JPH10198643A - Distributed computer system - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a distributed computer system capable of preventing the operation of a node from being stopped. SOLUTION: Resources to be consumed by respective jobs are previously defined in a consumption-predicted job resource definition table 12. At the time of executing jobs, the jobs are distributed so that a resource to be consumed in the future which is required for a job execution node 2 for executing the jobs does not exceed a resource included in the node 2 based on a consumption- predicted resource and consumption-predicted resource information stored in the table 12 to prevent the stop of operation.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a distributed computer system including a plurality of computer nodes.

[0002]

2. Description of the Related Art In recent years, a computer system has a case in which a distributed computer system in which a plurality of computers are connected via a network is configured due to a demand for high-speed processing, an increase in the amount of processing data, and the development of a network connecting the computers. Are increasing. When a job is executed on a distributed computer system, in order to effectively utilize computer resources (hardware and software necessary for executing a group of work in the computer system), the resource consumption state of each computer must be determined. It is necessary to allocate jobs optimally to each computer while managing them.

Conventionally, the job distribution in a distributed computer system is determined at the time of starting a new job, based on (i) the resource consumption state at the time of the job distribution of each computer, and (ii) the job distribution based on a predetermined policy. Was.

[0004]

However, the job distribution in the conventional distributed computer system is based on the fact that the resource information on each computer managed for controlling the job is such that the resource information on the computer, which changes every moment, in a certain time section. Based on data. For this reason, even when it is determined that the available resources of the computer are sufficient to execute the job at the time of starting the job, the resources are consumed by a plurality of jobs executed on the computer over time, and the computer is not used. May be stopped due to lack of resources.

When the resources of the computer are consumed beyond a preset threshold, the operator is notified by an alarm or the like, and the operator stops the computer by taking measures such as canceling the job. It is possible to prevent. However, there is inconvenience that the job whose execution has been stopped needs to be restarted by another computer, and when the computer runs all day, the operator must constantly monitor the job. This is because the job allocation is determined only by the information of the computer resources at the time of starting the job. To solve this, all the jobs running on the computer must It is necessary to manage resources consumed over time (time from start to stop of a job).

An object of the present invention is to provide a distributed computer system capable of preventing a node from stopping operation.

[0007]

In order to achieve the above-mentioned object, the present invention provides a distributed computer system comprising a plurality of computers. Definition means, and job distribution means for allocating a job such that, when the job is executed, future scheduled resources required on the node executing the job do not exceed the resources of the node. It has a configuration. According to this configuration, the resources to be used for each job are defined in advance in the job consumption scheduled resource defining means, and when the job is executed, the job is not allocated to the individual nodes at a time, that is, the job allocation is performed. If the resource is determined not to exceed the resources of the node, the resources on the computer required to execute the job at the time of starting the job are virtually reserved as predicted consumption resources over the lifetime of the job. As a result, the resources consumed by all the jobs executed on the computer over the lifetime can be managed, and the situation in which the computer resources run short due to the execution of each job does not occur.

[0008]

Embodiments of the present invention will be described below. FIG. 1 is a block diagram showing a job distribution processing part of the distributed computer system according to the present invention. FIG.
Shows a job control node 1 of the distributed computer system and a job execution node 2 connected to the job control node 1. The job control node 1 includes a job distribution control unit 11 that operates based on a job execution request from a user, a job consumption resource definition table 12 that provides the defined job consumption resource to the job distribution control unit 11, Execution node owned resource definition table 1 which holds the owned resource information of each execution node and provides the information to the job distribution control unit 11.
3. Collect the predicted resource consumption management table 14 for providing the predicted resource consumption on the node to the job distribution control unit 11 and the predicted resource consumption information from the job execution node 2 and reflect the information on the predicted resource management table 14. A predicted resource consumption collecting unit 15 is provided.

The job execution node 2 includes a job execution control unit 21 started by an execution request from the job control node 1, and a job consumption resource definition for providing the defined job consumption resource to the job execution control unit 21. Table 22,
The job execution control unit 21 updates the predicted consumption resource of the execution node with the predicted consumption resource management table 23 and the predicted consumption resource management unit 24 that manages the predicted consumption resource with reference to the contents of the predicted consumption resource management table 23. Have. Although a plurality of job execution nodes 2 can exist, in FIG.
Only one configuration is shown, and only one is described in the operation description. In FIG. 1, the job control node 1 and the job execution node 2 are configured as separate nodes, but the job control node 1 may include the job execution node 2.

In the above configuration, the job distribution control unit 1
Reference numeral 1 denotes a predicted consumption resource (on a job execution node) Rfc managed by a predicted consumption resource general management table 14 and a job consumption scheduled resource definition table 12 in response to a job execution request from a user.
(Rfc + R) of the scheduled resource Rcp of the job defined in
Select the job execution node 2 in which the ratio Rfc / (Rfc + Rcp) of the Rfc to the cp) is the minimum and the Rfc + Rcp does not exceed the resource owned by the job execution node defined in the execution node owned resource definition table 13. . A job execution request is made to the job execution control unit 21 on the selected job execution node 2.

The job execution control unit 21 that has received the job execution request from the job distribution control unit 11 activates the job and acquires the scheduled resource value of the job from the scheduled job resource definition table 22. This scheduled resource value is added to the predicted resource management table 23 by the job execution control unit 21, and when the execution of the job is completed, the scheduled resource value of the job is subtracted from the predicted resource management table 23. , The predicted consumption resource is updated, and the state of the predicted consumption resource is managed. After that, the forecast resource management unit 2
4, the predicted consumption resource management table 23 is referred to, the predicted consumption resource information is acquired based on the reference result, and the predicted consumption resource information is transmitted from the predicted consumption resource management unit 24 to the predicted consumption resource collection unit 15. You.

As described above, the resources to be used for each job are defined in advance in the job consumption scheduled resource definition tables 12 and 22, and the job distribution control unit prevents the jobs from being distributed to each node at one time when the job is executed. If determined by 11, the computer resources can be used effectively. As a result, the resources consumed by all the jobs executed on the computer over the lifetime can be managed, and the situation in which the computer resources run short due to the execution of each job does not occur.

The present invention is particularly effective in applications such as executing a large number of similar jobs that consume a large amount of resources, such as in the field of image processing and scientific and technical computing. That is, it is effective for application to a system in which excessive resources are required for computers by allocating too many jobs to each computer at once, which adversely affects the operation of the computers.

FIG. 2 shows the job distribution control unit 11 shown in FIG.
FIG. 4 is a block diagram showing the details of. Job distribution control unit 11
Is configured to include a job execution request management unit 16, a job execution request queue management table 17, and a job distribution processing unit 18. The job execution request management unit 16 registers the job execution request in the queue of the job execution request queue management table 17 when receiving the job execution request from the user. In addition, the job distribution processing unit 18 executes a job execution request from the job execution request queue management table 17 based on a start command from the job execution request management unit 16 at a predetermined cycle or when a user issues a job execution request. The request is taken out, the job consumption scheduled resource definition table 12, the execution node possessed resource definition table 13
Then, each job is distributed to the job execution nodes with reference to the three items of the predicted consumption resource general management table 14.

Further, the job distribution processing unit 18 suspends the distribution of the job for which the required predicted resource cannot be secured on any job execution node, and checks the next order in the job execution request queue management table 17. Try to distribute the jobs. This realizes efficient use of the job execution node. However, when such job distribution is performed, a case may occur in which a job requiring a large amount of resources to be consumed is not distributed to the job execution node 2 forever. Therefore, the job distribution processing unit 18
Each time the distribution of a job is skipped, the skip counter (not shown) of the job in the job execution request queue management table 17 is incremented. When the count value reaches a threshold value defined in advance (as a skip counter), processing is performed so that the job is executed with priority.
By employing this job distribution method, it is possible to realize fair resource distribution for jobs, that is, to execute jobs fairly. The job consumption scheduled resource definition table 22
When the job distribution control unit 11 requests the job execution control unit 21 to execute a job, the job distribution control unit 11 can also delete the resource by notifying the job execution control unit 21 of the resource to be consumed.

Next, a specific example of the distributed computer system according to the present invention will be described. FIG. 3 is a block diagram of a distributed computer system that executes a job for processing and processing a large amount of input data and outputting the job on a plurality of data processing servers. Since the size of the input data differs for each job execution request given by the user, the size of the data output by each job also differs for each job execution request. For this reason, in the conventional job distribution technique, as the execution of each job on the data processing server progresses, there is a possibility that the disk (storage medium) for storing the output data is used up. The inability to store the output data on the disk means that the operation of the data processing server is stopped. Therefore, such a situation has to be avoided during the job distribution process. However, by applying the present invention, it is possible for a job being executed on each data processing server to know the disk capacity to be consumed over its lifetime, and to perform job distribution based on the information. It is possible to significantly reduce the risk of stopping the operation of the server.

Hereinafter, the configuration and operation of FIG. 3 will be described. FIG. 3 shows a configuration of the management agent 101 and a plurality of data processing servers 201.
01 corresponds to the job control node 1 in FIG. 1, and the data processing server 201 corresponds to the job execution node 2 in FIG.
The management agent 101 has a job attribute definition file 1
1, a job attribute generator 122, a job attribute definition table 123 corresponding to the job consumption scheduled resource definition table 12 in FIG. 1, a node storage resource definition file 131, a node storage resource generator 132, and an execution node possession resource management table 13 in FIG. Corresponding node storage resource definition table 1
33, a storage resource general account table 141, a storage resource account collection master 151, a job execution reception processing unit 161 corresponding to the job execution request management unit 16 in FIG. 2, and a job management corresponding to the job execution request queue management table 17 in FIG. It is provided with a table 171, a job dispatcher 181 corresponding to the job distribution processing unit 18 in FIG. 2, and a job status update processing unit 182.

The data processing server 201 has a job activation processing unit 21 corresponding to the job execution control unit 21 in FIG.
1, a job shell 212, a storage resource account table 231, and a storage resource account collection slave 241 corresponding to the predicted consumption resource management unit 24 in FIG. Note that a plurality of data processing servers 201 can exist.
01 will be described.

In the management agent 101, the job attribute definition table 123 exists on a memory (not shown) of the management agent, whereas the job attribute definition file 121 and the job attribute generator 122 function as auxiliary mechanisms. More specifically, the user describes job attributes in the job attribute definition file 121 and executes the job attribute generator 122 to store the job attributes stored in the job attribute definition file 121 in the job attribute definition table on the memory. 123 can be reflected. In this embodiment, the job attribute is defined as (1) a relational expression between the size of input data given by a user at the time of a job execution request and the size of intermediate data and final output data generated in the course of job execution. , (2) a skip limit value for starting the job.

The node storage resource definition table 133 exists in the memory of the management agent 101, whereas the node storage resource definition file 131 and the node storage resource generator 132
33 functions as an auxiliary mechanism. Specifically,
The user describes information on the amount of storage resources possessed by each node in the node storage resource definition file 131, and operates the node storage resource generator 132 to store the amount of storage resources possessed by each node in the node storage resource definition table 133 on the memory. Can be reflected in In this embodiment, the node storage resources defined in the node storage resource definition file 131 include the disk capacity of each data processing server.

The job management table 171 exists on the memory of the management agent 101. The job dispatcher 181 is activated at a predefined cycle, requests the storage resource account collection master 151 to collect the contents of the storage resource account information table 231 of each data processing server 201, and stores the collection results. The job execution request is extracted from the job management table 171 with reference to the resource management account table 141, and the job is dispatched to one of the data processing servers 201.

Here, the job dispatcher 181
If there is no free storage resource required for executing the job in any of the data processing servers 201, the execution of the job is suspended, and the skip counter of the job in the job management table 171 is incremented. When the value of the skip counter reaches the skip limit value given from the job attribute definition table 123, the job dispatcher 181 does not skip the job thereafter, and stores enough data to execute the job in one of the data processing servers 201. Wait until resources are available. When dispatching a job, the job dispatcher 181 uses the storage resource account value of the job generated by the job dispatcher 181 based on the information acquired from the job attribute definition table 123 as a job start parameter, and transfers this to the job start processing unit 211. I do.
Thus, the job consumption scheduled resource definition table 22 required in FIG. 1 can be omitted.

The job management table 171 stores not only the queue management information of the job requested by the user but also the status of each job executed on the data processing server 201.
It is held through the job request reception processing unit 161. In addition,
Although not shown in FIG. 3, the job management table 171
Can be used as original data for providing the user with the execution state of each job through a screen or the like.

The storage resource general account table 141 corresponds to the predicted consumption resource general management table 14 in FIG. 1 and exists on the memory of the management agent. Further, the storage resource account collection master 151 corresponds to the predicted consumption resource collection unit 15 in FIG.
Is equivalent to This storage resource account collection master 151
Collects the storage resource account information of each data processing server 201 from the storage resource account collection slave 241 according to the request from the job dispatcher 181 and stores it in the storage resource general account table 141. The storage resource account table 231 corresponds to the predicted consumption resource management table 23 in FIG. 1, and exists on the memory of the data processing server. Also, the storage resource account collection slave 241
Transfers the contents of the storage resource account collection table 231 to the storage resource account collection master 151 in response to a request from the storage resource account collection master 151. The job shell 212 includes a job, and has the same function as that of the part of the job execution control unit 21 shown in FIG.

FIG. 4 is a flowchart showing the processing executed by the job shell 212. First, the storage resource account value passed as a parameter of the job start request from the job dispatcher 181 to the job start processing unit 211 is acquired (step S401). Next, the obtained storage resource account value is added to the storage resource account table 231 (step S402). Further, the job is executed (step S403), and after the job is completed, the storage resource account value is subtracted from the storage resource account table 231 (step S404). Finally,
A job completion notification is issued to the job status update processing unit 182 (step S405). As described above, the contents of the storage resource management account table 141 can be updated with the execution of the job by a simple process.

FIG. 5 shows the configuration of the job attribute definition table 123. Here, the estimated memory capacity, the estimated disk capacity, and the skip limit are defined for each job. In FIG. 5, the job "JOB-
The disk capacity to be consumed for “A” is 5 fixedly used.
It is calculated by adding 2.2 times the input data size to 0 MB (50 + 2.2 × input data size (MB)).

FIG. 6 shows the configuration of the job management table 171. In FIG. 6, the job "A01" of the first entry indicates that the startup is skipped because there is currently no data processing host capable of securing the estimated memory capacity and the estimated disk capacity.
In this example, since the skip counter has reached the skip limit, the job dispatcher 181 does not skip starting the job “A01” any more, and
01 is not issued until the required memory capacity and the required disk capacity required to execute “01” can be secured in one of the data processing hosts.

FIG. 7 shows the node storage resource definition table 133.
Shows the contents of the configuration. Here, the possessed memory capacity including the swap capacity and the possessed disk capacity are defined for each data processing server. FIG. 8 shows the configuration of the storage resource account table 231. In this example,
The estimated memory capacity and the estimated disk capacity of each job executed on the data processing host are stored.

FIG. 9 shows the storage resource management account table 1
41 shows the configuration contents. In this example, the total memory capacity and the total disk capacity of the job are stored for each data processing server. As mentioned above,
Even in a system including the data processing server 201,
By defining the resources to be used for each job in the job attribute definition table 123 and the data processing server 201 in advance, it becomes possible to manage the resources consumed by all the jobs executed on the computer over their lifetime. In addition, a situation in which computer resources are insufficient due to the execution of each job does not occur.

FIG. 10 is a block diagram showing another embodiment of the present invention obtained by developing the configuration of FIG. In FIG. 10, the management agent 101 and the data processing server 20
1 shows a system configuration in which information that can be shared by the server 1 is managed by the database server 301. Information sharing on the database server 301 includes file sharing by NFS (Network File / System) or the like, and SQL (Structured Query Language: Structured Query L).
An DBMS is based on a DBMS (Data Base Management System) using an anguage or the like.

The configuration of the distributed computer system shown in FIG. 10 will be described. A management agent 101 corresponding to the job control node shown in FIG. 1, a plurality of data processing servers 201,
And management agent 101 and data processing server 201
And a database server 301 connected to the server. Here, the three are configured independently, but the database server 301
1 or the data processing server 201.

The management agent 101 is a job execution reception processing unit 1 corresponding to the job execution request management unit 16 in FIG.
61 and a job dispatcher 181 corresponding to the job distribution processing unit 18 in FIG. Further, the data processing server 201 is provided with the job execution control unit 21 shown in FIG.
And a job shell 212 corresponding to the above. In addition, the database server 301 performs the job execution request queue management table 17 shown in FIG.
1, a job attribute definition table 312 corresponding to the job consumption scheduled resource definition table 12 in FIG. 1, a node storage resource definition table 313 corresponding to the execution node possessed resource management table 13 in FIG. 1, and FIG. The storage resource management account table 314 corresponding to the predicted consumption resource management table 14 of FIG.

The job execution reception processing unit 161 of the management agent 101 receives a job execution request from a user and adds an entry of the job execution request to the job management table 311 in the database server 301. The job dispatcher 181 starts at a predefined cycle,
The job execution request is extracted from the job management table 311, the predicted resource consumption information of the job defined in the job attribute definition table 312, the node storage resource definition table 31
3 based on the storage resource information of each data processing server defined in 3 and the storage resource consumption information of each data processing server managed by the storage resource general account table 314. decide. Further, the job dispatcher 181 sends a job activation request to the job activation processing unit 211 using the storage resource account value of the job calculated based on the information defined in the job attribute definition table 312 as a parameter. In response to this request, the job activation processing unit 211 activates the job shell 212. This job shell 21
The process 2 will be described below.

FIG. 11 is a flowchart showing the processing of the job shell 212. First, job dispatcher 1
The storage resource account value passed as a parameter of the job start request from 81 to the job start processing unit 211 is acquired (step S601). Next, the storage resource account value is added to the storage resource management account table 314 (step S602). Further, the activated job is executed (step S603), and after the execution of this job is completed, the storage resource account value in the storage resource general account table 314 is subtracted (step S604). Next, the end status of the job is reflected in the job management table 311 (step 605). With the processing in FIG. 11, the contents of the storage resource management account table 314 can be updated with the execution of the job.

As described above, according to the configuration of FIG. 10, even in a system in which information that can be shared by the management agent 101 and the data processing server 201 is managed by the database server 301, as in the case of FIG. The resources to be used for the job are defined in the job attribute definition table 312, and the resources are appropriately allocated using the job attribute definition table 312, whereby the resources on the computer required to execute the job at the time of starting the job are predicted over the lifetime of the job. It can be virtually reserved as a consumed resource, and all the jobs executed on the computer can manage the resources consumed over their lifetime, so that there is no shortage of computer resources due to the execution of each job .

[0036]

As described above, according to the present invention, in a computer system, resources to be consumed of a job to be executed are defined in advance as job attributes, so that the resources required on a node at the time of job start-up are required. Jobs can be distributed such that future scheduled resources do not exceed the resources of the nodes, and a distributed computer system can be realized in which the occurrence of job execution node stoppages is significantly reduced.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a job distribution processing part of a distributed computer system according to the present invention.

FIG. 2 is a block diagram illustrating details of a job distribution control unit in FIG. 1;

FIG. 3 is a block diagram showing a configuration of a first specific example of the present invention.

FIG. 4 is a flowchart illustrating a job shell process of FIG. 3;

FIG. 5 is a configuration example diagram of a job attribute definition table 123.

FIG. 6 is a diagram illustrating a configuration example of a job management table 171.

FIG. 7 is a configuration example diagram of a node storage resource definition table 133.

FIG. 8 is a configuration example diagram of a storage resource account table 231.

FIG. 9 is a configuration example diagram of a storage resource management account table 141.

FIG. 10 is a block diagram showing a configuration of a second specific example of the present invention.

11 is a flowchart illustrating processing of the job shell in FIG.

[Explanation of symbols]

 REFERENCE SIGNS LIST 1 job control node 2 job execution node 11 job distribution control unit 12 job consumption scheduled resource definition table 13 execution node possessed resource management table 14 estimated resource consumption management table 15 estimated resource consumption collection unit 16 job execution request management unit 17 job execution request Queue management table 18 Job distribution processing unit 21 Job execution control unit 22 Job consumption scheduled resource definition table 23 Predicted resource consumption management table 24 Predicted resource consumption management unit 101 Management agent 123, 312 Job attribute definition table 133, 313 Node storage resource definition table 141, 314 Storage resource management account table 181 Job dispatcher 201 Data processing server 211 Job start processing unit 212 Job shell 301 Database server

 ──────────────────────────────────────────────────続 き Continuing on the front page (72) Inventor Norio Tanaka 5-2-1 Omika-cho, Hitachi City, Ibaraki Prefecture Inside the Hitachi, Ltd. Omika Plant

Claims (6)

[Claims] 1. In a distributed computer system comprising a plurality of computers, a scheduled job resource defining means for defining in advance a scheduled resource of a job to be executed, and a node for executing the job when executing the job. And a job distribution means for allocating a job such that the resource to be consumed in the future does not exceed the resource of the node. 2. The method according to claim 1, wherein the job distribution unit calculates a sum of the scheduled resource of the job defined by the scheduled resource consumption defining unit and the predicted consumption resource on the node managed by the predicted consumption resource general management unit. 2. The distributed computer system according to claim 1, wherein jobs are allocated so as not to exceed the resources held on the nodes defined by the node owned resources defining means. 3. The job distribution unit further includes, in addition to the job distribution, a scheduled consumption resource of the job defined by the scheduled job consumption resource definition unit and a prediction on the node managed by the predicted consumption resource general management unit. Allocating jobs so that the ratio of the predicted resource consumption on the node to the sum with the consumed resource is minimized and the sum does not exceed the resource held on the node defined by the node owned resource defining means. 3. The distributed computer system according to claim 2, wherein: 4. The distributed computer according to claim 1, wherein the job distribution unit causes the skipped job to be executed preferentially when the number of times of activation of the skipped job exceeds a preset value. system. 5. The distributed computer system according to claim 1, wherein a database server is used instead of said node. 6. The node or the data processing server acquires a scheduled resource consumption value or a storage resource account value in response to a job execution request, and adds this value to a predicted resource consumption management unit or a storage resource account table. 2. The apparatus according to claim 1, further comprising means for initiating a job and subtracting the scheduled resource value from said predicted resource management means or subtracting said storage resource account value from said storage resource account table after execution. Item 6. The distributed computer system according to Item 5.