JP6485343B2 - Control program, control device, and control method - Google Patents

Description

  The present invention relates to a control program, a control device, and a control method.

  Conventionally, there is a technique called in-memory type data processing in which data processing is performed using an internal memory without an external I / O (Input / Output) such as a disk. Furthermore, there is a technology for causing JavaVM (JVM: Java Virtual Machine Java is a registered trademark) to execute in-memory data processing. In in-memory data processing, the processing area allocated by the memory of the machine that executes the data processing task and used by the task when the task is repeatedly executed, and the transfer area for transferring the task execution result Region is used.

  As a related prior art, for example, based on the measurement result obtained by measuring changes in the usage status of the internal memory and the external memory, the internal memory configuration is changed, the data stored in the internal memory and the external memory is moved, the internal memory or Some change the area of external memory. In addition, CPU memory demand prediction amount is obtained based on the amount of memory secured from a portion that can be assigned to all CPUs by a CPU (Central Processing Unit) at a predetermined time, and secured from the portion that can be assigned to all CPUs based on the obtained prediction amount There is a technique for transferring the memory that has been allocated to a portion allocated only to a predetermined CPU. Also, when transferring data by remote memory writing on a parallel computer, the data transfer memory area on each computer is divided into an area that is fixed in advance for each communication partner and an area that is dynamically allocated. There is a technique for selectively using these areas depending on the data length. There is a technique for securing a page fixed area for use as a transmission / reception data area in the physical address space of each node based on the size set when the operating system of each node is activated.

International Publication No. 2011/045931 JP-A-10-320358 Japanese Patent Laid-Open No. 10-143486 JP-A-6-110845

  However, according to the prior art, it is difficult to determine whether or not it is necessary to change memory allocation in in-memory data processing. For example, if the data size of the execution result of the data processing task exceeds the size of the transfer area, writing to the disk is performed, and the time required for data processing increases by the amount written to the disk. Therefore, even if the allocation amount of the transfer area is changed, it takes time to restart the JVM that executes the task to change the allocation. Therefore, even if the memory allocation is changed, an increase in time required for data processing cannot always be suppressed.

  In one aspect, an object of the present invention is to provide a control program, a control apparatus, and a control method that can appropriately determine whether or not a memory allocation change is necessary.

  According to one aspect of the present invention, among the storage areas of the memory, the size of the processing area used by the task when the task is repeatedly executed and the size of the transfer area for transferring the task execution result are obtained, Based on the usage amount, change the size of the transfer area based on the usage amount of the processing area used by the task that was used when the task was executed and the acquired size of the processing area and the size of the transfer area In this case, it calculates a reduction time that shortens the time required to repeatedly execute the task, and determines whether to change the memory allocation based on the calculated reduction time and the time required to change the memory allocation. A control program, a control device, and a control method are proposed.

  According to one aspect of the present invention, it is possible to appropriately determine whether or not it is necessary to change memory allocation.

FIG. 1A is an explanatory diagram (part 1) of an operation example of the control device 101 according to the present embodiment. FIG. 1B is an explanatory diagram (part 2) of an operation example of the control device 101 according to the present embodiment. FIG. 2 is an explanatory diagram showing a configuration example of the in-memory data processing system 200. FIG. 3 is an explanatory diagram illustrating a hardware configuration example of the master machine 201. FIG. 4 is a block diagram illustrating a functional configuration example of the master machine 201. FIG. 5 is an explanatory diagram showing an example of assignment of tasks 111 when a single machine performs data processing. FIG. 6 is a flowchart illustrating an example of a data processing procedure in a single machine. FIG. 7 is an explanatory diagram showing an example of assignment of tasks 111 when data processing is performed on a plurality of machines. FIG. 8 is an explanatory diagram showing an execution example of the task 111 when data processing is performed on a plurality of machines. FIG. 9 is a flowchart illustrating an example of a data processing procedure in a plurality of machines.

  Hereinafter, embodiments of a disclosed control program, control device, and control method will be described in detail with reference to the drawings.

  1A and 1B are explanatory diagrams illustrating an operation example of the control device 101 according to the present embodiment. The control device 101 is a computer that controls a machine that executes in-memory data processing. The control device 101 is, for example, a server. Here, in-memory type data processing is data processing using an internal memory without accompanying external I / O such as a disk. In in-memory data processing, the processing result existing in the memory can be used as it is for the next processing.

  Here, one processing of data processing is referred to as “task”. In application software that implements machine learning that repeats the same task, in-memory data processing is effective. Hereinafter, the application software is referred to as “application”. Here, the task indicates a minimum execution unit obtained by dividing a job that is an application execution unit.

Examples of machine learning are shown in References 1 and 2 below.
(Reference Document 1: Machine Learning Algorithm in the Age of Big Data: Online Learning, [Search September 24, 2015], http://www.nttdata.com/jp/ja/insights/trend_keyword/2014110601.html)
(Reference 2: Introduction to classic machine learning 7. Online learning, [searched on September 24, 2015], http://www.slideshare.net/hiroshinakagawa3/ss-322774089)

  For this reason, in-memory type data processing is assigned so as to be distributed in parallel to machines that execute in-memory type data processing by using repetition of tasks between the same machine and different machines. In-memory data processing integrates the processing area used when executing the application and the task execution results stored in multiple processing areas, or part or all of the task execution results. And a transfer area used for transferring the file to another machine.

  The processing contents of the machine learning application can be arbitrarily specified by the user. Further, the data size of the processing area used for the task differs depending on the task. The sum of the processing area and the transfer area does not exceed the total amount of memory of the device that executes the task. If the processing area is insufficient during task execution, task execution is delayed. In addition, when the transfer area is insufficient during the execution of the task, a part of the task execution result may be written to a disk of a device that executes the task, for example, a magnetic disk drive (HDD). is there.

  Further, the control apparatus 101 causes the JVM to execute as a machine that executes in-memory data processing. The apparatus that operates the JVM may be one or more apparatuses different from the control apparatus 101, may be the control apparatus 101, or may be one or more apparatuses different from the control apparatus 101 and the control apparatus 101. When the JVM executes in-memory data processing, a processing area and a transfer area are allocated when the JVM is activated.

  However, in in-memory data processing, it is difficult to determine whether it is necessary to change the allocation between the processing area and the transfer area. Hereinafter, the allocation change between the processing area and the transfer area is referred to as “memory allocation change”. For example, if the data size of the execution result of the data processing task exceeds the size of the transfer area, writing to the disk is performed, and the time required for data processing increases by the amount written to the disk. Therefore, even if the allocation amount of the transfer area is changed, it takes time to restart the JVM that executes the task to change the allocation. Therefore, even if the memory allocation is changed, an increase in time required for data processing cannot always be suppressed.

  Here, in the case of machine learning, if the amount of memory executed in the first task is known, almost the same amount of memory is used in the second and subsequent tasks. Therefore, in the present embodiment, when an in-memory type data processing task is executed, it is possible to shorten the time by increasing the transfer area based on the usage amount of the processing area and the time required for changing the memory allocation. Then, a method of determining whether or not the memory allocation needs to be changed is performed.

  An operation example of the control apparatus 101 will be described with reference to FIGS. 1A and 1B. In the example of FIGS. 1A and 1B, it is assumed that the device that operates the JVM is the execution device 102. As illustrated in FIG. 1A, the control device 101 is connected to a plurality of execution devices 102. The execution device 102 executes the task 111 n times in response to an instruction from the control device 101. n is an integer of 2 or more.

  As illustrated in (1) of FIG. 1A, the control device 101 acquires from the execution device 102 the size of the processing region mp and the size of the transfer region mt allocated from the storage area of the memory of the execution device 102. To do. As the acquisition timing, the control device 101 may acquire the data before performing the data processing, or may acquire the data during the data processing.

  Next, as shown by (2) in FIG. 1B, the control device 101 calculates the execution time Tb as the first time and the execution time Ta as the second time, and subtracts the execution time Ta from the execution time Tb. The calculated time rt is calculated. The time rt is a time during which the task 111 is repeatedly executed when the size of the transfer area mt is changed based on the usage amount u, that is, the time required for executing the data processing is shortened. The time rt is also a time for shortening the time required for executing the data processing in consideration of the time required for the memory allocation change. The execution time Ta is the execution time of data processing when the memory allocation is changed. The execution time Tb is the execution time of data processing when the memory allocation is not changed.

  The control device 101 calculates an execution time Tb based on the time T1 taken when the task 111 is executed. For example, the control device 101 calculates a value obtained by multiplying the time T1 by n as Tb.

  In addition, the control device 101 uses T1, the usage amount u of the processing area mp used by the task 111 used when the task 111 is executed, the time rat for changing the memory allocation, the size of the processing area mp, and An execution time Ta is calculated based on the size of the transfer area mt. The time rat required to change the memory allocation is the time required to restart the JVM in order to change the memory allocation. For example, the control apparatus 101 calculates Ta according to the following formula (1).

  Ta = T1 + T2 × (n−1) + rat (1)

  Here, T2 is the time taken to execute the task 111 when the memory allocation is changed. For example, the control apparatus 101 calculates, as T2, a value obtained by multiplying T1 by the reciprocal of the magnification at which the transfer area mt increases due to the memory allocation change. In the example of FIG. 1A, the amount of increase in the transfer area mt is a hatched shaded area portion that is a free space obtained by removing the usage amount u from the processing area mp.

  The graph 112 in FIG. 1B illustrates the amounts indicated by the execution times Ta, Tb, the time rt, and the time rat required to change the memory allocation. As shown in the graph 112, the time rt is a value obtained by subtracting Ta from Tb.

  The execution times Ta, Tb, and time rt may be calculated at any time after the task 111 is executed once and before the n-th execution of the task 111. After running once.

  Then, as indicated by (3) in FIG. 1B, the control device 101 determines whether to change the memory allocation based on the time rt. For example, if the time rt is 0 or more, the control apparatus 101 determines that the memory allocation is to be changed, and otherwise determines that the memory allocation is not to be changed.

  As described above, the control device 101 can appropriately determine whether or not the memory allocation needs to be changed in consideration of the time rat required for the memory allocation change, and as a result, the time required for the data processing can be reduced. it can.

  Note that the method for determining a change in memory allocation is not limited to the method described with reference to FIG. 1B. For example, the control device 101 uses the usage u of the processing area mp used by the task 111 used when the task 111 is executed, and the acquired size of the processing area mp and the size of the transfer area mt. Thus, the shortening time may be calculated. This shortening time is a time for which the task 111 is repeatedly executed when the size of the transfer area mt is changed based on the usage amount u, that is, the time required for executing the data processing is shortened. In addition, this shortening time is a time for shortening the time required for executing the data processing when the time rat for changing the memory allocation is not taken into consideration. For example, the shortening time is obtained by the following equation (2).

  Reduction time = (T1−T2) × (n−1) (2)

  Here, the expression (2) is an expression that can be derived from the expression (1) and Tb = T1 × n as follows.

Reduction time = Tb- (Ta-rat)
⇔ Shortening time = T1 × n− (T1 + T2 × (n−1))
⇔ Shortening time = T1 × (n−1) −T2 × (n−1)
⇔ shortening time = (T1-T2) × (n−1)

  Then, the control device 101 determines whether or not to change the memory allocation based on the calculated reduction time and the time rat required for the memory allocation change. As a specific determination example, the control apparatus 101 compares the shortening time and the time rat required for changing the memory allocation. If the shortening time is small, the control device 101 determines that the memory allocation is to be changed. It is determined that the assignment of is not changed.

  In the description of FIGS. 1A and 1B, the JVM executes in-memory data processing, but the present invention is not limited to this. For example, a dedicated application may execute in-memory data processing. The dedicated application cannot change the assignment between the processing area mp and the transfer area mt unless it is once ended. In this case, the time rat required to change the memory allocation is the time required to restart the dedicated application. Next, an example in which the control device 101 is applied to an in-memory data processing system will be described with reference to FIG.

  FIG. 2 is an explanatory diagram showing a configuration example of the in-memory data processing system 200. The in-memory type data processing system 200 includes a master machine 201, slave machines 202-1 to m, and a client terminal 203. The master machine 201 and slave machines 202-1 to 202-1m are connected via a network 211, respectively. The master machine 201 and the client terminal 203 are connected via a network 212. Here, the master machine 201 corresponds to the control device 101 shown in FIG. 1A. The slave machines 202-1 to 202-1m correspond to the execution device 102 shown in FIG. 1A.

  The master machine 201 is a computer that controls the in-memory data processing system 200. The slave machines 202-1 to 202-m are computers that execute the task 111 in accordance with instructions from the master machine 201. The client terminal 203 is a computer operated by a user who uses the in-memory data processing system 200.

  Next, FIG. 3 shows a hardware configuration example of the master machine 201 as computer hardware included in the in-memory data processing system 200. Since the hardware configuration of the slave machine 202 is the same as the hardware configuration of the master machine 201, description thereof is omitted.

(Hardware of master machine 201)
FIG. 3 is an explanatory diagram illustrating a hardware configuration example of the master machine 201. In FIG. 3, the master machine 201 includes a CPU 301, a ROM (Read-Only Memory) 302, and a RAM (Random Access Memory) 303. The master machine 201 includes a disk drive 304 and a disk 305, and a communication interface 306. The CPU 301 to the disk drive 304 and the communication interface 306 are connected by a bus 307, respectively.

  The CPU 301 is an arithmetic processing device that controls the entire master machine 201. The CPU 301 may have a plurality of processor cores.

  The ROM 302 is a nonvolatile memory that stores programs such as a boot program. A RAM 303 is a volatile memory used as a work area for the CPU 301.

  The disk drive 304 is a control device that controls reading and writing of data with respect to the disk 305 according to the control of the CPU 301. As the disk drive 304, for example, a magnetic disk drive, an optical disk drive, a solid state drive, or the like can be adopted. The disk 305 is a nonvolatile memory that stores data written under the control of the disk drive 304. For example, when the disk drive 304 is a magnetic disk drive, a magnetic disk can be adopted as the disk 305. Further, when the disk drive 304 is an optical disk drive, an optical disk can be adopted as the disk 305. When the disk drive 304 is a solid state drive, a semiconductor memory formed by a semiconductor element, that is, a so-called semiconductor disk can be used as the disk 305.

  The communication interface 306 is a control device that controls an internal interface with the networks 211 and 212 and controls input / output of data from other devices. Specifically, the communication interface 306 is connected to another device via the networks 211 and 212 through a communication line. As the communication interface 306, for example, a modem or a LAN adapter can be employed.

  When the administrator of the in-memory data processing system 200 directly operates the master machine 201, the master machine 201 may have hardware such as a display, a keyboard, and a mouse. The client terminal 203 also includes hardware equivalent to the hardware included in the master machine 201 and hardware such as a display, a keyboard, and a mouse.

(Example of functional configuration of master machine 201)
FIG. 4 is a block diagram illustrating a functional configuration example of the master machine 201. The master machine 201 has a control unit 400. The control unit 400 includes an acquisition unit 401, a determination unit 402, a calculation unit 403, a determination unit 404, and an assignment change unit 405. The control unit 400 realizes the functions of the respective units when the CPU 301 executes a program stored in the storage device. Specifically, the storage device is, for example, the ROM 302, the RAM 303, the disk 305, etc. shown in FIG. The processing results of each unit are stored in a register of the CPU 301, a cache memory of the CPU 301, a RAM 303, and the like.

  Hereinafter, each part will be described separately for a case where a single machine performs data processing and a case where a plurality of machines perform data processing. Here, whether the data processing is performed by a single machine or a plurality of machines may be determined by the in-memory data processing system 200 or determined by an application that performs data processing. Alternatively, the master machine 201 may determine whether a single machine or a plurality of machines performs data processing based on the load status of the slave machines 202-1 to 202-m. For example, it is assumed that an application that performs data processing can be executed by a plurality of machines. Then, if the loads on the slave machines 202-1 to m are substantially equal, the master machine 201 may determine that the slave machines 202-1 to m are to perform data processing. Alternatively, the master machine 201 may determine that one of the slave machines 202-1 to 202-m that has a smaller load than the other machine performs data processing.

  First, each part when a single machine performs data processing will be described. The acquisition unit 401 acquires the size of the processing area mp and the size of the transfer area mt of the slave machine 202.

  Based on the execution history of the task 111, the determination unit 402 determines whether or not the execution result of the task 111 is written to another storage area different from the storage area of the memory during the execution of the task 111. The other storage area different from the storage area included in the memory is, for example, a storage area on a disk different from the RAM included in the slave machine 202. Further, the other storage area may be an external storage device other than the disk, which can be accessed by the slave machine 202. In the execution history of the task 111, the amount of data written to the disk without the execution result of the task 111 being written in the transfer area mt is described. Accordingly, for example, if the amount of data written to the disk without the execution result of the task 111 being recorded in the transfer area mt is larger than 0 in the execution history of the task 111, the determination unit 402 stores the disk as another storage area. It is determined that the execution result of task 111 has been written.

  The calculation unit 403 uses the amount u of the processing area mp used by the task 111 used when the task 111 is executed, the size of the processing area mp and the size of the transfer area mt acquired by the acquisition unit 401. Based on the above, the shortening time is calculated. Here, the usage amount u is the usage amount when the first task is executed if the first task is executed. If the x-th task between 1 and n−1 is executed, the usage amount u may be the usage amount when the x-th task is executed, or when the first task is executed. The average value from the usage amount to the usage amount when the x-th task is executed may be used. Further, the calculation unit 403 may calculate the shortened time when the determination unit 402 determines that the execution result of the task 111 has been written to another storage area.

  The calculation unit 403 calculates the execution time Tb based on the time T1 taken when the task 111 is executed, and also calculates the time T1, the usage u, the time rat, the size of the processing area mp, The execution time Ta may be calculated based on the size of the transfer area mt. A specific calculation example is as described in FIG. 1B.

  Further, the calculation unit 403 may calculate the time required for repeatedly executing the task 111 when the size of the transfer area mt is changed based on the usage amount as the third time. Specifically, the calculation unit 403 calculates the time T1, the usage amount of the processing area mp used by the task 111 used when the task 111 is executed, and the acquired processing area mp and transfer area mt. The third time may be calculated based on the size. The third time corresponds to T1 + T2 × (n−1), that is, Ta-rat in the graph 112 of FIG. 1B.

  The determination unit 404 determines whether to change the memory allocation based on the shortened time calculated by the calculation unit 403 and the time rat required for the memory allocation change.

  The determination unit 404 may determine whether to change the memory allocation based on a time rt obtained by subtracting the execution time Ta from the execution time Tb. Specifically, the determination unit 404 determines that the memory allocation is to be changed if the time rt is 0 or more, and otherwise determines that the memory allocation is not to be changed.

  Further, the determination unit 404 determines whether to change the memory allocation based on a comparison result between the time obtained by subtracting the third time from the execution time Tb and the time rat required for the memory allocation change. Also good. The time obtained by subtracting the third time from the execution time Tb corresponds to (T1−T2) × (n−1). For example, the determination unit 404 determines that the memory allocation is changed if the time obtained by subtracting the third time from the execution time Tb is equal to or longer than the time rat, and otherwise determines that the memory allocation is not changed. .

  The allocation changing unit 405 changes the memory allocation of the slave machine 202 when the determining unit 404 determines to change the memory allocation. Specifically, the allocation changing unit 405 stops the JVM operating on the slave machine 202, sets the memory of the JVM, and restarts the JVM.

  Next, each part when a plurality of machines perform data processing will be described. First, in the case of a plurality of machines, each of the slave machines 202-1 to 202-m executes the task 111 using the processing area mp of each slave machine 202. Here, the processing area mp has the same size among the slave machines 202. Then, after each slave machine 202-1 to m completes the task 111, each slave machine 202 uses the transfer area mt of each slave machine 202 to execute an execution result of the task 111 between the computers. Execute transfer processing to transfer. Here, the transfer area mt has the same size between the slave machines 202. Each slave machine 202 executes the task 111 again based on the transferred execution result.

  The acquisition unit 401 acquires the size of the processing area mp and the size of the transfer area mt from any one of the slave machines 202-1 to 202-m.

  The determination unit 402 determines whether any of the slave machines 202-1 to 20-m has written the execution result of the task 111 to another storage area different from the storage area included in the memory included in the execution of the task 111. Determine whether. Here, based on the execution history of the task 111 of each slave machine 202, the determination unit 402 writes the execution result to another storage area even if any one of the slave machines 202-1 to 202-m is written. It is determined that the execution result of the task 111 is written in the storage area.

  Assume that the determination unit 402 determines that any slave machine 202 has written the execution result of the task 111 to the other storage area. In this case, the calculation unit 403, based on the usage amount of the processing area mp used by the task 111 used when the task 111 is executed, and the acquired sizes of the processing area mp and the transfer area mt, Calculate the reduction time. The shortened time here is a time to shorten the time taken when the slave machines 202-1 to 20-m repeatedly execute the task 111 when the time rat for changing the memory allocation is not taken into consideration.

  The allocation changing unit 405 changes the memory allocation of all the slave machines 202-1 to m when the determination unit 404 determines to change the memory allocation.

(When a single machine processes data)
Next, a case where a single machine performs data processing as the first embodiment will be described with reference to FIGS.

  FIG. 5 is an explanatory diagram showing an example of assignment of tasks 111 when a single machine performs data processing. FIG. 5 shows an assignment example of the task 111 when the slave machine 202-1 performs data processing as a unit machine. Here, when an application execution instruction is issued by the client terminal 203, the master machine 201 generates a job 501 and performs data processing by assigning a task 111 to the slave machine 202-1 as one slave machine 202. Run. Here, one process of data processing is referred to as a “stage”. In the example of FIG. 5, the slave machine 202-1 reads data as stage 0, a task 111, which is a process for one data process, is assigned by the master machine 201 as stages 1 to n, and the assigned task 111 is assigned. Run.

  Here, in FIG. 5 and FIG. 6, for simplification of description, when the processing area mp and the transfer area mt are described, the processing area mp and the transfer area mt of the slave machine 202-1 are indicated. Show. Further, a symbol to which “_sx” is added after the processing area mp and the transfer area mt indicates the processing area mp and the transfer area mt in the stage x in a single machine. x is an integer of 0 or more and n or less.

  While executing the data processing, the master machine 201 monitors whether or not the data to be transferred has been written to the disk of the slave machine 202-1 without being transferred to the transfer area mt. If it is determined that the data to be transferred cannot fit in the transfer area mt, the master machine 201 sets an execution time Ta when performing a method of executing data processing by changing the memory allocation accompanying the restart of the JVM. calculate. Further, the master machine 201 calculates an execution time Tb when performing a method of executing data processing without changing the memory allocation. An example of calculating the execution times Ta and Tb will be described after the description of FIG. Then, the master machine 201 compares the execution times Ta and Tb, selects the shorter one, and continues data processing by the selected method.

  FIG. 5 shows an example in which the memory allocation is changed after stage 1 ends. Specifically, the master machine 201 determines that the data to be transferred cannot be placed in the transfer area mt_s1 when the stage 1 is completed. Then, the master machine 201 calculates execution times Ta and Tb, respectively. In FIG. 5, assuming that the execution time Ta is shorter than the execution time Tb, the master machine 201 restarts the JVM operating on the slave machine 202-1 and changes the memory allocation to the slave machine 202-1. Make it. As a result of the memory allocation change, the storage capacity of the transfer area mt_s2 increases more than the storage capacity of the transfer area mt_s1.

  After performing the memory allocation change, the slave machine 202-1 performs the transfer process, and then, as stage 2, performs one process using the process area mp_s2 for which the allocation change has been performed, and transfers the area mt_s2 Store data in. Subsequently, the slave machine 202-1 executes stages 3 to n.

(Example of calculating execution time on a single machine)
Next, an example of calculating the execution times Ta and Tb in a single machine will be described. First, an example of calculating the execution time Ta will be described. In order to calculate the execution time Ta, the master machine 201 calculates the time T2 required for one task 111 after stage 2 in the data processing according to the following expression (sa1).

  T2 = T1 × (M1 / M2) (sa1)

  Here, T1 indicates the time taken for stage 1 in data processing. In the example of a single machine, T1 and T2, which are times required for one stage, match the time required for the task 111 in the corresponding stage. M1 indicates the size of the JVM transfer area mt_s1 in stage 1. M2 indicates the size of the JVM transfer area mt_s2 after stage 2 when the transfer area mt is changed. For example, the master machine 201 may set a value obtained by adding a value obtained by subtracting the used amount used by the task 111 to the size of mt_s1 to M1 as M2. Alternatively, the master machine 201 may set M2 as a value obtained by adding a smaller value to M1 among the value obtained by subtracting the usage amount from the size of mt_s1 and the size of the data actually written to the disk 305. Note that the expression (sa1) is only an approximation and may be different from the actual time required for the stage 2.

  Then, the master machine 201 calculates an execution time Ta in a single machine according to the following formula (sa2).

  Ta = T1 + (Ts + Tm + Tr) + (T2 × (n−1)) (sa2)

  Here, Ts indicates the time required for stopping the JVM. Tm indicates the time required for the process of setting the JVM memory. Tr indicates the time required for restarting the JVM. Therefore, the second term in the expression (sa2) indicates the time required for the process of changing the memory allocation. Ts, Tm, and Tr are measured in advance and are given as fixed values.

  Next, an example of calculating the execution time Tb in a single machine is shown. The master machine 201 calculates the execution time Tb according to the equation (sb1).

  Tb = T1 × n (sb1)

  Next, a flowchart of data processing in a single machine will be described with reference to FIG.

  FIG. 6 is a flowchart illustrating an example of a data processing procedure in a single machine. Data processing in a single machine is processing in which data processing is performed by one slave machine 202.

  The master machine 201 starts executing the application (step S601). Then, the master machine 201 acquires the size of the processing area mp and the size of the transfer area mt from the slave machine 202 (step S602). Next, according to an instruction from the master machine 201, the slave machine 202 executes the first task 111 included in the data processing (step S603).

  Then, after receiving the execution result of the first task 111 from the slave machine 202, the master machine 201 determines whether or not the slave machine 202 has written data to the transfer area mt and the data has been written on the disk (step). S604). When the slave machine 202 writes the data to the transfer area mt without writing data to the disk (step S604: Yes), the master machine 201 calculates the execution time Ta for changing the memory allocation (step S605). . Further, the master machine 201 calculates an execution time Tb when the memory allocation is not changed (step S606).

  Then, the master machine 201 determines whether Ta is equal to or less than Tb (step S607). When Ta is equal to or less than Tb (step S607: Yes), after the master machine 201 instructs the slave machine 202, the slave machine 202 changes the memory allocation (step S608).

  When Ta is larger than Tb after the process of step S608 is completed (step S607: No), or when the slave machine 202 cannot write data in the transfer area mt and does not write it to the disk (step S604: No). In accordance with the instruction 201, the slave machine 202 executes the second and subsequent tasks 111 (step S609). Next, the master machine 201 determines whether or not the task 111 has been repeated n times (step S610). If the task 111 has not been repeated n times (step S610: No), the master machine 201 proceeds to the process of step S609. Note that the slave machine 202 executes the task 111 once by the process of step S603, and therefore executes the task 111 n-1 times by the process of step S609.

  On the other hand, when the task 111 is repeated n times (step S610: Yes), the master machine 201 ends the data processing in the single machine. Here, in each process of steps S605 to S607, Ta and Tb are calculated and then Ta and Tb are compared. However, the present invention is not limited to this. For example, the master machine 201 may execute the following inequality instead of the processes of steps S605 to S607, and if the inequality is satisfied, step S607: Yes, and if the inequality is not satisfied, step S607: No may be performed.

  Tb-Ta ≧ 0

  If the (sa2) and (sb1) equations are used for Tb and Ta, the above inequality can be modified as follows.

(T1 × n) − (T1 + (Ts + Tm + Tr) + (T2 × (n−1))) ≧ 0
⇔ (T1-T2) × (n−1) − (Ts + Tm + Tr) ≧ 0

  Here, (Ts + Tm + Tr) corresponds to the time rat required for the memory allocation change shown in FIG. 1B.

(When multiple machines work together to process data)
Next, as a second embodiment, a case where a plurality of machines cooperate to perform data processing will be described with reference to FIGS. Here, for simplification of explanation, it is assumed that one JVM is operated on each of the slave machines 202. In practice, multiple JVMs may be run on each of the slave machines 202. When a plurality of JVMs are operated, for example, the JVM may be operated on each core of the CPU of the slave machine 202.

  Here, in FIG. 7 to FIG. 9, for simplification of explanation, when the processing area mp and the transfer area mt are described, the respective processing areas mp and transfer areas mt of the slave machines 202-1 to 20-m. Indicates that Further, a symbol to which “_mx” is added after the processing area mp and the transfer area mt indicates the processing area mp and the transfer area mt in the stage x in a plurality of machines. x is an integer of 0 or more and n or less.

  FIG. 7 is an explanatory diagram showing an example of assignment of tasks 111 when data processing is performed on a plurality of machines. In the case of a plurality of machines, when an application execution instruction is issued from the client terminal 203, the master machine 201 generates a job 501 and assigns a task 111 to the slave machines 202-1 to m as a plurality of slave machines 202. Perform data processing. In the case of a plurality of machines, a stage from when the slave machines 202-1 to 20-m start executing one task 111 to when all of the slave machines 202-1 to 20-m complete one task 111 is defined as one stage.

  In the present embodiment, it is assumed that the in-memory data processing system 200 executes up to n stages. In the example of FIG. 7, the state of the stage 1 is shown. In the example of FIG. 7, the slave machines 202-1 and 202-m indicate that data could not be written to the transfer area mt_m <b> 1 during execution of the task 111, and the slave machines 202-1 and m were written to the respective disks.

  FIG. 8 is an explanatory diagram showing an execution example of the task 111 when data processing is performed on a plurality of machines. FIG. 8 shows an example of assignment of the task 111 at each stage when the slave machines 202-1 to 202-m cooperate as a plurality of machines to perform data processing.

  In the example of FIG. 8, stage 1 starts at time t1s, and stage 1 ends at time t1e when all of the slave machines 202-1 to 20-m have finished the task 111. In the example of FIG. 8, the slave machine 202-1 starts executing the task 111 at time t1s, and ends the task 111 before time t1e. Further, the slave machine 202-m starts executing the task 111 after the time t1s and ends the task 111 at the time t1e. As shown in FIG. 8, when all of the slave machines 202-1 to 20-m finish the task 111 by the time t1e, the slave machines 202-1 to 20-m execute a transfer process.

  In the example of FIG. 8, the slave machines 202-1 to m have completed the transfer process by time t2s. Then, stage 2 starts from time t2s, and stage 2 ends at time t2e when all of the slave machines 202-1 to 202-m have finished the task 111.

(Example of calculating execution time on multiple machines)
Next, an example of calculating the execution times Ta and Tb in a plurality of machines will be described. First, an example of calculating the execution time Ta will be described. In order to calculate the execution time Ta, the master machine 201 calculates the time T2 required for one task 111 after stage 2 in the data processing according to the following equation (ma1).

  T2 = T1 × (M1 / M2) (ma1)

  Here, T1 indicates the time taken for stage 1 in data processing. In the example of a plurality of machines, since parallel processing is performed in one stage, T1 and T2 which are times required for one stage are not simple calculations such as processing time of one task × number of machines. M1 represents the total size of the transfer area mt_m1 of the JVM operating on the slave machines 202-1 to 20-m in stage 1. Similarly, M2 represents the sum of the sizes of the transfer areas mt_m2 of the JVM operating on the slave machines 202-1 to m in stage 2 and subsequent stages when the transfer area mt is changed. For example, the master machine 201 calculates a value obtained by subtracting the usage amount used for the task 111 when the JVM running on the slave machines 202-1 to 202-m executes the task 111 from the sum of the sizes of mt_m1. Then, the master machine 201 may set a value obtained by adding the calculated value to M1 as M2. Note that the formula (ma1) is only an approximation and may be different from the actual time required for the stage 2.

  Then, the master machine 201 calculates an execution time Ta in a single machine according to the following formula (ma2).

  Ta = T1 + (Ts + Tm + Tr) + (T2 × (n−1)) (ma2)

  Here, Ts indicates the time required for stopping the JVM. Tm indicates the time required for the process of setting the JVM memory. Tr indicates the time required for restarting the JVM. Therefore, the second term in the equation (ma2) indicates the time required for the process of changing the memory allocation. Ts, Tm, and Tr are measured in advance and are given as fixed values.

  Next, an example of calculating the execution time Tb in a plurality of machines is shown. The master machine 201 calculates the execution time Tb according to the equation (mb1).

  Tb = T1 × n (mb1)

  Next, a flowchart of data processing in a plurality of machines will be described with reference to FIG.

  FIG. 9 is a flowchart illustrating an example of a data processing procedure in a plurality of machines. Data processing in a plurality of machines is processing in which data processing is performed by a plurality of slave machines 202.

  The master machine 201 starts executing the application (step S901). Then, the master machine 201 acquires the size of the processing area mp and the size of the transfer area mt from any of the slave machines 202-1 to 202-1m (step S902). Next, according to an instruction from the master machine 201, the slave machines 202-1 to 202-1 to m execute the first task 111 included in the data processing (step S903).

  Then, the master machine 201 receives the execution result of the first task 111 from all of the slave machines 202-1 to 202-m, and then has the slave machine 202 that has written data to the disk without being able to place data in the transfer area mt. It is determined whether or not (step S904). If there is a slave machine 202 that has written data to the disk without data being stored in the transfer area mt (step S904: Yes), the master machine 201 changes the memory allocation to all of the slave machines 202-1 to m. An execution time Ta for performing is calculated (step S905). Further, the master machine 201 calculates an execution time Tb when the memory allocation is not changed (step S906).

  Then, the master machine 201 determines whether Ta is equal to or less than Tb (step S907). When Ta is equal to or less than Tb (step S907: Yes), after the master machine 201 instructs the slave machines 202-1 to m, the slave machines 202-1 to m change the memory allocation (step S908).

  When Ta is larger than Tb after the process of step S908 is completed (step S907: No), or when there is no slave machine 202 that has not written data in the transfer area mt and has not been written to the disk (step S904: No). In accordance with the instruction 201, the slave machines 202-1 to 202-1m execute the second and subsequent tasks 111 (step S909). Next, the master machine 201 determines whether or not the task 111 has been repeated n times (step S910). If the task 111 has not been repeated n times (step S910: No), the master machine 201 proceeds to the process of step S909. Since the slave machines 202-1 to 20-m have executed the task 111 once by the process of step S903, the task 111 is executed n-1 times by the process of step S909.

  On the other hand, when the task 111 is repeated n times (step S910: Yes), the master machine 201 ends data processing in a plurality of machines. Here, similarly to the case where a single machine performs data processing, even when a plurality of machines cooperate to perform data processing, Ta and Tb are calculated after Ta and Tb are calculated in steps S905 to S907. However, the present invention is not limited to this. For example, the master machine 201 may execute the following inequality instead of the processes of steps S905 to S907, and if the inequality is satisfied, step S907: Yes, and if the inequality is not satisfied, step S907: No may be performed.

  Tb-Ta ≧ 0

  Since the transformation of the inequality is the same as that in the case where a single machine performs data processing, a description thereof will be omitted.

  As described above, when the master machine 201 executes the first task of in-memory data processing, the master machine 201 can reduce the time required for increasing the transfer area based on the usage amount u, and change the memory allocation. From this time rat, it is determined whether or not the memory allocation needs to be changed. The master machine 201 may determine whether or not the memory allocation needs to be changed based on a time rt obtained by subtracting the execution time Ta from the execution time Tb. Accordingly, the master machine 201 can appropriately determine whether or not the memory allocation needs to be changed in consideration of the time rat required to change the memory allocation, and as a result, the time required for data processing can be reduced. .

  Further, when the master machine 201 determines that the execution result of the task 111 is written to the disk during the execution of the task 111, the master machine 201 may calculate the time rt. As a result, the time rt is calculated only when the execution result of the task 111 is written to the disk during the execution of the task 111, so that the load on the master machine 201 can be suppressed.

  Further, the master machine 201 causes each of the plurality of slave machines 202 to execute the task 111, and after all of the plurality of slave machines 202 complete the task 111, causes each of the plurality of slave machines 202 to execute transfer processing, and again The task 111 may be executed. Thereby, the in-memory type data processing system 200 can execute in-memory type data processing in parallel using a plurality of slave machines 202 while being distributed.

  Further, the master machine 201 calculates the shortened time when a plurality of machines perform data processing, and when any of the plurality of slave machines 202 writes the execution result of the task 111 to the disk while the task 111 is being executed. Also good. As a result, even when a plurality of machines perform data processing, if there is no slave machine 202 that has written the execution result of the task 111 to the disk during the execution of the task 111, it is not necessary to calculate the shortened time. The load can be suppressed.

  Further, the master machine 201 can determine whether or not it is necessary to change memory allocation in consideration of the time required for memory allocation. Further, it is possible to suppress a decrease in processing speed that occurs when data overflow occurs in the transfer area.

  The control method described in this embodiment can be realized by executing a program prepared in advance on a computer such as a personal computer or a workstation. This control program is recorded on a computer-readable recording medium such as a hard disk, a flexible disk, a CD-ROM (Compact Disc-Read Only Memory), a DVD (Digital Versatile Disk), etc., and is read from the recording medium by the computer. Executed. The control program may be distributed via a network such as the Internet.

  The following additional notes are disclosed with respect to the embodiment described above.

(Supplementary note 1)
Obtaining the size of the processing area used by the task when repeatedly executing the task and the size of the transfer area for transferring the execution result of the task, allocated from the storage area of the memory,
Based on the usage amount of the processing area used by the task used when the task is executed, and the acquired size of the processing area and the size of the transfer area, based on the usage amount When the size of the transfer area is changed, the time required for repeatedly executing the task is calculated.
Determining whether to change the memory allocation based on the calculated shortened time and the time taken to change the memory allocation;
A control program characterized by causing a process to be executed.

(Supplementary note 2)
Based on the execution history of the task, it is determined whether or not the execution result of the task has been written to another storage area different from the storage area of the memory during the execution of the task.
The calculation process is as follows:
When it is determined that the execution result of the task has been written to the other storage area, the transfer is performed based on the usage amount based on the usage amount, the acquired size of the processing area, and the size of the transfer area. When the size of the area for use is changed, a reduction time for reducing the time taken to repeatedly execute the task is calculated.
The control program according to supplementary note 1, wherein:

(Supplementary note 3) The calculation process is as follows.
Based on the time taken when the task is executed, the first time required for repeatedly executing the task when the size is not changed is calculated, and the time taken when the task is executed, Based on the usage amount, the size of the transfer area is changed based on the usage amount, the time taken to change the allocation of the memory, and the acquired size of the processing area and the size of the transfer area. Calculating the second time it takes to repeatedly execute the task of the case,
The determination process is as follows:
Determining whether to change the memory allocation based on a time obtained by subtracting the second time from the calculated first time;
The control program according to supplementary note 2, characterized by:

(Supplementary Note 4) Each of the plurality of computers connected to the computer uses the processing area having the same size among the computers in the storage area of the memory of each of the computers. The task is executed between the computers using a transfer area having the same size among the computers in the storage area after the plurality of computers complete the task. Transfer the execution result of and execute the task again based on the transferred execution result.
The control program according to appendix 2 or 3, characterized by the above.

(Supplementary note 5)
Obtaining the size of the processing area and the size of the transfer area from any one of the plurality of computers,
The determination process is as follows.
Based on the task execution history of each of the computers, any one of the plurality of computers is different from the storage area of the memory of any of the computers during execution of the task. Determine whether the task execution result has been written to the area,
The calculation process is as follows:
When it is determined that one of the computers has written the execution result of the task to the other storage area, the usage amount of the processing area used by the task used when the task is executed, and the acquired Based on the size of the processing area and the size of the transfer area, calculate a reduction time for reducing the time required for the plurality of computers to repeatedly execute the task,
The determination process is as follows:
Determining whether to change the memory allocation based on the calculated shortened time and the time taken to change the memory allocation;
The control program according to supplementary note 4, characterized in that:

(Additional remark 6) The said process to calculate is
Based on the time taken when the task is executed, the first time required for repeatedly executing the task when the size is not changed is calculated, and the time taken when the task is executed, Repetitively executing the task when the size of the transfer area is changed based on the usage amount based on the usage amount, the acquired size of the processing area, and the size of the transfer area. Calculate the third time,
The determination process is as follows:
Determining whether to change the memory allocation based on a comparison result between a time obtained by subtracting the third time from the first time and a time required for the memory allocation change;
The control program according to any one of appendices 1 to 5, characterized in that:

(Appendix 7) Obtaining the size of the processing area allocated by the storage area included in the memory and used by the task when the task is repeatedly executed and the size of the transfer area for transferring the execution result of the task,
Based on the usage amount of the processing area used by the task used when the task is executed, and the acquired size of the processing area and the size of the transfer area, based on the usage amount When the size of the transfer area is changed, the time required for repeatedly executing the task is calculated.
Determining whether to change the memory allocation based on the calculated shortened time and the time taken to change the memory allocation;
A control device comprising a control unit.

(Appendix 8) The computer
Obtaining the size of the processing area used by the task when repeatedly executing the task and the size of the transfer area for transferring the execution result of the task, allocated from the storage area of the memory,
Based on the usage amount of the processing area used by the task used when the task is executed, and the acquired size of the processing area and the size of the transfer area, based on the usage amount When the size of the transfer area is changed, the time required for repeatedly executing the task is calculated.
Determining whether to change the memory allocation based on the calculated shortened time and the time taken to change the memory allocation;
A control method characterized by executing processing.

mp processing area mt transfer area u usage T1, T2 execution time 101 control device 102 execution device 111 task 201 master machine 202-1 to m slave machine 400 control unit 401 acquisition unit 402 determination unit 403 calculation unit 404 determination unit 405 Allocation change part

Claims (7)

On the computer,
Obtaining the size of the processing area used by the task when repeatedly executing the task and the size of the transfer area for transferring the execution result of the task, allocated from the storage area of the memory,
Based on the usage amount of the processing area used by the task used when the task is executed, and the acquired size of the processing area and the size of the transfer area, based on the usage amount When the size of the transfer area is changed, the time required for repeatedly executing the task is calculated.
Determining whether to change the memory allocation based on the calculated shortened time and the time taken to change the memory allocation;
A control program characterized by causing a process to be executed. In the computer,
Based on the execution history of the task, it is determined whether or not the execution result of the task has been written to another storage area different from the storage area of the memory during the execution of the task.
The calculation process is as follows:
When it is determined that the execution result of the task has been written to the other storage area, the transfer is performed based on the usage amount based on the usage amount, the acquired size of the processing area, and the size of the transfer area. When the size of the area for use is changed, a reduction time for reducing the time taken to repeatedly execute the task is calculated.
The control program according to claim 1, wherein: The calculation process is as follows:
Based on the time taken when the task is executed, the first time required for repeatedly executing the task when the size is not changed is calculated, and the time taken when the task is executed, Based on the usage amount, the size of the transfer area is changed based on the usage amount, the time taken to change the allocation of the memory, and the acquired size of the processing area and the size of the transfer area. Calculating the second time it takes to repeatedly execute the task of the case,
The determination process is as follows:
Determining whether to change the memory allocation based on a time obtained by subtracting the second time from the calculated first time;
The control program according to claim 2. Each of a plurality of computers connected to the computer executes the task using a processing area having the same size among the respective computers in a storage area of a memory of each of the computers. The execution result of the task between the computers using the transfer area having the same size among the computers in the storage area after the plurality of computers complete the task. Transfer, and execute the task again based on the transferred execution result,
The control program according to claim 2 or 3, wherein The process to obtain is
Obtaining the size of the processing area and the size of the transfer area from any one of the plurality of computers,
The determination process is as follows.
Based on the task execution history of each of the computers, any one of the plurality of computers is different from the storage area of the memory of any of the computers during execution of the task. Determine whether the task execution result has been written to the area,
The calculation process is as follows:
When it is determined that one of the computers has written the execution result of the task to the other storage area, the usage amount of the processing area used by the task used when the task is executed, and the acquired Based on the size of the processing area and the size of the transfer area, calculate a reduction time for reducing the time required for the plurality of computers to repeatedly execute the task,
The determination process is as follows:
Determining whether to change the memory allocation based on the calculated shortened time and the time taken to change the memory allocation;
The control program according to claim 4. Obtaining the size of the processing area used by the task when repeatedly executing the task and the size of the transfer area for transferring the execution result of the task, allocated from the storage area of the memory,
Based on the usage amount of the processing area used by the task used when the task is executed, and the acquired size of the processing area and the size of the transfer area, based on the usage amount When the size of the transfer area is changed, the time required for repeatedly executing the task is calculated.
Determining whether to change the memory allocation based on the calculated shortened time and the time taken to change the memory allocation;
A control device comprising a control unit. Computer
Obtaining the size of the processing area used by the task when repeatedly executing the task and the size of the transfer area for transferring the execution result of the task, allocated from the storage area of the memory,
Based on the usage amount of the processing area used by the task used when the task is executed, and the acquired size of the processing area and the size of the transfer area, based on the usage amount When the size of the transfer area is changed, the time required for repeatedly executing the task is calculated.
Determining whether to change the memory allocation based on the calculated shortened time and the time taken to change the memory allocation;
A control method characterized by executing processing.

Images