JP2011076304A - Device, method and program for allocating job - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To improve use efficiency of a calculating node by avoiding a bottleneck due to an I/O processing standby state to the same data in a parallel type batch processing system including a plurality of nodes. <P>SOLUTION: When a sub-job which is executed by a sub-job execution part 302 of a job allocation device 3 is excluded from its access to shared data 422, processing is interrupted without being retried, and the processing state is suspended and stored in a shared data I/O processing standby queue 314, and the execution of a new sub-job is started. When the processing of the shared data of a certain new sub-job ends, a job including processing to the same shared data is extracted from the shared data I/O processing standby queue 314, and the shared data in the processing standby state are executed in a batch. When the suspended processing of the sub-job is resumed, the processing is started from the point where the processing of the sub-job is interrupted. <P>COPYRIGHT: (C)2011,JPO&INPIT

Description

  The present invention relates to a job assignment apparatus, a job assignment method, and a job assignment program used for batch processing systems such as finance, distribution, and industry.

  In recent years, grid technology that connects a plurality of computers via a network and provides powerful computing resources has attracted attention, and a method for speeding up the technology by applying this technology to batch processing has been realized. For the purpose of further speeding up the batch processing, a method has been proposed in which the execution order of jobs is controlled to effectively use the computation nodes (see, for example, Patent Document 1).

  In the technique described in Patent Document 1, when a job is processed in parallel by a plurality of computers, a task of a job with a low priority is saved in a task queue and a job with a high priority is assigned in advance. It improves the utilization efficiency of the computation node.

JP 2008-226023 A

  However, the technique for controlling the job execution order described in Patent Document 1 is aimed at parallelization of processes, and parallelism of data access is not considered. If all the compute nodes have a lot of processing for shared data to be read and written, the I / O (Input / Output) processing of the shared data becomes a bottleneck, and the CPU (Central Processing Unit) of the compute node cannot be used up effectively. .

  The present invention has been made in view of such a background, and the present invention avoids a bottleneck caused by waiting for I / O processing and can improve the utilization efficiency of a computation node, and a job allocation apparatus and a job allocation method And providing a job allocation program.

The present invention controls a process when access to shared data is exclusive in a parallel batch processing system in which jobs are parallelized, distributed to a plurality of computers (job allocation devices), and a process involving data access is executed. To do.
When a job is locked out of access to shared data, the processing is interrupted without retrying, the processing state is suspended and saved in the queue, and execution of a new job is started.
When the shared data processing for a new job is completed, a job including processing for the same shared data is extracted from a plurality of suspended jobs with reference to the queue table, and the shared data waiting for processing is collectively executed.

  According to the present invention, it is possible to provide a job allocation device, a job allocation method, and a job allocation program that can avoid bottlenecks due to waiting for I / O processing and improve the utilization efficiency of a computation node.

1 is a functional block diagram illustrating a configuration example of a job allocation system including a job allocation device according to the present embodiment. It is a figure which shows an example of a data structure of the sub job management table which concerns on this embodiment. It is a figure which shows an example of a data structure of the sub job progress management table concerning this embodiment. It is a figure which shows an example of a data structure of the shared data I / O process waiting queue table which concerns on this embodiment. It is a figure which shows an example of a data structure of the shared data I / O process waiting queue which concerns on this embodiment. It is a figure which shows an example of the data structure of the lock table which concerns on this embodiment. It is a sequence diagram for demonstrating the job allocation method which concerns on this embodiment. It is a sequence diagram for demonstrating the job allocation method which concerns on this embodiment. It is a sequence diagram for demonstrating the job allocation method which concerns on this embodiment. It is a sequence diagram for demonstrating the job allocation method which concerns on this embodiment. It is a figure for demonstrating the time-sequential change of the subjob presence location which concerns on this embodiment. It is a figure for demonstrating the total process of the financial institution which is an example to which this invention is applied.

  Next, modes for carrying out the present invention (referred to as “embodiments”) will be described in detail with reference to the drawings as appropriate.

  In the present embodiment, as an example of batch processing in which all sub-jobs perform shared data processing, tabulation processing for obtaining the total value of specific items will be described as an example. This aggregation batch process is executed in a grid environment.

FIG. 12 is a diagram for explaining the aggregation processing of a financial institution that is an example to which the present invention is applied.
Here, the job net shown in FIG. 12A is a collection of jobs associated with the execution order. When the job net is executed, the jobs in the job net are executed according to the execution order. Further, the job shown in FIG. 12B is a minimum unit of elements constituting a business, and specifically, a collection of commands, shell scripts, Windows (registered trademark) execution files, and the like.
In FIG. 12A, for example, a branch of a financial institution (bank) aggregates the transaction results on the business day (aggregation process), creates slips and invoice data (slip creation and bill creation), The processing order of each job of calculating sales (sales calculation) and outputting a form (form output) is shown as a job net. FIG. 12B shows the contents of “total processing”, which is one job in the job net of FIG. 12A. “Total processing” is further divided into “data extraction” and “accounting calculation”. Loop processing is performed in which each job step of “personal master update” and “department store master update” is repeated for each customer. The sub-job shown in FIG. 12C is obtained by dividing the job into predetermined processing units by paying attention to loop processing in order to parallelize the processes of this job. For example, if there are jobs that process customer units from 0001 to 3000, they are divided into sub-jobs from 0001 to 1000, 1001 to 2000, and 2001 to 3000, and parallel processing is performed.

The present invention divides a job into sub-jobs, distributes the sub-jobs to a plurality of computers (job allocation devices), and executes job input / output data in a parallel batch processing system that executes processing involving a large amount of data access. Controls processing when access to shared data in a managed DB server is exclusive.
For example, in the aggregation process of a financial institution, access is made to the divided data that manages transactions for individual customers, and the shared data that manages data such as daily processing totals for each branch, provided in the DB server. It will be necessary. In the conventional parallel type aggregation process, when a computer accesses shared data and an exclusive process is performed because another computer is accessing the shared data, a processing error and It was. In the present invention, when a sub job by a computer is exclusively accessed for shared data, the processing is suspended without retrying, the processing state is suspended (temporarily suspended), the queue is saved, and the next new sub job is saved. Start execution.
When access to the shared data of the new sub job is successful, the sub job that requests access to the same shared data is extracted from the suspended sub job with reference to the saved queue table, and the shared data that has been waiting for processing is extracted. Aggregation processing is performed by executing the processing in a batch.

FIG. 1 is a functional block diagram illustrating a configuration example of a job allocation system 5 including a job allocation device 3 according to the present embodiment.
The job allocation system 5 is a system that executes batch processing, and includes a job input device 1, a sub job distribution device 2, a plurality of grid nodes (job allocation devices) 3 (3a, 3b, 3c), and a DB (DataBase ) The server 4 is connected via the communication network 6. Note that the number of grid nodes (job allocation devices) 3 (3a, 3b, 3c) is not limited to three, and a plurality of grid nodes 3 may be provided.

First, an outline of job assignment processing of the job assignment system 5 will be described.
The sub job distribution device 2 to which a job is input from the job input device 1 divides the job into predetermined processing units, and distributes the divided sub job to each grid node 3 (3a, 3b, 3c). Then, the grid node 3 to which the sub job is distributed executes the sub job and accesses the divided data 421 and the shared data 422 of the DB server 4.
When accessing the shared data, the grid node 3 suspends (temporarily suspends) the processing state by suspending the processing without retrying, if the access to the shared data 422 of the sub job is excluded. It is registered in the I / O processing wait queue table (shared data I / O processing wait queue information) 313, and the execution of the next new sub job is started.
Next, when the grid node 3 succeeds in processing the shared data of the new sub-job, the grid node 3 refers to the shared data I / O processing wait queue table 313 to execute a sub-job including processing for the same shared data from a plurality of suspended sub-jobs. Extract and execute shared data processing waiting for processing in a batch. At this time, resuming the processing of the suspended sub job starts from the point where the processing is interrupted, not from the beginning of the sub job. By doing so, a bottleneck caused by waiting for I / O processing is avoided.

Next, each device will be specifically described.
The job input device 1 stores the job, and transmits the stored job to the sub job distribution device 2. The job input device 1 includes a control unit 10, a memory unit 11, a storage unit 12, an input / output unit 13, and a communication unit 14.

The control unit 10 includes a job input unit 101, and the job input unit 101 extracts a job from the job net stored in the job net storage unit 122 in the storage unit 12 and stores it in the job storage unit 121. Then, the stored job is transmitted to the sub job distribution apparatus 2 via the communication unit 14.
The memory unit 11 includes storage means such as a RAM (Random Access Memory), and temporarily stores information necessary for the processing of the control unit 10.
The storage unit 12 includes a job storage unit 121 and a job net storage unit 122. The storage unit 12 includes storage means such as a hard disk and a flash memory, and stores a program and the like for performing processing of the control unit 10.
The input / output unit 13 receives an instruction regarding processing of the control unit 10 from an external device. Moreover, the information which the control part 10 produced | generated is output to an external device. The input / output unit 13 includes an input interface and an output interface.
The communication unit 14 includes a communication interface for transmitting / receiving information to / from each device via a LAN (Local Area Network) or a dedicated line.
The function of the control unit 10 is realized, for example, when the CPU develops and executes a program stored in the storage unit 12 of the job input apparatus 1 in the memory unit 11.

Next, the sub job distribution device 2 divides the job acquired from the job input device 1 into predetermined processing units, and distributes the divided sub job to each grid node 3. The sub job distribution apparatus 2 includes a control unit 20, a memory unit 21, a storage unit 22, an input / output unit 23, and a communication unit 24.
Note that the memory unit 21, the input / output unit 23, and the communication unit 24 have the same functions as the memory unit 11, the input / output unit 13, and the communication unit 14 of the job input device 1, and thus description thereof is omitted.

The control unit 20 controls the entire sub job distribution device 2 and includes a job dividing unit 201, a sub job distribution unit 202, and a sub job management unit 203. The function of the control unit 20 is realized, for example, when the CPU develops and executes a program stored in the storage unit 22 of the sub job distribution apparatus 2 in the memory unit 21.
The job dividing unit 201 divides the job acquired from the job input device 1 into predetermined processing units, generates a sub job, and stores it in the sub job storage unit 221 in the storage unit 22.
The sub job distribution unit 202 determines the grid node 3 (3a, 3b, 3c) to which the sub job is distributed, and transmits the sub job to the determined grid node 3 (3a, 3b, 3c) via the communication unit 24. In addition, a predetermined number of jobs that can be suspended (temporarily suspended) is set in the sub job distribution unit 202 when determining the distribution of sub jobs, and the sub job distribution unit 202 does not exceed the predetermined number of jobs. Sort out. Thereby, the dispersion | variation in the job execution end time between the grid nodes 3 can be suppressed.
The sub job management unit 203 stores and manages the progress information of the data processing of each sub job distributed to each grid node 3 (3a, 3b, 3c) in the sub job management table 222 (see FIG. 2 described later) in the storage unit 22. To do.

  The storage unit 22 includes a sub-job storage unit 221 that stores the sub-jobs divided by the job dividing unit 201 and a sub-job management table 222, and includes storage means such as a hard disk and a flash memory.

FIG. 2 is a diagram illustrating an example of a data configuration of the sub job management table 222 according to the present embodiment.
The sub job management table 222 is information including one record for each sub job, and includes data items of a job ID 1001, a sub job ID 1002, a distribution destination node 1003, and a progress status 1004 as shown in FIG.

  The job ID 1001 is a unique identifier for each job before being divided into sub-jobs. The sub job ID 1002 is a unique identifier assigned to each sub job divided by the job dividing unit 201. The distribution destination node 1003 is a node name of each grid node 3 (3a, 3b, 3c) to which the sub job distribution unit 202 distributes the sub job. Also, the progress status 1004 is registered by the sub job management unit 203 as “unprocessed”, “running”, or “completed” in accordance with the progress status of the data processing of the sub job.

Returning to FIG. 1, the grid node (job allocation apparatus) 3 will be described. The grid node 3 is a device that executes a sub job by accessing the DB server 4, and includes a control unit 30, a memory unit 31, a storage unit 32, an input / output unit 33, and a communication unit 34. Composed.
Note that the input / output unit 33 and the communication unit 34 have the same functions as the input / output unit 13 and the communication unit 14 of the job input apparatus 1, and thus description thereof is omitted.

The control unit 30 controls the entire grid node 3 and includes a sub job control unit 301, a sub job execution unit 302, a queue operation unit 303, and a queue management unit 304.
The function of the control unit 30 is realized, for example, when the CPU expands and executes a program stored in the storage unit 32 of the grid node 3 in the memory unit 31.

  The sub job control unit 301 refers to a sub job progress management table (sub job progress management information) 312 (FIG. 3) in the memory unit 31 described later, and searches for an unprocessed sub job. Then, the sub job control unit 301 allocates the searched unprocessed sub job to the sub job execution unit 302 via the queue operation unit 303. In addition, the sub job control unit 301 updates the sub job progress management table 312.

  The sub job execution unit 302 executes data input, data processing, data output, and the like according to the description of the sub job. At this time, the sub job execution unit 302 first executes the process for the “half-part data” in the first half of the process of the entire sub job, and then executes the process for the “shared data” as the second half. Further, the sub job execution unit 302 updates a shared data I / O processing wait queue table 313 (see FIG. 4) described later.

The queue operation unit 303 instructs to store and retrieve a sub job, and collects information related to the execution state of the sub job when the sub job enters a shared data processing waiting state.
Specifically, the queue operation unit 303 stores the sub job received from the sub job distribution device 2 via the communication unit 34 in the sub job queue 311 in the memory unit 31. The queue operation unit 303 receives a command from the sub job control unit 301, acquires the sub job from the sub job queue 311, and transmits the sub job to the sub job execution unit 302. Further, the queue operation unit 303 collects information necessary for suspending the execution state of the sub job and storing it in the queue when the sub job execution unit 302 fails to process the shared data, and performs shared data I / O processing. The data is stored in the waiting queue 314 (see FIG. 5 described later). In addition, when the sub job execution unit 302 succeeds in processing the shared data, the queue operation unit 303 assigns a sub job that processes the shared data and a sub job that requests access to the same shared data to the shared data I / O processing waiting queue. It acquires from 314 and performs processing to restore its execution state.

  The queue management unit 304 uses the sub job progress management table 312 in the memory unit 31 to check the number of sub jobs that are unprocessed or waiting for shared processing, and transmits the information to the sub job distribution device 2 via the communication unit 34. To do.

  Next, the memory unit 31 stores a sub job queue 311 in which a sub job transmitted from the sub job distribution apparatus 2 is stored, a sub job progress management table 312 (see FIG. 3), and a shared data I / O processing wait queue table 313 ( 4) and a shared data I / O processing wait queue 314 (see FIG. 5). Note that these pieces of information stored in the memory unit 31 may be stored in the storage unit 32.

FIG. 3 is a diagram showing an example of the data configuration of the sub job progress management table (sub job progress management information) 312 according to the present embodiment.
The sub job progress management table 312 is information consisting of one record for each sub job arranged in the grid node 3, and as shown in FIG. 3, data of job ID 1001, sub job ID 1002, distribution source 1005, and sub job progress status 1006 Consists of items.

  Here, the distribution source 1005 is the node name of the sub job distribution device 2 that is the distribution source of the acquired sub job. Also, the sub job progress status 1006 stores information of “unprocessed”, “running”, “waiting for shared processing”, or “completed” by the sub job control unit 301 according to the progress status of the data processing of the sub job. . This “waiting for shared processing” indicates that the sub job execution unit 302 cannot process the shared data, and the sub job is in the processing waiting state stored in the shared data I / O processing waiting queue 314 (see FIG. 5). It is shown.

FIG. 4 is a diagram showing an example of the data configuration of the shared data I / O processing wait queue table (shared data I / O processing wait queue information) 313 according to the present embodiment.
The shared data I / O processing waiting queue table 313 is information including one record for each sub job for which access to the shared data is excluded. As shown in FIG. 4, the job ID 1001, the sub job ID 1002, the queue ID 1007, and the shared data are stored. It includes a data item of ID1008.

  Here, the queue ID 1007 is a queue identifier for distinguishing between a queue on its own grid node 3 to which a sub-job is distributed and a queue on another grid node 3. For example, the queue ID 1007 of “3a_SpdQ” is assigned to the sub job assigned to the grid node 3 (3a), and “3b_SpdQ” is assigned to the sub job assigned to the grid node 3 (3b). The shared data ID 1008 indicates the identifier of the shared data 422 for which the sub job has requested processing and is waiting for processing.

FIG. 5 is a diagram showing an example of the data configuration of the shared data I / O processing wait queue 314 according to the present embodiment.
The shared data I / O processing waiting queue 314 is information for storing the suspended state of the sub job, and includes a header unit 1009 and a main body unit 1010.

  The header portion 1009 includes data items of a job ID 1001, a sub job ID 1002, a queue ID 1007, and a shared data ID 1008. These pieces of information are data items having the same contents as one record of the shared data I / O processing waiting queue table 313 shown in FIG.

The main body 1010 stores sub job 1011, checkpoint 1012, input source data name 1013, output destination data name 1014, and unoutput data value 1015.
Here, the sub job 1011 stores calculation instructions and processing described in the sub job. The check point 1012 is information indicating how far the sub job 1011 has been completed in the job step. The input source data name 1013 indicates the data name input in the sub job. The output destination data name 1014 indicates the name of data output in the sub job. An unoutput data value 1015 indicates an unoutput data item and an unoutput data value at the time of suspension.
It should be noted that by storing each piece of information in the main body 1010, it is possible to resume and process information at the time of suspension in the middle of processing from that point.

Returning to FIG. 1, the DB server 4 will be described. The DB server 4 manages job input / output data, and includes a control unit 40, a memory unit 41, a storage unit 42, an input / output unit 43, and a communication unit 44.
Note that the memory unit 41, the input / output unit 43, and the communication unit 44 have the same functions as the memory unit 11, the input / output unit 13, and the communication unit 14 of the job input device 1, and thus description thereof is omitted.

The control unit 40 includes a data control unit 401 and controls the entire DB server 4. Note that the function of the control unit 40 is realized by, for example, the CPU storing and executing a program stored in the storage unit 42 of the DB server 4 in the memory unit 41.
The data control unit 401 processes the divided data 421 and the shared data 422 in the storage unit 42 based on the sub job execution instruction from the sub job execution unit 302 of the grid node 3. The data control unit 401 manages data input / output using the data management table 423 in the storage unit 42 and updates the lock table 424 (see FIG. 6 described later) in the data management table 423.
Specifically, when the shared data 422 is locked (inaccessible) by a certain grid node 3, the data control unit 401, when there is an access request to the shared data 422 from another grid node 3, The exclusive control of the sub job is performed, and the information of the sub job subjected to the exclusive control is stored in the lock table 424 (see FIG. 6) described later. Further, when the shared data 422 is not locked, the data control unit 401 registers the queue ID 1007 of the grid node 3 that is the transmission source of the sub job that has accessed the shared data 422 in the lock table 424.
Since the queue ID 1007 is registered in the lock table 424, the data control unit 401 confirms that the shared data 422 is locked by the grid node 3.

FIG. 6 is a diagram illustrating an example of a data configuration of the lock table 424 according to the present embodiment.
The lock table 424 stores, for each shared data 422, information on a queue that has acquired the lock when the shared data 422 is locked, and information on a sub-job requesting the lock on the shared data 422. The lock table 424 includes data items of a shared data ID 1008, a lock acquisition queue 1016, and a lock request sub job 1017.

Here, the shared data ID 1008 is an identifier of the shared data 422 that has been waiting for processing. The lock acquisition queue 1016 stores the queue ID 1007 of the grid node 3 that holds the lock authority for the shared data 422. Also, in the lock request sub job 1017, the job ID, sub job ID, and node name of the sub job waiting to release the lock for processing the shared data 422 are registered. The lock request sub-job 1017 stores the above-described information of the sub-jobs waiting for processing because the shared data is locked, in the order of waiting for processing.
In this way, by managing the authority to lock and unlock the shared data 422 with the queue ID 1007 stored in the lock acquisition queue 1016 of the grid node 3, it is possible to avoid interrupt processing from other grid nodes 3. it can.

  Returning to FIG. 1, the divided data 421 that is exclusively processed by any one of the grid nodes 3 (3 a, 3 b, 3 c) and all the grid nodes 3 are commonly read into the storage unit 42 of the DB server 4. , Shared data 422 to be processed and written, and a data management table 423. The data management table 423 stores the lock table 424 used when the data control unit 401 performs exclusive data control.

  Next, a job allocation method by the job allocation system 5 including the grid node (job allocation apparatus) 3 according to the present embodiment will be specifically described with reference to FIGS. 7 to 10 and FIGS. To do.

  7 to 10 are sequence diagrams for explaining the job assignment method according to the present embodiment. Note that the transmission of the job to the sub job distribution device 2 by the job input device 1, the division of the job into sub jobs by the job division unit 201 of the sub job distribution device 2, and the storage in the sub job storage unit 221 have been completed. explain.

  First, as illustrated in FIG. 7, the queue management unit 304 of the grid node 3 (3 a) refers to the sub job progress management table 312 (FIG. 3), and confirms the number of sub jobs that are unprocessed or waiting for shared processing (step S 701). ). Then, the queue management unit 304 transmits the confirmation result to the sub job distribution unit 202 of the sub job distribution device 2 (step S702).

Next, the sub job distribution unit 202 determines a grid node to which the sub job is distributed when the confirmation result in step S701 does not exceed a predetermined upper limit value (step S703). The sub job distribution unit 202 does not perform sub job distribution processing when a predetermined upper limit is exceeded.
Subsequently, the sub job distribution unit 202 acquires the sub job from the sub job storage unit 221 (step S704), distributes the sub job to each grid node 3 (3a, 3b, 3c) (step S705), and distributes the sub job. The sub job is transmitted to each grid node 3 (3a, 3b, 3c) (step S706). Here, the processing will be described on the assumption that sub-jobs are distributed to the grid node 3 (3a). Further, the sub job distribution unit 202 updates the sub job management table 222 (see FIG. 2) via the sub job management unit 203 (step S707). For example, for the record corresponding to the distributed sub-job, the distribution destination node 1003 is updated to the distribution-destination grid node “3a”, and the progress status 1004 is updated from “unprocessed” to “running” (step S707).

  Next, the queue operation unit 303 of the grid node 3 (3 a) stores the sub job received in step S 706 in the sub job queue 311 (see FIG. 1) (step S 708), and the new sub job is displayed in the grid for the sub job control unit 301. An update request for information associated with the placement at the node 3 (3a) is made (step S709). Then, the sub job control unit 301 adds a new record to the sub job progress management table 312 (see FIG. 3), and registers the job ID 1001, the sub job ID 1002, the distribution source 1005, and the sub job progress status 1006 based on the acquired data. (Step S710). At this time, the sub job progress status 1006 of the newly added record is “unprocessed”.

  Subsequently, the sub job control unit 301 refers to the sub job progress management table 312 (FIG. 3) and searches for the presence or absence of an unprocessed sub job (step S711). If there is an unprocessed sub job, the sub job control unit 301 instructs the queue operation unit 303 to allocate the sub job to the sub job execution unit 302 (step S712), and transmits this sub job allocation command (step S713). Further, the sub job control unit 301 updates the sub job progress status 1006 from “unprocessed” to “running” for the record corresponding to the sub job in the sub job progress management table 312 (step S714).

  The queue operation unit 303 acquires the sub job instructed in step S713 from the sub job queue 311 (see FIG. 1) (step S715), and transmits the sub job to the sub job execution unit 302 (step S716).

  Next, the sub job execution unit 302 activates the sub job (step S717), starts the divided data processing of the first half of the sub job (step S718), and transmits the divided data 421 via the data control unit 401 of the DB server 4 (FIG. 1) is requested (step S719). Since the divided data 421 can be exclusively used by the grid node 3 (3a) and lock contention does not occur, the divided data 421 is promptly transmitted to the sub job execution unit 302 (step S720), and the processing of the divided data 421 is executed by the sub job execution unit 302. (Step S721).

  Subsequently, as shown in FIG. 8, the sub job execution unit 302 starts the shared data processing of the second half of the sub job (step S722), and the shared data 422 (see FIG. 1) via the data control unit 401 of the DB server 4. ) Is requested (step S723). Here, when the shared data 422 is locked by another grid node 3 (3b, 3c), the data control unit 401 of the DB server 4 performs exclusive control (step S724). Subsequently, the data control unit 401 registers, in the lock request sub job 1017 of the lock table 424 (see FIG. 6), the job ID, sub job ID, and node name related to the sub job that requested access to the shared data 422 (step S725). ). Next, the data control unit 401 transmits, as a lock processing wait notification, to the sub job execution unit 302 of the grid node 3 (3a) that the shared data 422 is locked (step S726).

  Upon receiving the lock processing wait notification, the sub job execution unit 302 stops the execution of the sub job (step S727), the job ID 1001, the sub job ID 1002, the queue ID 1007 in the shared data I / O processing wait queue table 313 (see FIG. 4), Then, the shared data ID 1008 is registered (step S728). Then, the sub job execution unit 302 generates a message indicating that the shared data processing has failed (step S729), and transmits the message to the sub job control unit 301 (step S730).

  Next, the sub job control unit 301 generates sub job registration command information using as arguments the job ID 1001, sub job ID 1002, queue ID 1007, and shared data ID 1008 of the sub job that failed in the shared data processing acquired in step S730 (step S731). ) To the queue operation unit 303 (step S732).

  Subsequently, the queue operation unit 303 suspends the execution status of the sub job specified by the sub job registration command information in step S732 and stores the information in the queue, that is, check point 1012, sub job input source data name 1013, the output destination data name 1014, and the non-output data value 1015 of the sub job are collected (step S733). Then, the queue operation unit 303 generates information including the header unit 1009 and the main body unit 1010, and registers the information as the shared data I / O processing wait queue 314 (see FIG. 5) (step S734). Subsequently, the queue operation unit 303 transmits the completion of the registration process to the shared data I / O processing waiting queue 314 to the sub job control unit 301 (step S735).

  Next, when receiving the registration completion notification in step S735, the sub job control unit 301 updates the sub job progress status 1006 in the sub job progress management table 312 (see FIG. 3) from “in progress” to “waiting for shared processing” ( Step S736).

  With the above processing, since one sub job has been suspended due to the shared data processing failure, processing for the next new sub job is started (step S737).

  FIG. 9 shows the processing from step S711 to step S721 in FIG. 7 in the processing for the new sub job started in step S737 in FIG. 8, that is, the processing after the execution of the divided data 421 in the first half of the sub job is completed. Show. Hereinafter, processing when the shared data 422 of the DB server 4 is not locked will be described with reference to FIGS. 9 and 10.

  As shown in FIG. 9, the sub job execution unit 302 starts shared data processing for the latter half of the sub job in the same manner as in FIG. 7 (step S722), and transfers the shared data 422 to the shared data 422 via the data control unit 401 of the DB server 4. Access is requested (step S723). If the shared data 422 is not locked, the data control unit 401 of the DB server 4 registers the queue ID 1007 of the grid node 3 (3a) in the lock acquisition queue 1016 of the lock table 424 (see FIG. 6) ( Step S738). This means that the queue ID 1007 of the grid node 3 (3a) is locked with respect to the shared data 422 of the DB server 4 is locked. Then, the data control unit 401 transmits the shared data 422 to the sub job execution unit 302 (step S739).

  The sub job execution unit 302 that has received the shared data 422 executes shared data processing (step S740). When the sub job is completed due to the completion of the shared data processing, the sub job execution unit 302 generates a sub job processing completion message (step S741) and transmits it to the sub job control unit 301 (step S742).

  Next, the sub job control unit 301 updates the sub job progress status 1006 of the sub job progress management table 312 (see FIG. 3) from “in progress” to “completed” (step S743). Then, the sub job control unit 301 generates a sub job completion notification (step S744) and transmits it to the sub job distribution unit 202 of the sub job distribution device 2 (step S745).

  Subsequently, the sub job distribution unit 202 of the sub job distribution apparatus 2 that has received the sub job completion notification changes the progress status 1004 of the sub job management table 222 (see FIG. 2) from “in progress” to “ It is updated to “completed” (step S746).

  On the other hand, the sub job control unit 301 of the grid node 3 (3a) searches the shared data I / O processing waiting queue table 313 (see FIG. 4), and the shared data 422 that is the same as that acquired in step S738 for the shared data ID 1008. It is searched whether or not an access request for is included (step S747). In other words, the shared data I / O processing waiting queue table 313 is searched for the same shared data ID 1008 as the shared data 422 that has been processed. If there is a sub job that matches the shared data ID 1008, the sub job control unit 301 generates sub job acquisition command information (step S748) and transmits it to the queue operation unit 303 (step S749). The sub job acquisition command information is information including a job ID, a sub job ID, a queue ID, and a shared data ID as arguments.

  Next, the queue operation unit 303 acquires the sub job specified by the sub job acquisition command information in step S749 from the shared data I / O processing wait queue 314 (see FIG. 5), and performs processing for restoring the execution state (step S749). S750). Specifically, the check point 1012 stored in the main body 1010 of the shared data I / O processing waiting queue 314 is referred to, and the execution position of the job step is skipped before suspending. Then, the input source data name 1013, output destination data name 1014, and unoutput data value 1015 of the sub job stored in the main body 1010 are set in the restored sub job. Then, the queue operation unit 303 notifies the sub job control unit 301 that the sub job execution state restoration processing has been completed (step S751).

  Next, as shown in FIG. 10, the sub job control unit 301 receives that the sub job restoration processing has been completed, and issues a command to the queue operation unit 303 to allocate the restored sub job to the sub job execution unit 302 ( Step S752). Thereafter, the execution processing of the shared data 422 in the latter half of the sub-job is performed in the same manner as steps S713 to S717 in FIG. 7 and steps S722 to S746 in FIG. 9 (step S753). In step S745, after transmitting the sub job completion notification, the sub job control unit 301 performs an update to delete the record indicating the sub job for which processing has been completed from the shared data I / O processing wait queue table 313 (see FIG. 4) ( Step S754).

Thereafter, the shared data I / O processing in steps S747 to S754 is repeatedly executed until there is no corresponding sub job in the shared data I / O processing waiting queue table 313 (step S755). Then, in step S747, the shared data I / O processing wait queue table 313 is searched, and if the corresponding sub job does not exist, the shared data I / O loop processing is terminated (step S756).
Subsequently, the sub job control unit 301 generates unlock command information of the shared data 422 (step S757) and transmits it to the data control unit 401 of the DB server 4 (step S758).

  Next, the data control unit 401 updates the lock table 424 (see FIG. 6) based on the acquired unlock command information (step S759). Specifically, the data control unit 401 clears the lock acquisition queue 1016 of the record indicating the shared data ID 1008 accessed by the sub job, and deletes the sub job having the node name 3a from the lock request sub job 1017. Then, the data control unit 401 transmits the completion of the unlock process to the sub job control unit 301 of the grid node 3 (3a) (step S760).

  Upon receiving the completion of the unlock process, the sub job control unit 301 resumes the process for the new sub job (step S761). Specifically, the process returns to step S711 in FIG.

  The sub job distribution unit 202 of the sub job distribution device 2 receives the confirmation result of the number of sub jobs waiting for unprocessed or shared processing from the queue management unit 304 of the grid node 3 at a predetermined interval (step S702). The sub job distribution process in steps S703 to S707 is repeated. In step S704, when there is no sub job to be distributed, that is, when the job processing is completed, a job end notification is transmitted to the job input device 1, and the sub job distribution is awaited.

Next, the flow of sub job processing when the job allocation program according to the present embodiment is executed will be described by showing the location of data relating to each sub job.
FIG. 11 is a diagram for explaining a time-series change in the location of the sub job according to the present embodiment.

In FIG. 11, T0 to T13 indicate time axes that continue from T0 before execution of the job assignment program to T13 after completion of execution. The blocks in FIG. 11 include, in order from the left, a sub job storage unit 221 of the sub job distribution device 2, a sub job queue 311 of the grid node 3 (3a), a shared data I / O processing wait queue 314, and a sub job execution unit 302. Show. A circled number indicates a sub job stored in the computer, and the number represents a sub job ID 1002.
In FIG. 11, the location of the sub job at a certain point is expressed by the circled characters. Hereinafter, an example of the flow of sub job processing will be described along the time axis.

At time T0, the sub jobs “1” to “10” are stored in the sub job storage unit 221, that is, the sub job is not yet distributed to the grid node 3.
At time T1, the sub jobs “1” and “3” are distributed to the grid node 3 (3a) and stored in the sub job queue 311 (FIG. 1).
Next, at time T <b> 2, the sub job “1” is executed by the sub job execution unit 302, while the sub job “5” is further stored in the sub job queue 311.
Subsequently, since the shared data of the sub job “1” could not be processed at time T3, it is suspended and stored in the shared data I / O processing waiting queue 314 (FIG. 5). Next, the new sub job “3” is processed by the sub job execution unit 302.
At time T4, since the shared data 422 of the sub job “3” could not be processed, it is suspended and stored in the shared data I / O processing waiting queue 314. Next, a new sub job “5” is processed by the sub job execution unit 302.
At time T5, the sub job “5” disappears because the shared data processing is completed, and the sub job “1” is then reassigned to the sub job execution unit 302.
At time T6, the sub job “1” disappears because the shared data processing is completed. Then, the sub job “3” is processed by the sub job execution unit 302, and the shared data processing is repeatedly executed until the shared data I / O processing waiting queue 314 becomes empty.
Next, at time T7, a new sub job “7” is processed by the sub job execution unit 302. Thereafter, processing from time T8 to time T12 is performed, and processing of all the sub jobs is completed at time T13.

  As described above, according to the job allocation apparatus, the job allocation method, and the job allocation program according to the present invention, in a parallel batch processing system including a plurality of nodes, bottles due to waiting for I / O processing for the same data are performed. The bottleneck can be avoided and the utilization efficiency of the computation node can be improved.

1 Job submission device 2 Job distribution device 3 Grid node (job allocation device)
4 DB server 5 Job allocation system 6 Communication network 10, 20, 30, 40 Control unit 11, 21, 31, 41 Memory unit 12, 22, 32, 42 Storage unit 13, 23, 33, 43 Input / output unit 14, 24 , 34, 44 Communication unit 101 Job input unit 121 Job storage unit 122 Job net storage unit 201 Job division unit 202 Sub job distribution unit 203 Sub job management unit 221 Sub job storage unit 222 Sub job management table 301 Sub job control unit 302 Sub job execution unit 303 Queue Operation unit 304 Queue management unit 311 Sub job queue 312 Sub job progress management table (sub job progress management information)
323 Shared data I / O processing wait queue table (shared data I / O processing wait queue information)
324 Shared data I / O processing waiting queue 401 Data control unit 421 Division data 422 Shared data 423 Data management table 424 Lock table

Claims (6)

A job distribution device that divides a job into sub-jobs of a predetermined processing unit and distributes the jobs to a plurality of job allocation devices, and processes the distributed sub-jobs by accessing divided data and shared data stored in a DB server The job allocation apparatus provided in a job allocation system in which a plurality of job allocation apparatuses and the DB server that manages access of the divided data and the shared data for each job allocation apparatus are connected via a communication network. And
Sub job queue in which the sub job is stored, sub job progress management information in which the progress status of the processing state of the sub job is stored, and shared data I / O processing wait queue information in which a sub job of the sub job waiting for shared data processing is stored A storage unit for storing
Updating the sub job progress management information and referring to the sub job progress management information to search for an unprocessed sub job;
A sub job execution unit that acquires the searched unprocessed sub job from the sub job queue and executes the sub job by accessing the divided data and the shared data of the DB server;
When the sub job execution unit fails to process the shared data, the execution of the sub job is temporarily suspended, and the progress status of the suspended sub job waiting for processing is displayed as the shared data I / O processing queue information. A queue operation unit registered in
The sub job execution unit, after temporarily suspending the execution state of the sub job, acquires the new unprocessed sub job from the sub job queue and processes the shared data,
When the processing is successful, the sub job control unit waits for the shared data I / O processing to wait for a sub job that requests access to the same shared data as the shared data for which the processing has succeeded among the sub jobs waiting for processing. Extracted from queue information,
The job assignment apparatus, wherein the sub job execution unit executes processing of the shared data of the extracted sub job. The storage unit further includes a shared data I / O processing waiting queue for storing information related to the suspended execution state of the sub job when the execution of the sub job is suspended.
The queue operation unit, when the sub job execution unit fails to process the shared data, collects information related to the execution state of the sub job and stores it in the shared data I / O processing waiting queue.
Further, when the sub job control unit extracts the sub job waiting for processing from the shared data I / O processing wait queue information, the queue operation unit displays information regarding the execution state of the extracted sub job waiting for processing. Retrieve from the shared data I / O processing queue, restore the execution status of the sub job when it was interrupted,
The job assignment apparatus according to claim 1, wherein the sub job execution unit executes processing of the sub job waiting for processing from the time point of the temporarily suspended execution state.   The information about the execution status of the sub job that has been interrupted includes a checkpoint indicating the degree of progress of the entire processing of the sub job, an input source data name and an output destination data name of the sub job, and information at the time of the temporary suspension is not stored. The job assignment apparatus according to claim 2, wherein the job assignment apparatus is an output data item and a non-output data value. A job distribution device that divides a job into sub-jobs of a predetermined processing unit and distributes the jobs to a plurality of job allocation devices, and processes the distributed sub-jobs by accessing divided data and shared data stored in a DB server A plurality of job allocation devices and the DB server that manages access to the divided data and the shared data for each job allocation device are used for the job allocation device provided in a job allocation system connected via a communication network. A job assignment method, comprising:
The job allocation device includes:
Sub job queue in which the sub job is stored, sub job progress management information in which the progress status of the processing state of the sub job is stored, and shared data I / O processing wait queue information in which a sub job of the sub job that is waiting for shared data processing is stored A storage unit for storing
Searching for an unprocessed sub-job with reference to the sub-job progress management information;
Obtaining the retrieved unprocessed sub job from the sub job queue and executing the sub job by accessing the divided data and the shared data of the DB server;
When the processing of the shared data fails, temporarily suspending the execution of the sub job, and registering the progress status of the suspended sub job waiting for processing in the shared data I / O processing waiting queue information;
After temporarily suspending the execution state of the sub job, obtaining a new unprocessed sub job from the sub job queue and processing the shared data;
A step of extracting, from the shared data I / O processing waiting queue information, a sub job that requests access to the same shared data as the shared data for which the processing has succeeded among the sub jobs waiting for processing when the processing is successful When,
Executing the processing of the shared data of the extracted sub-job;
A job assignment method characterized by executing The storage unit further includes a shared data I / O processing waiting queue for storing information related to the suspended execution state of the sub job when the execution of the sub job is suspended.
Collecting information related to the execution state of the sub-job when the shared data processing fails, and storing the information in the shared data I / O processing waiting queue;
When the sub job waiting for processing is extracted from the shared data I / O processing waiting queue information, information on the execution state of the extracted processing waiting sub job is acquired from the shared data I / O processing waiting queue, and A step of restoring the execution state of the sub job at the time of interruption;
Executing the processing of the sub-job waiting for the processing from the time of the suspended execution state;
The job assignment method according to claim 4, further comprising:   A job assignment program for causing a computer to execute the job assignment method according to claim 4 or 5.

Images