after the configuration change or even if the configuration is newly constructed, each explanatory information It is possible to determine that the root cause is that there is a high possibility that a problem has occurred in the explanatory information that has the largest fluctuation range. In addition, this makes it possible to automatically identify the root cause immediately when a failure occurs.

Also, when using a function for What-if analysis, even if the configuration is changed immediately or newly constructed, by assigning the value you want to try to the function, it will appear in the function in the situation of the assigned value. It is possible to simulate the value of the metric of.

As described above, according to the present invention, it is possible to obtain effects such as prevention of failure occurrence or failure to satisfy management requirements even immediately after configuration change or newly constructed configuration, and quick failure recovery in the event of failure occurrence. Is possible. The present invention can be applied to the above-mentioned various cloud forms, and can also be applied to a form in which the management software SaaS and operation management work are undertaken as a service.

The computer system according to the second embodiment basically has the same configuration as that of the first embodiment, and performs the same operation. However, the second embodiment is to use not only the related information related to the target information but also the information acquired in the computer system of the business whose target information has characteristics similar to the target business to generate the prediction formula. Different from Example 1.

FIG. 11 is a diagram showing an example of the business characteristic management table 1800 according to the second embodiment. The business characteristic management table 1800 manages business characteristic information for each business unit.

The business characteristic management table 1800 stores data of the business unit 18011 and the business characteristic 18012. In this embodiment, an example is shown in which business units are associated with volumes and information related to I / O such as the number of I / Os and the ratio of each I / O pattern is managed as business characteristics for each business. Referring to FIG. 11, the business characteristic management table 1800 is associated with the fields of the business unit 18011 and the business characteristic 18012. The business unit 18011 includes the volume ID 1801. The business characteristics 18012 include fields of I / O number 1802, I / O increase / decrease rate 1803, high frequency access 1804, and I / O pattern 1805.

The volume ID 1801 stores an identifier for uniquely identifying the volume 2210. The I / O number is recorded in the I / O number 1802. For example, record the average value or median value of IOPS of the previous month. In the I / O increase / decrease rate 1803, the rate of how much IOPS has changed in the past fixed period is recorded. For example, in half a year or one year, the average of IOPS for one month is calculated, and the rate of increase / decrease of the average value of each month with respect to the average value of the previous month is calculated. In the I / O pattern 1805, the generation ratio of each I / O pattern of Random Read, Random Write, Sequential Read, and Sequential Write is recorded. The I / O pattern with the highest proportion is recorded in the frequent access 1804. Here, an example in which a business unit corresponds to a volume is given, but the present invention is not limited to this. As another example, the business unit may be a VM, a business program on the server 3000, or information on a processing unit executed by the business program.

FIG. 12 is a flowchart of a process for executing learning data sharing according to the second embodiment. The learning-required data sharing process 400 is the process in Example 2 corresponding to step 105 of the predictive expression generation process 100 in FIG. 8 in Example 1. This process is performed by the processor 1300 of the management server 1000 executing the predictive expression generation program 1170 expanded on the memory 1100. A specific example of this flowchart is shown below.

The prediction formula generation program 1170 first acquires the information of the business characteristic management table 1800 (step 401). Next, the prediction formula generation program 1170 checks whether or not there is a component of the prediction formula generation target used in a similar business similar to the business of the target for which the prediction formula is generated (step 402). Here, the information in the business characteristic management table 1800 acquired in step 401 is used to determine whether or not the product is used in similar business. Similar businesses may be grouped in advance, and the presence or absence of components used in businesses belonging to the same group may be checked.

As an example of grouping, businesses with the same high-frequency access information may be grouped as business-like groups. Alternatively, regarding the I / O increase / decrease rate, a decrease rate of 5% or more, an increase / decrease rate of plus or minus 5% or less, an increase rate of 5% or more, etc. may be set as business-like groups. Alternatively, the k-means method may be used to classify the work into any number of groups. Alternatively, the number of groups may be input in advance and grouped appropriately according to the number of sono groups. Alternatively, the business may be grouped by combining the above grouping methods. As described above, the grouping may be performed by any method and is not particularly limited.

In the case of the business characteristic management table 1800 illustrated in FIG. 11, Volume 1 and Volume 3 have the same high-frequency access 1804 for "RW", the same I / O increase / decrease rate 1803 for "5% or more", and I / O. The number of O's 1802 is "10000 or more" and is the same. Businesses may be grouped so that Volume 1 and Volume 3 are determined to be the same business-like group.

In step 402, when there are components used in the similar business, the prediction formula generation program 1170 uses the information in the prediction formula source information table 1140 of each component used in the similar business to formulate the prediction formula. Is generated and stored in the prediction expression table 1150 (step 403). In step 402, when there is no component used in the similar business, the prediction formula generation program 1170 generates a prediction formula from the information in the prediction formula source information table for each component and stores it in the prediction formula table. (Step 404).

As described above, by sharing learning data between groups with similar tasks, it is possible to shorten the learning time and quickly implement highly accurate and efficient management based on machine learning technology. For example, when creating a new environment, it is usually necessary to acquire various historical information such as performance information and capacity information for a long period of time, for example, several months. However, if there is a business in an environment similar to the environment of the newly created business, the data of the similar business group can be used to determine the similar business group in a short period of time, for example, 3 days, and the machine learning technology can be used. Efficient management based on this becomes possible. Further, also in step 203 of the setting change content determination process 200 shown in FIG. 9, since the missing domain can be searched based on the configuration taken by each similar business group, the missing data can be searched. It is possible to shorten the time required for collection, and it is possible to carry out efficient management based on machine learning technology.

The computer system according to each of the above-described embodiments can be organized in the following aspects.

(Aspect 1)
The relationship between the information collecting unit that acquires learning data from the components of the target system while the target system is operated and the components of the target system based on the learning data is expressed by the relationship between the target information and the explanatory information. It has a prediction formula generation unit that generates the prediction formula, a setting change content determination unit that determines the setting content that changes the configuration of the target system, and a configuration change unit that changes the configuration of the target system. The setting change content determination unit extracts a state in which the training data is not sufficiently acquired within the range of the states of the components that can be taken when changing the configuration of the target system as a data shortage state, and the components. Determines the setting content to change the configuration of the target system so that the data is insufficient, the configuration changing unit changes the configuration of the target system according to the determined setting content, and the information collecting unit , A learning data processing device that acquires training data when the data is insufficient from the constituent elements of the target system.

Since the learning data that will be insufficient when the configuration of the target system is changed in the future can be acquired by temporarily changing the configuration of the target system in advance, the lack of learning data that occurs when the configuration is actually changed Is reduced, and the accuracy of machine learning is improved at an early stage. When machine learning in a target system in which the configuration is changed relatively frequently is applied to this, it is possible to suppress a decrease in the accuracy of machine learning that occurs when the configuration is changed.

(Aspect 2)
The setting change content determination unit extracts a configuration in which learning data is not acquired in excess of a predetermined amount of data from the range of states of the component that the target system can take, and determines a setting change corresponding to the configuration. The learning data processing apparatus according to the first aspect.

It is possible to temporarily set a configuration in which the learning data is not sufficient among the configurations that the target system can take and collect the learning data, and it is possible to cover the learning data of the configuration that the target system can take.

(Aspect 3)
The setting change content determination unit temporarily changes the configuration of the target system so that the component is in the data shortage state, and the configuration change unit configures the target system according to the determined setting content. When the information collecting unit acquires and accumulates the learning data in the data shortage state from the component of the target system and the configuration of the target system is changed, the predictive expression generation unit Is the learning data processing apparatus according to the first aspect, wherein the prediction formula is generated using the learning data accumulated by the information collecting unit in a configuration in which the component is in the data shortage state.

When the configuration of the target system is changed, the training data in the changed configuration is acquired in advance and the prediction formula is generated from the learning data in the changed configuration, so when the configuration is changed, it takes a short time. A highly accurate prediction formula can be obtained.

(Claim 4)
The prediction formula uses the performance of the target component, which is the target component of the prediction formula, as the target information, and the performance of one or more related components logically associated with the target component as explanatory information. The target information is indicated by a function of the explanatory information, the information collecting unit accumulates the performance exhibited by the components of the target system as performance history information, and the prediction formula generation unit generates the performance history. The learning data processing device according to aspect 1, wherein the function is calculated using the information as the learning data.

Since the prediction formula is generated based on the measured and accumulated performance history information, it is possible to generate a good prediction formula by accumulating sufficient performance history information.

(Aspect 5)
The learning data processing apparatus according to aspect 4, wherein the target information is the response time of the volume in the storage, and the explanatory information includes the utilization rate of the processor used to access the volume and the size of the cache.

(Aspect 6)
In the prediction formula, the target information is represented by the sum of the products of the explanatory information and the coefficients, and the prediction formula generation unit uses the performance history information as the learning data and calculates the coefficients for each of the related components. The learning data processing device according to aspect 4, which is calculated.

Since the prediction formula is represented by the sum of products of the explanatory information and the coefficient and the coefficient is calculated, a function showing the relationship between the target information and the explanatory information can be easily calculated.

(Aspect 7)
The learning data processing device according to aspect 1, wherein the prediction formula generation unit generates the prediction formula using learning data acquired by a similar system that satisfies a predetermined similarity condition with respect to the target system.

When the target system has a similar system, the prediction formula is generated using the learning data of the similar system. Therefore, the learning data that can be used is increased, and highly accurate machine learning becomes possible from an early stage after the configuration change.

(Aspect 8)
The target information is the performance of the volume of the storage, and the similarity condition is defined by the degree of similarity of the I / O pattern including random read, random write, sequential read, and sequential write to the volume, according to the seventh aspect. Learning data processing device.

Since the similarity is judged based on the similarity of the I / O patterns, it is possible to use the learning data collected in the tasks with similar I / O patterns in the configuration change or new construction of other tasks.

(Aspect 9)
The learning data processing device according to the fifth aspect, wherein the setting change content determining unit determines to change the cache size to a size in which sufficient learning data is not obtained within a possible range.

If the explanatory information has a cache size, the training data within the range that the cache size can take can be covered in advance, so even if the cache size of the target system is changed, the training data is insufficient and the accuracy of the prediction formula is maintained high. can do.

1000 ... Management server, 1100 ... Memory, 1110 ... Related information table, 1120 ... Performance history information table, 1121 ... Time, 1124 ... Metric, 1125 ... Performance value, 1130 ... Configuration information table, 1133 ... Metric, 1134 ... Value, 1140 ... Prediction formula source information table, 1160 ... Information collection program, 1170 ... Prediction formula generation program, 1180 ... Setting change content determination program, 1190 ... Configuration change program, 1200 ... Communication device, 1300 ... Processor, 1400 ... Output device, 1500 ... Input device, 1600 ... Storage device, 1700 ... Internal bus, 1800 ... Business characteristic management table, 2000 ... Storage device, 2100 ... Memory, 2110 ... Disk cache, 2120 ... Configuration performance information collection program, 2130 ... Configuration change program, 2200 ... Logical volume provider, 2210 ... Volume, 2220 ... Disk pool, 2230 ... Physical area,
2300 ... Disk I / F controller, 2500 ... Processor, 2600 ... Data I / F, 2700 ... Communication path, 3000 ... Server, 3100 ... Memory, 3110 ... Configuration information collection program, 3120 ... Business program, 3130 ... Configuration change program, 3300 ... processor, 3500 ... communication path, 4000 ... SAN, 5000 ... management network

Claims (10)

An information collection unit that acquires learning data from the components of the target system while the target system is in operation.
A prediction formula generation unit that generates a prediction formula that expresses the relationship between the components of the target system based on the learning data by the relationship between the target information and the explanatory information.
The setting change content determination unit that determines the setting content that changes the configuration of the target system,
It has a configuration change unit that changes the configuration of the target system.
The setting change content determination unit extracts a state in which the learning data is not sufficiently acquired within the range of the states of the components that can be taken when changing the configuration of the target system as a data shortage state, and the configuration is described. Determine the setting contents to change the configuration of the target system so that the element is in the data shortage state.
The configuration change unit changes the configuration of the target system according to the determined setting contents.
The information collecting unit acquires learning data in the data shortage state from the component of the target system.
Learning data processing device. The setting change content determination unit extracts a configuration in which learning data is not acquired in excess of a predetermined amount of data from the range of states of the component that can be taken by the target system, and determines a setting change corresponding to the configuration. To do
The learning data processing device according to claim 1. The setting change content determination unit temporarily changes the configuration of the target system so that the component is in the data shortage state.
The configuration change unit changes the configuration of the target system according to the determined setting contents.
The information collecting unit acquires and accumulates learning data in the data shortage state from the component of the target system.
When the configuration of the target system is changed
The prediction formula generation unit generates the prediction formula using the learning data accumulated by the information collection unit in a configuration in which the component is in the data shortage state.
The learning data processing device according to claim 1. The prediction formula uses the performance of the target component, which is the target component of the prediction formula, as the objective information, and the performance of one or more related components logically associated with the target component as explanatory information. , The target information is indicated by a function of the explanatory information.
The information gathering unit accumulates the performance exhibited by the components of the target system as performance history information.
The prediction formula generation unit calculates the function using the performance history information as the learning data.
The learning data processing device according to claim 1. The learning data processing apparatus according to claim 4, wherein the target information is the response time of the volume in the storage, and the explanatory information includes the usage rate of the processor used for accessing the volume and the size of the cache. In the prediction formula, the target information is represented by the sum of the products of the explanatory information and the coefficients.
The prediction formula generation unit calculates the coefficient for each of the related components using the performance history information as the learning data.
The learning data processing device according to claim 4. The learning data processing device according to claim 1, wherein the prediction formula generation unit generates the prediction formula using learning data acquired by a similar system that satisfies a predetermined similarity condition with respect to the target system. A performance of a volume of the object information gas storage,
The similarity condition is determined by the degree of similarity of the I / O pattern including random read, random write, sequential read, and sequential write to the volume.
The learning data processing device according to claim 7. The learning data processing device according to claim 5, wherein the setting change content determination unit determines to change the cache size to a size within which sufficient learning data has not been obtained. A prediction formula that acquires learning data from the components of the target system while the target system is in operation and expresses the relationship between the components of the target system based on the learning data by the relationship between the target information and the explanatory information. It is a training data processing method for generating
The means for determining the content of setting changes realized by a computer is
Within the range of states of the components that can be taken when the configuration of the target system is changed, a state in which the training data is not sufficiently acquired is extracted as a data shortage state.
The setting content for changing the configuration of the target system is determined so that the component is in the data shortage state.
The configuration changing means realized by the computer changes the configuration of the target system according to the determined setting contents.
The information collecting means realized by the computer acquires the learning data in the data shortage state from the component of the target system.
Training data processing method.

Images