JP2013088863A - Parallel distributed processing method and parallel distributed processing system - Google Patents

Abstract

PROBLEM TO BE SOLVED: To enable a master device to ascertain processing efficiency of each slave device without introducing an external monitoring computer or a monitoring program, in a master/slave type system.SOLUTION: A parallel distributed processing method to be executed by a parallel distributed processing system 4, includes the steps in which: slaves 1 execute tasks and also measure execution of the tasks by using a measurement unit predetermined in the parallel distributed processing system 4, store measurement results in a storage unit 11, and transmit the stored measurement results to a master 2; the master 2 stores the received measurement results in association with information for identifying respective transmission source slaves 1 in a collection unit 21 as a collection result, generates instruction information for the slaves 1 on the basis of the stored collection result, and transmits the instruction information to the slaves 1; and the respective slaves 1 control operation of its own slave 1 on the basis of the received instruction information.

Description

  The present invention relates to a parallel distributed processing method and a parallel distributed processing system for performing parallel distributed processing of tasks.

  Today, a very large number of small individual tasks are distributed and processed by a plurality of computers. In such a method, a plurality of computers that process individual tasks are often called “slaves”, and a computer that assigns tasks to each slave and aggregates the processing results of each slave is often called a “master”.

  As such a master / slave type system, MapReduce technology proposed by Google (registered trademark), Hadoop, which is an open source implementation thereof, and Dryad (registered trademark) proposed by Microsoft (registered trademark) are included. Furthermore, OpenMPI etc. are mentioned.

  In a master / slave system, it is essential to monitor the operating status of each computer. One of the purposes of monitoring is to find a computer that does not exhibit the planned processing efficiency for some reason such as failure or overload. If the cause of the reduction in processing efficiency is that the specific computer is a bottleneck, it will be resolved, and if it is a failure, measures such as repair or replacement will be taken.

  As techniques for monitoring the operation status of a computer, for example, techniques described in Patent Document 1 and Non-Patent Document 1 below are known. Patent Document 1 discloses that a dedicated monitoring computer for monitoring a monitoring target computer is prepared, and the monitoring target computer is monitored from outside the monitoring target computer. In Non-Patent Document 1, a dedicated monitoring software is installed and operated on a monitoring target computer, and a monitoring target process (a state in which the monitoring target software is operating) is executed on the monitoring target computer. The point of monitoring the process to be monitored from the outside is disclosed.

  On the other hand, in a server / client type system, as a technique for performing monitoring without the aid of an external monitoring computer or monitoring software, for example, techniques described in Patent Document 2 and Patent Document 3 below are known. Patent Document 2 discloses that the server changes the behavior of the server by monitoring the communication speed with the client and reducing the resolution of image data provided to the client, for example, according to the communication speed. . Patent Document 3 discloses that a server polls each client, monitors the activity (dead or alive) status of each client, and automatically excludes a client determined to be inactive from the system.

JP 2001-325161 A JP 2002-32295 A JP-A-7-271701

The Ganglia Distributed Monitoring System: Design Implementation, and Experience, Parallel Computing Volume30, Issue 7

  In a master / slave system, the master is required to devise assignment (scheduling) of tasks to slaves so that processing of all tasks can be completed earlier. In order to realize such scheduling, the master monitors the processing efficiency of tasks of each slave, and assigns tasks to higher-performance slaves preferentially or is not used in the computer pool in the system. When there is a computer, it is necessary to devise such as stopping a slave with low processing efficiency and making another computer in the computer pool a new slave.

  Here, with the techniques described in Patent Document 1 and Non-Patent Document 1, general-purpose and objectively monitorable indicators such as CPU usage rate and memory usage rate are monitored from outside the computer or process to be monitored. become. With the techniques described in Patent Literature 1 and Non-Patent Literature 1, it can be detected that a high load is applied to a computer or process having a high CPU usage rate or a high memory usage rate. Therefore, the task processing efficiency is low. It cannot be judged. This is because the CPU usage rate and the memory usage rate may be high because the task is processed with high efficiency. In other words, with the techniques described in Patent Document 1 and Non-Patent Document 1, it is only possible to indirectly estimate the task processing efficiency only from an index that can be monitored from the outside.

  In addition, the technique described in Patent Document 1 requires costs and labor for preparing and operating a monitoring computer, and the technique described in Non-Patent Document 1 may generate an extra load due to operating the monitoring software. It is a problem.

  On the other hand, the techniques described in Patent Document 2 and Patent Document 3 do not require an external monitoring computer or monitoring software. However, the technique described in Patent Document 2 only monitors the communication speed between the server and the client, and the task processing efficiency cannot be measured only by the communication speed. The technique described in Patent Document 3 only monitors the activity status of the client, and cannot measure the task processing efficiency only by the activity status of the client.

  The present invention has been made to solve the above-described problems. In a master / slave type system, the master grasps the processing efficiency of the slave without introducing an external monitoring computer or monitoring program. An object of the present invention is to provide a parallel distributed processing method and a parallel distributed processing system.

  A parallel and distributed processing method according to one aspect of the present invention includes a parallel and distributed processing system that includes one or more slaves and a master connected to the slaves via a network, and performs parallel and distributed processing of tasks on the slaves. The parallel distributed processing method executed by the slave, the slave executes the task assigned to the slave, and measures the execution of the task using a predetermined measurement unit in the parallel distributed processing system, A processing step for storing the measurement result in the storage unit of the slave, a first reporting step in which the slave transmits the measurement result stored in the storage unit of the slave to the master, and the master receiving the measurement result from the slave The measurement result is associated with the information for identifying the slave at the transmission source, A communication step for storing in the master, a master generating instruction information for the slave based on the counting result stored in the counting means, and a determination step for transmitting to the slave, and a step in which the slave receives the instruction information from the master. And a control step in which the slave controls the operation of the slave based on the instruction information received in the second reporting step.

  In this way, the execution of a task is measured by a slave using a predetermined measurement unit in the parallel distributed processing system, and the measurement result is transmitted to the master. It is possible to grasp the processing efficiency of a slave without introducing a program. Then, the master issues an instruction to the slave based on the processing efficiency of the slave, and the slave dynamically controls the operation of the entire system based on the processing efficiency of the slave by controlling the operation based on the instruction. can do.

  In addition, the determination step includes a summation result per reception from the transmission source slave in the communication step among the aggregation results stored in the aggregation means, and one from a slave other than the transmission source slave in the communication step. It is also possible to compare the aggregated results per reception, generate instruction information for the transmission source slave based on the comparison result, and transmit the instruction information to the transmission source slave. In such a case, the master can issue an instruction to the target slave more accurately based on the comparison result of the aggregated result per reception of the target slave and the aggregated result per reception of other slaves. The processing efficiency of the slave can be grasped, and the operation of the entire system can be dynamically controlled more efficiently.

  In the communication step, the master further associates the elapsed time from the reception of the previous measurement result from the transmission source slave, and stores it in the totaling means of the master as a totaling result. Of the aggregated results stored in, the value per unit time of the aggregated results of one reception from the transmission source slave in the communication step, and the unit of the aggregated results of one reception from slaves other than the transmission source slave in the communication step It may be compared with a value per time, instruction information for the transmission source slave may be generated based on the comparison result, and transmitted to the transmission source slave. In such a case, the master instructs the target slave based on the comparison result between the value per unit time of the aggregate result of one reception of the target slave and the value per unit time of the aggregate result of one reception of the other slave. Therefore, the processing efficiency of the target slave can be grasped more accurately and the operation of the entire system can be dynamically controlled more efficiently.

  In addition, in the determination step, the master calculates an average value of all the aggregation results of the transmission source slave among the aggregation results stored in the aggregation means, and an average value of all the aggregation results of the slaves other than the transmission source slave. May be generated, instruction information for the transmission source slave may be generated based on the comparison result, and transmitted to the transmission source slave. In such a case, the master can issue an instruction to the target slave based on the comparison result between the average value of all the aggregation results of the target slave and the average value of all the aggregation results of the other slaves. In addition, the processing efficiency of the target slave can be grasped, and the operation of the entire system can be dynamically controlled more efficiently.

  In the communication step, the master further associates the elapsed time from the reception of the previous measurement result from the transmission source slave, and stores it in the totaling means of the master as a totaling result. Of the aggregation results stored in, the average value per unit time of all aggregation results of the transmission source slave and the average value of unit aggregation values of all the aggregation results of slaves other than the transmission source slave May be generated, instruction information for the transmission source slave may be generated based on the comparison result, and transmitted to the transmission source slave. In such a case, the master determines the target based on the comparison result between the average value of all the aggregate results of the target slaves per unit time and the average value of all the aggregate results of other slaves per unit time. Since an instruction can be issued to the slave, the processing efficiency of the target slave can be grasped more accurately and the operation of the entire system can be dynamically controlled more efficiently.

  In addition, in the determination step, the master performs statistical processing on the aggregation results for all slaves of the aggregation results received from the slave in the communication step among the aggregation results stored in the aggregation means. The instruction information for the transmission source slave may be generated based on the result and transmitted to the transmission source slave. In such a case, the master can issue an instruction to the slave based on the statistical processing results for the totaling results for all slaves of the totaling results per reception from the slave, so the slave processing can be performed more accurately. The efficiency can be grasped, and the operation of the entire system can be dynamically controlled more efficiently.

  In the communication step, the master further associates the elapsed time from the reception of the previous measurement result from the transmission source slave, and stores it in the totaling means of the master as a totaling result. The statistical processing is performed on the values related to all slaves of the aggregated results of one reception from the slave in the communication step among the aggregated results stored in the communication step, and based on the statistical processing results, Instruction information for the slave may be generated and transmitted to the slave at the transmission source. In such a case, the master can issue an instruction to the slave based on the statistical processing results for the values related to all slaves of the values per unit time of the total result of one reception from the slave, and thus more accurate. In addition, the processing efficiency of the slave can be grasped, and the operation of the entire system can be dynamically controlled more efficiently.

  In addition, the determination step includes, among the aggregation results stored in the aggregation means, when the average value for all slaves of the aggregation results per reception from the slave in the communication step exceeds a predetermined threshold, The master may generate instruction information for the slave to reduce the load on the master and transmit the instruction information to the slave. In such a case, the master can detect that, for example, the number of slaves exceeding the master's capacity is connected, and the number of slaves connected according to the master's capacity or the conditions set according to the capacity. Can be controlled dynamically.

  By the way, the invention according to the parallel distributed processing method can be regarded as the invention of the system, and has the same operation and effect. The invention of the system can be described as follows.

  A parallel distributed processing system according to an aspect of the present invention is a parallel distributed processing system that includes one or more slaves and a master connected to the slaves via a network, and performs parallel distributed processing of tasks on the slaves. The slave executes the task assigned to the slave, measures the execution of the task using a predetermined measurement unit in the parallel distributed processing system, and stores the measurement result in the storage unit of the slave. Based on the processing means for storing, the reporting means for transmitting the measurement result stored in the storage means of the slave to the master, and receiving the instruction information from the master as a response to the transmission, and the instruction information received by the reporting means A control means for controlling the operation of the slave, and the master receives the measurement result from the slave and performs the measurement. The communication means for storing the result and the information for identifying the slave of the transmission source and storing it in the totaling means of the master as the totaling result, and generating the instruction information for the slave based on the totaling result stored in the totaling means, Determining means for transmitting to the slave.

  According to the present invention, in the master / slave type system, the master can grasp the processing efficiency of the slave without introducing an external monitoring computer or a monitoring program.

It is a conceptual diagram of the parallel distributed processing system which concerns on one Embodiment of this invention. It is a block diagram which shows the structure of the slave which concerns on one Embodiment of this invention. It is a block diagram which shows the structure of the master which concerns on one Embodiment of this invention. It is a figure which shows an example of the hardware constitutions of the slave and master which concern on one Embodiment of this invention. It is a flowchart which shows the processing operation of the parallel distributed processing method in the slave which concerns on one Embodiment of this invention. It is a flowchart which shows the processing operation (the 1) of the parallel distributed processing method in the master which concerns on one Embodiment of this invention. It is a flowchart which shows the processing operation (the 2) of the parallel distributed processing method in the master which concerns on one Embodiment of this invention. It is a flowchart which shows the processing operation (the 3) of the parallel distributed processing method in the master which concerns on one Embodiment of this invention. It is a figure which shows an example (the 1) of the table data of the total part which concerns on one Embodiment of this invention. It is a figure which shows an example (the 2) of the table data of the total part which concerns on one Embodiment of this invention. It is a figure which shows an example (the 3) of the table data of the total part which concerns on one Embodiment of this invention. It is a figure which shows an example (the 4) of the table data of the total part which concerns on one Embodiment of this invention. It is a figure which shows an example (the 5) of table data of the total part which concerns on one Embodiment of this invention.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings. In the description of the drawings, the same or equivalent elements are denoted by the same reference numerals, and redundant description is omitted.

(Overview of parallel distributed processing system)
First, an overall view of a parallel distributed processing system according to an embodiment of the present invention will be described with reference to FIG. FIG. 1 is a diagram showing an overview of a parallel distributed processing system 4. The parallel distributed processing system 4 includes one or more slaves 1 and a master 2. In the parallel distributed processing system 4, each slave 1 and master 2 can communicate with each other via a network 3.

  Each of the slave 1 and the master 2 is a server device, but is not limited thereto, and may be a PC (Personal Computer), a notebook PC, a mobile communication terminal such as a mobile phone, or the like. The network 3 is configured by, for example, the Internet network or a mobile communication network, but is not limited thereto.

  The parallel distributed processing system 4 is a so-called master / slave type cluster (a set of computers). In the parallel distributed processing system 4, there are a large number of tasks (processing contents), each of which is a small uniform task, Parallel distributed processing is executed by performing distributed processing among one or more slaves 1. In the parallel distributed processing system 4, the master 2 manages all the slaves 1.

(Configuration of slave 1)
FIG. 2 is a diagram illustrating a configuration of the slave 1 according to the embodiment of the present invention. The slave 1 shown in FIG. 2 includes a processing unit 10 (processing unit), a storage unit 11 (storage unit), a reporting unit 12 (reporting unit), and a control unit 13 (control unit).

  FIG. 4 shows an example of the hardware configuration of the slave 1. As a hardware configuration, the slave 1 includes a CPU 50, a RAM 51, a ROM 52, an input unit 53 including a keyboard and a numeric keypad, a communication unit 54 that communicates with the outside, an auxiliary storage device 55, and a display. 56. The function of each functional block of the slave 1 described above is realized by causing the RAM 51 or the like to read a program or data and executing the program under the control of the CPU 50. The master 2 described later has the same hardware configuration as that shown in FIG.

  Hereinafter, each functional block of the slave 1 shown in FIG. 2 will be described. The processing unit 10 executes the task assigned to the slave 1, measures the execution of the task using a predetermined measurement unit in the parallel distributed processing system 4, and stores the measurement result in the storage unit 11. . The processing unit 10 repeats this operation until there is no task assigned to the slave 1.

  The measurement unit needs to be determined in advance between all the slaves 1 and the master 2 in the parallel distributed processing system 4. The measurement unit may be anything as long as it can be measured (or counted) equally in all the slaves 1, and examples thereof include “number of processed tasks” and “seconds”. In this embodiment, the measurement unit is described as “the number of processed tasks”. It is intended for situations where the total number of tasks is very large and each task is small and uniform.

  When storing the measurement result in the storage unit 11, the processing unit 10 stores the measurement result at regular intervals. Here, the fixed interval may be a fixed time or every time a certain number of tasks are processed, but needs to be determined in advance between all the slaves 1 and the master 2 in the parallel distributed processing system 4. is there. When the processing unit 10 stores the processing result in the storage unit 11, the processing unit 10 resets the measurement result held by the processing unit 10 to zero.

  Further, the processing unit 10 may determine whether to store the measurement result when storing the measurement result in the storage unit 11. Examples of the determination criterion include “store if one second or more has elapsed since the previous storage”, “store every time one task is processed”, and the like. The determination criteria need to be unified for all slaves 1 in the parallel distributed processing system 4, but may be fixed programmatically or may be set using a setting file or the like. In the present embodiment, the processing unit 10 determines whether to store the measurement result in the storage unit 11 using the determination criterion “store every time one task is processed”.

  The measurement result is stored in the storage unit 11 by the processing unit 10. Each time the measurement result is stored in the storage unit 11, the stored measurement result is added to the previously stored measurement result without being overwritten.

  The report unit 12 extracts the measurement result stored in the storage unit 11 and transmits the extracted measurement result to the master 2. Then, the reporting unit 12 receives the instruction information from the master 2 as a response to the transmission, and outputs the received instruction information to the reporting unit 12.

  When the report unit 12 takes out the measurement result stored in the storage unit 11, the report unit 12 resets the measurement result in the storage unit 11 to zero. The reporting unit 12 determines whether or not the taken measurement result should be reported to the master 2 when transmitting the taken measurement result to the master 2, that is, whether or not the measurement result is “1” or more. Also good. In this case, when the measurement result is zero, the reporting unit 12 waits because there is no measurement result to be reported, and when there is a measurement result to be reported, the reporting unit 12 actually transmits the extracted measurement result to the master 2. . Further, when transmitting the extracted measurement result to the master 2, the report unit 12 may transmit it together with the sequential report of the processing result from the slave to the master in the conventional parallel distributed processing technique.

  The control unit 13 acquires the instruction information output from the report unit 12, and controls the operation of the slave 1 based on the instruction information. Specific examples of the instruction information include “do nothing special”, “stop the operations of the main thread in which processing by the processing unit 10 is executed, and another thread in which processing by the reporting unit 12 is executed” “ The processing unit 10 changes the determination criterion used when storing the measurement result in the storage unit 11 ”.

(Configuration of Master 2)
FIG. 3 is a diagram showing a configuration of the master 2 according to an embodiment of the present invention. 3 includes a communication unit 20 (communication unit), a totaling unit 21 (totaling unit), a determination unit 22 (determination unit), and a display unit 23.

  Hereinafter, each functional block of the master 2 shown in FIG. 3 will be described. The communication unit 20 receives the measurement result from the slave 1, associates the measurement result with information for identifying the slave 1 as the transmission source, and stores the result in the aggregation unit 21 as the aggregation result. A specific example of information for identifying the slave 1 is the IP address of the slave 1. The communication unit 20 further associates the elapsed time from the reception of the previous measurement result from the transmission source slave 1 or the reception time of the measurement result from the transmission source slave 1 and stores the result in the aggregation unit 21 as the aggregation result. Also good.

  The totaling result is stored in the totaling unit 21 by the communication unit 20. For example, the counting unit 21 may store and count the counting result for each piece of information for identifying the slave 1 included in the counting result.

  The determination unit 22 determines the status of the slave 1, the master 2, and the parallel distributed processing system 4, the prediction and instruction associated with the status based on the total result stored in the total unit 21, and determines the slave based on the determination result. 1 is generated and transmitted to the slave 1. The contents of the determination include detection of the slave 1 with low processing efficiency, the degree of load of the master 2 itself, or some prediction from the processing efficiency of the entire parallel distributed processing system 4. Hereinafter, various processing patterns of the determination unit 22 are listed. Specific examples of these processing patterns will be described in the processing of various embodiments of the master 2 described later.

  The determination unit 22 is configured such that the master 2 is not included in the totaling results stored in the totaling unit 21, except for the totaling result per reception from the transmission source slave 1 in the communication unit 20 and the transmission source slave 1 in the communication unit 20. May be compared with the total result per reception from the slave 1, the instruction information for the transmission source slave 1 may be generated based on the comparison result, and may be transmitted to the transmission source slave 1. In addition, the determination unit 22 determines the value per unit time of the total result of one reception from the transmission source slave 1 in the communication unit 20 among the total results stored in the total unit 21 by the master 2, and the communication unit 20. Are compared with the values per unit time of the total result of one reception from the slaves 1 other than the transmission source slave 1, and based on the comparison result, instruction information for the transmission source slave 1 is generated, and the transmission source slave 1 May be sent to.

  In addition, the determination unit 22 determines that the master 2 calculates the average value of all the aggregation results of the transmission source slave 1 among the aggregation results stored in the aggregation unit 21 and all the slaves 1 other than the transmission source slave 1. It may be compared with the average value of the tabulation results, and the instruction information for the transmission source slave 1 may be generated based on the comparison result and transmitted to the transmission source slave 1. In addition, the determination unit 22 determines that the master 2 is an average value of all the aggregation results of the transmission source slave 1 among the aggregation results stored in the aggregation unit 21 and other than the transmission source slave 1. May be compared with the average value of the values per unit time of all the summation results of the slave 1, and the instruction information for the slave 1 of the transmission source may be generated based on the comparison result and transmitted to the slave 1 of the transmission source .

  In addition, the determination unit 22 performs statistics on the aggregation results for all slaves 1 of the aggregation results per reception from the slave 1 in the communication unit 20 among the aggregation results stored in the aggregation unit 21. Processing may be performed to generate instruction information for the transmission source slave 1 based on the statistical processing result and transmit the instruction information to the transmission source slave 1. In addition, the determination unit 22 is a value related to all the slaves 1 among the totaling results stored in the totaling unit 21 by the master 2 as a result of the totaling result of one reception from the slave 1 in the communication unit 20. May be subjected to statistical processing, instruction information for the transmission source slave 1 may be generated based on the statistical processing result, and transmitted to the transmission source slave 1.

  In addition, the determination unit 22 determines that the average value for all slaves 1 of the total results per reception from the slave 1 in the communication unit 20 among the total results stored in the total unit 21 is a predetermined threshold value. When it exceeds, the master 2 may generate instruction information for the slave 1 to reduce the load on the master and transmit the instruction information to the slave 1.

  The display unit 23 displays information related to the parallel distributed processing system 4 to the user of the parallel distributed processing system 4.

(Description of slave 1 processing)
Next, the processing procedure of the parallel distributed processing method in the slave 1 will be described with reference to FIG. First, when a task is assigned, the slave 1 starts a main thread and another thread. On the main thread, processing by the processing unit 10, specifically, steps 101 to 104 are executed. On the other hand, the processing by the reporting unit 12, specifically, steps 201 to 205 are executed on another thread.

  First, processing on the main thread of the slave 1 will be described. First, the processing unit 10 determines whether there are still tasks assigned to the slave 1 (step 101). If no task remains in step 101, the operation of the slave 1 is terminated, and both the main thread and another thread are stopped. On the other hand, if a task remains in step 101, the processing unit 10 executes one task and measures the execution of the task using a predetermined measurement unit beside the task (step 102). Next, the processing unit 10 determines whether or not to store the measurement result in the storage unit 11 (step 103). When it is determined in step 103 that the data is stored in the storage unit 11, the processing unit 10 stores the measured measurement result in the storage unit 11, and resets the measurement result of the processing unit 10 to zero (step 104). If the previous measurement result remains in the storage unit 11, the current measurement result is added without being overwritten. If it is determined in step 103 that the data is not stored in the storage unit 11, the process proceeds to step 101. The main thread of slave 1 repeats the above steps 101 to 104 until there is no task assigned to slave 1.

  Next, processing on another thread of the slave 1 will be described. First, the reporting unit 12 retrieves the measurement result from the storage unit 11 and resets the measurement result in the storage unit 11 to zero (step 201). Next, the reporting unit 12 determines whether or not the taken measurement result should be reported to the master 2, that is, whether or not the measurement result is “1” or more (step 202). If the measurement result is zero in step 202, the determination result is “NO” because there is no measurement result to be reported, and the process proceeds to step 201. If there is a measurement result to be reported in step 202, the reporting unit 12 transmits the measurement result to the master 2 (step 203). Next, the reporting unit 12 receives an instruction from the master 2 as a response to the measurement result transmitted to the master 2 (step 204). Next, the reporting unit 12 hands over the instruction received from the master 2 to the control unit 13 (step 205). Upon receiving the instruction, the control unit 13 controls the operation of the slave 1 according to the content of the instruction. Another thread of the slave 1 repeats the above steps 201 to 205 until the main thread is finished.

  Here, the main thread and another thread in the slave 1 operate asynchronously with each other. The main thread repeats task processing, measurement, and storage of measurement results at best effort, regardless of the operation of another thread. The other thread also performs a best-effort sequence of operations such as taking out the measurement result, sending the measurement result to the master, receiving the instruction from the master, and passing the instruction to the control unit 13 regardless of the operation of the main thread. Repeat until the main thread ends. By the way, although there are one or more slaves 1 in the entire parallel distributed processing system 4, there is no causal relationship between the number of slaves 1 and the processing efficiency of the main thread of one slave 1, which is irrelevant. It is. On the other hand, there is a causal relationship between the number of slaves 1 and the efficiency of another thread of a certain slave 1. This is because, as the number of slaves 1 increases, the number of slaves 1 that transmit masters to which the master 2 should receive and respond increases, so the load on the master 2 increases. This is because the time until the response is returned (hereinafter referred to as response time) becomes longer. The point of this embodiment is that the ratio of the efficiency of the main thread and another thread in the slave 1 is used.

(Description of the first example of processing of the master 2)
Next, a first example of the parallel distributed processing method performed by the master 2 will be described with reference to FIG. First, the communication unit 20 receives a measurement result from each slave 1 (step 301). Next, the communication unit 20 determines whether to use time information (step 302). In the following description, instead of using only the latest measurement result for each transmission source slave 1 of the measurement result, the history of the measurement result for each transmission source slave 1 is stored in the counting unit 21 and the average value thereof is used. I will do it.

  If it is determined in step 302 that the time information is not used, the communication unit 20 associates the received measurement result with the transmission source ID for identifying the slave 1 that is the transmission source of the measurement result, and calculates the totaling unit as a totaling result. 21 (step 303). In the case where the time information is not used, the processing capability of all slaves 1 is uniform, for example, the hardware configuration of all slaves 1 is uniform, and the connection path between each slave 1 and master 2 is also uniform. This is an effective case in a situation where the difference in communication delay can be ignored.

  FIG. 9 is a diagram illustrating a table example of the totaling result stored in the totaling unit 21 immediately after step 303. In FIG. 9 (and FIGS. 10 to 13 described later), the unit of measurement is “number of processed tasks”. When the average value of the measurement results is calculated for each transmission source slave 1 from the table example shown in FIG. 9, slave 1 whose transmission source ID is AAAAA (hereinafter referred to as AAAAA. The same applies to slaves 1 of other transmission IDs), The average value of the measurement results of CCCCC and EEEEEE is “10”, the average value of BBBBB is “5”, and the average value of DDDDD is “15”.

  Subsequent to step 303, the determination unit 22 determines whether or not the aggregation result is abnormal with respect to the average value of the aggregation results (step 304). The determination unit 22 may make the following determination regarding BBBBB, for example. In other words, if the connection paths between all the slaves 1 and the master 2 are uniform, the response time seen from each slave 1 should be almost uniform, but the number of tasks that can be processed during that time is about half of that in the BBBBB. It is. In this embodiment, since it is assumed that each task is uniform, it can be determined that the processing efficiency of BBBBB is reduced due to some cause. This is because it is strange that the hardware configuration is extremely low due to the assumption that the hardware configuration is uniform. On the other hand, for example, regarding the DDDDD, the determination unit 22 may make the following determination. In other words, it is difficult to consider that only DDDDDD can be skipped and the processing efficiency is increased on the assumption that the tasks are uniform and the hardware configuration is uniform. That is, it can be determined that the response time as viewed from DDDDD is long due to some cause, for example, malfunction of the network 3 connection. Note that the policy for determining “how many (or less) the counting results are considered abnormal” may be fixed programmatically or may be set in a policy file or the like.

  Based on the above determination by the determination unit 22, if the measurement result received by the determination unit 22 is from BBBBB or DDDDD, it is determined as abnormal in step 304 (determination “YES” in step 304 ]), For example, corresponding instruction information such as immediately stopping the operation of the slave 1 is generated and returned to the slave 1 (step 308). If the measurement result received by the determination unit 22 is not from BBBBB or DDDDD (determination “NO” in step 304), a normal response such as instruction information “do nothing” is generated, for example. It returns to the slave 1 (step 307).

  When the determination unit 22 determines that some abnormality has occurred in the specific slave 1, the master 2 stores the information, for example, devise assignment (scheduling) of tasks to the slave 1, etc. Information is presented to the user of the parallel distributed processing system 4, for example, for the control of the entire parallel distributed processing system 4 thereafter, for example, reporting to the user the slave 1 in which an abnormality has occurred by the display unit 23 Can do.

  Next, a case where it is determined in step 302 that time information is used will be described. The case where it is determined that the time information is used is effective in a situation where it is impossible to assume that the processing capability of all the slaves 1 is uniform and the connection path between each slave 1 and the master 2 is uniform. It is.

  When it is determined in step 302 that the time information is used, the communication unit 20 receives the measurement result, the transmission source ID for identifying the slave 1 that is the transmission source of the measurement result, and the time when the measurement result is received. Are stored in the totaling unit 21 as a totaling result (step 305).

  FIG. 10 is a diagram illustrating a table example of a totaling result stored in the totaling unit 21 immediately after step 305. In the example of the table shown in FIG. 10, at first glance, the average values of the measurement results per reception by the communication unit 20 are all “10”, and it seems that there is no problem. However, when looking at the average value of measurement results per time (in the following embodiments, “seconds”), it is “2” for AAAAA, BBBBB, DDDDD, and IEEE, and “1” for CCCCC. I understand.

  Subsequent to step 205, the determination unit 22 determines whether or not the aggregation result is abnormal with respect to the average value of the aggregation results (step 306). The determination unit 22 receives the measurement result from the CCCCC twice as long as the other slave 1 because the network connection of the CCCCC is originally inferior to the other, or the network of the CCCCC It cannot be judged whether something is wrong with the connection. However, the determination unit 22 determines that the task processing efficiency per unit time is inferior to the other for CCCCC. Note that the determination unit 22 cannot determine whether the CCCCC processing efficiency is originally inferior to the other, or because some failure has occurred in the CCCCC, but anyway, the CCCCC processing efficiency is higher than the others. Judge that it is inferior.

  Based on the above determination by the determination unit 22, if the measurement result received by the determination unit 22 is from the CCCCC, the determination unit 22 determines that there is an abnormality in step 306 (in step 306). If “No” (determination “YES”), the process proceeds to step 308, and if it is not from the CCCCC, the determination unit 22 determines that there is no abnormality in step 306 (determination “NO” in step 306) and proceeds to step 307.

(Description of second example of processing of master 2)
Subsequently, a second example of the parallel distributed processing method performed by the master 2 will be described with reference to FIG. First, the communication unit 20 receives a measurement result from each slave 1 (step 401). Next, the communication unit 20 associates the received measurement result with the transmission source ID for identifying the slave 1 that is the transmission source of the measurement result, and stores it in the totaling unit 21 as the totaling result (step 402). FIG. 11 and FIG. 12 are diagrams showing examples of tabulation result tables stored in the tabulation unit 21 immediately after step 402, respectively. In the example of the table shown in FIG. 11, the number of slaves 1 is five, and in the example of the table shown in FIG. 12, the number of slaves 1 is ten.

  Next, the determination unit 22 calculates a total result per report (step 403). The tabulated result per report calculated here is not the one for each slave 1 but the average value for all the slaves 1. In this calculation, only the latest measurement result of each slave 1 may be used, or all measurement results of each slave 1 may be used. In the table example shown in FIG. 11, the calculation result is “10”, and in the table example shown in FIG. 12, “20”. In the table example shown in FIG. 12, the number of slaves 1 to which the master 2 should respond is doubled compared to the table example shown in FIG. 11, and the response time viewed from each slave 1 is doubled. On the other hand, this is because the efficiency with which the main thread of each slave 1 processes a task does not depend on the number of slaves 1 in the parallel distributed processing system 4 as a whole.

  Next, the determination unit 22 determines whether or not the calculated total result per report exceeds a threshold value (step 404). This threshold value may be fixed programmatically or may be set by a setting file or the like. Note that the threshold here is set to “15”. This means that if the total result per report exceeds 15, the response time is too long, that is, the number of slaves 1 is too large.

  If the counting result stored in the counting unit 21 in step 404 is the table example shown in FIG. 11, the determination result by the determining unit 22 is “NO”, and the determining unit 22 sets the measurement result transmission source slave 1 to A normal response, for example, a response including an instruction “do nothing” is returned (step 405). On the other hand, in step 404, when the total result stored in the total unit 21 is the table example shown in FIG. 12, the determination result by the determination unit 22 is “YES”, and the determination unit 22 reduces the load on the master 2. To determine how to. For example, the determination unit 22 may select the slave 1 with the lowest processing efficiency (step 406) and return a stop instruction to the slave 1. In the case of the table example shown in FIG. 12, JJJJJJ having the lowest processing efficiency is selected.

  Subsequent to step 406, the determination unit 22 determines whether the measurement result of the transmission source slave 1 is from the slave 1 selected in step 406 (step 407). If it is determined in step 407 that the data is from the slave 1 (the determination “YES” in step 407), the determination unit 22 forcibly terminates the corresponding instruction information, for example, the slave 1 as a reply to the transmission source slave 1. Instruction information such as to do is returned (step 408). If it is determined in step 407 that the slave unit 1 is not from the slave 1 (determination “NO” in step 407), a normal response to the measurement result transmission source slave 1, for example, a response including an instruction “do nothing”. A reply is made (step 405). The slave selected in step 406 (here, JJJJJ) is stored in the master 2, and when a report is next received from JJJJJJ, the process proceeds to step 408, and the determination unit 22 returns a corresponding instruction. You may do it.

(Explanation of the third example of processing of the master 2)
Next, a third example of the parallel distributed processing method performed by the master 2 will be described with reference to FIG. First, the communication unit 20 receives a measurement result from each slave 1 (step 501). Next, the communication unit 20 determines whether to use time information (step 502). If it is determined in step 502 that the time information is not used, the communication unit 20 associates the received measurement result with the transmission source ID for identifying the slave 1 that is the transmission source of the measurement result, and calculates the totaling unit as the totaling result. 21 (step 503). Next, the determination unit 22 performs statistical processing of the total result per reception of the communication unit 20, for example, calculation of an average value (step 504).

  If it is determined in step 502 that the time information is used, the communication unit 20 receives the measurement result, the transmission source ID for identifying the slave 1 that is the transmission source of the measurement result, and the time when the measurement result is received. Are stored in the totaling unit 21 as a totaling result (step 505). FIG. 13 is a diagram illustrating a table example of a totaling result stored in the totaling unit 21 immediately after step 505. Next, the determination unit 22 performs statistical processing of the counting results per time, for example, calculates an average value (step 506). For example, from the table example shown in FIG. 13, the determination unit 22 determines that an average of 10 tasks can be processed per second in the entire parallel distributed processing system 4.

  Next, the determination unit 22 performs predictive determination and processing based on the statistical processing in step 504 or 506 (step 508). For example, assume that the total number of tasks to be processed in the entire parallel distributed processing system 4 is 1000. From the table example shown in FIG. 13, the determination unit 22 has processed 150 tasks in 15 seconds and processed 10 tasks on average per second. For example, the determination unit 22 may output the prediction result to the display unit 23, and the display unit 23 may output the prediction result to the user of the parallel distributed processing system 4. It becomes possible. Next, the determination unit 22 returns an instruction to the transmission source slave 1 (step 505). The determination unit 22 may instruct a specific process according to the determination based on the prediction, or may indicate “do nothing special”.

  Hereinafter, the effect of this embodiment is demonstrated.

  According to the parallel distributed processing method of this embodiment, the slave 1 measures task execution using a predetermined measurement unit in the parallel distributed processing system 4 and transmits the measurement result to the master 2. Thus, the master 2 can grasp the processing efficiency of each slave 1 without introducing an external monitoring computer or a monitoring program. Then, the master 2 issues an instruction to each slave 1 based on the processing efficiency of each slave 1, and each slave 1 controls the operation based on the instruction, so that the master 2 is based on the processing efficiency of each slave 1. The overall operation of the parallel distributed processing system 4 can be dynamically controlled.

  In addition, the determination unit 22 determines that the master 2 includes a total result per reception from the transmission source slave 1 by the communication unit 20 among the total results stored in the totalization unit 21 and a transmission source slave by the communication unit 20. It is also possible to compare the aggregated results per reception from the slaves 1 other than 1, generate instruction information for the transmission source slave 1 based on the comparison result, and transmit the instruction information to the transmission source slave 1. In such a case, the master 2 can issue an instruction to the target slave 1 based on a comparison result between the total result per reception of the target slave 1 and the total result per reception of the other slave 1. The processing efficiency of the target slave 1 can be grasped more accurately, and the operation of the entire parallel distributed processing system 4 can be dynamically controlled more efficiently.

  In addition, the communication unit 20 further associates the elapsed time from the reception of the previous measurement result from the transmission source slave 1, and stores it in the totaling unit 21 of the master 2 as a totaling result. , Out of the totaling results stored in the totaling unit 21, the master 2 has a value per unit time of the totaling result of one reception from the transmission source slave 1 by the communication unit 20 and the transmission source slave 1 by the communication unit 20. May be compared with a value per unit time of a total result of one reception from a slave 1 other than the slave 1, and instruction information for the transmission source slave 1 may be generated based on the comparison result and transmitted to the transmission source slave 1. . In such a case, the master 2 determines that the target slave 1 is based on a comparison result between the value per unit time of the total reception result of the target slave 1 and the value per unit time of the total reception result of the other slave 1. 1 can be instructed, the processing efficiency of the target slave 1 can be grasped more accurately, and the operation of the entire parallel distributed processing system 4 can be dynamically controlled more efficiently.

  In addition, the determination unit 22 determines that the master 2 calculates the average value of all the aggregation results of the transmission source slave 1 among the aggregation results stored in the aggregation unit 21 and all the slaves 1 other than the transmission source slave 1. It may be compared with the average value of the tabulation results, and the instruction information for the transmission source slave 1 may be generated based on the comparison result and transmitted to the transmission source slave 1. In this case, the master 2 can issue an instruction to the target slave 1 based on a comparison result between the average value of all the aggregation results of the target slave 1 and the average value of all the aggregation results of the other slaves 1. Therefore, the processing efficiency of the target slave 1 can be grasped more accurately, and the operation of the entire parallel distributed processing system 4 can be dynamically controlled more efficiently.

  In addition, the communication unit 20 further associates the elapsed time from the reception of the previous measurement result from the transmission source slave 1, and stores it in the totaling unit 21 of the master 2 as a totaling result. , Of the total results stored in the totaling unit 21 by the master 2, the average value of all the total results of the transmission source slave 1 and all of the slaves 1 other than the transmission source slave 1 It may be compared with the average value of the aggregated values per unit time, and the instruction information for the transmission source slave 1 may be generated based on the comparison result and transmitted to the transmission source slave 1. In such a case, the master 2 is based on the comparison result between the average value of all the aggregation results of the target slave 1 per unit time and the average value of all the aggregation results of other slaves 1 per unit time. Since the instruction can be issued to the target slave 1, the processing efficiency of the target slave 1 can be grasped more accurately and the operation of the entire parallel distributed processing system 4 can be dynamically controlled more efficiently. Can do.

  In addition, the determination unit 22 performs statistics on the aggregation results for all slaves 1 of the aggregation results per reception from the slave 1 by the communication unit 20 among the aggregation results stored in the aggregation unit 21. Processing may be performed to generate instruction information for the transmission source slave 1 based on the statistical processing result and transmit the instruction information to the transmission source slave 1. In such a case, since the master 2 can issue an instruction to each slave 1 based on the statistical processing result for the totaling results for all the slaves 1 in the totaling result per reception from the slave 1, The processing efficiency of each slave 1 can be accurately grasped, and the operation of the entire parallel distributed processing system 4 can be dynamically controlled more efficiently.

  In addition, the communication unit 20 further associates the elapsed time from the reception of the previous measurement result from the transmission source slave 1, and stores it in the totaling unit 21 of the master 2 as a totaling result. The master 2 performs statistical processing on the values related to all slaves 1 of the aggregated results of one reception from the slave 1 by the communication unit 20 among the aggregated results stored in the aggregating unit 21. The instruction information for the transmission source slave 1 may be generated based on the statistical processing result and transmitted to the transmission source slave 1. In such a case, the master 2 can issue an instruction to each slave 1 based on the statistical processing results of the values per unit time of the total result of one reception from the slave 1 with respect to the values related to all the slaves 1. Therefore, the processing efficiency of each slave 1 can be grasped more accurately, and the operation of the entire parallel distributed processing system 4 can be dynamically controlled more efficiently.

  In addition, the determination unit 22 determines that the average value for all slaves 1 of the total results per reception from the slave 1 in the communication unit 20 among the total results stored in the total unit 21 is a predetermined threshold value. When it exceeds, the master 2 may generate instruction information for the slave 1 to reduce the load on the master 2 and transmit the instruction information to the slave 1. In such a case, for example, the master 2 can detect that the number of slaves 1 exceeding the capacity of the master 2 is connected, and according to the capacity of the master 2 or a condition set according to the capacity. It becomes possible to dynamically control the number of slaves 1 connected.

  The master 2 needs to manage all the slaves 1 in the parallel distributed processing system 4, but since the processing capability of the master 2 is limited, it is not possible to manage a large number of slaves 1 without limit. Therefore, when management of the number of slaves 1 exceeding the capacity of the master 2 is requested, the master 2 can detect this by monitoring the load of the master 2 itself, and can select an appropriate behavior. For example, the master 2 may stop the slave 1 with low processing efficiency and maintain the best parallel distributed processing system 4 within a range that the master 2 can appropriately manage.

  Further, according to the parallel distributed processing system 4 of the present embodiment, the master 2 knows the processing efficiency of the entire parallel distributed processing system 4 and then notifies the user of predictive operations, for example, predicted completion times of all tasks. It can be presented.

  As an application example of the parallel distributed processing system 4 of the present embodiment, in the conventional parallel distributed processing system, the master 2 monitors the processing efficiency of each slave 1 without the help of an external monitoring computer or monitoring software, The master 2 monitors the degree of load of the master 2 itself, or the master 2 grasps the processing efficiency of the entire parallel distributed processing system 4 and predicts the operation of the entire parallel distributed processing system 4 in a predictive manner. 1 can be generated and used to control the behavior of the entire parallel distributed processing system 4.

  As described above, according to the parallel distributed processing system 4 of the present embodiment, the master 2 grasps the processing efficiency of each slave 1 in the master / slave type system without introducing an external monitoring computer or a monitoring program. The information can be used for assignment (scheduling) of tasks to each slave 1, for example. In addition, it becomes possible for the master 2 itself to detect that the number of slaves 1 exceeding the capability of the master 2 is connected. For example, according to the capability of the master 2 (or conditions set according to the capability) Thus, the number of connected slaves 1 can be dynamically controlled. Furthermore, by applying the fact that the master 2 can grasp the processing efficiency of each slave 1, it is possible to make a prediction regarding the overall operation of the parallel distributed processing system 4. For example, the estimated time until all tasks are completed is given to the user. By providing a service such as presenting, it is possible to give an instruction based on the prediction to the slave 1.

  DESCRIPTION OF SYMBOLS 1 ... Slave, 2 ... Master, 3 ... Network, 4 ... Parallel distributed processing system, 10 ... Processing part, 11 ... Storage part, 12 ... Report part, 13 ... Control part, 20 ... Communication part, 21 ... Total part, 22 ... judgment part, 23 ... display part.

Claims (9)

A parallel distributed processing method executed by a parallel distributed processing system that includes one or more slaves and a master connected to the slaves via a network and performs parallel distributed processing of tasks on the slaves. ,
The slave executes the task assigned to the slave, measures the execution of the task using a predetermined measurement unit in the parallel distributed processing system, and stores the measurement result in the storage unit of the slave Processing steps;
A first reporting step in which the slave transmits a measurement result stored in the storage means of the slave to the master;
The master receives a measurement result from the slave, associates the measurement result with information for identifying the slave of the transmission source, and a communication step of storing in the aggregation means of the master as the aggregation result;
Based on the counting result stored in the counting means, the master generates instruction information for the slave, and transmits to the slave.
A second reporting step in which the slave receives instruction information from the master;
A control step in which the slave controls the operation of the slave based on the instruction information received in the second reporting step;
A parallel distributed processing method. In the determination step, the master determines a total result per reception from the transmission source slave in the communication step out of the total results stored in the totalization unit, and other than the transmission source slave in the communication step. Comparing the aggregated result per reception from the slave, generating instruction information for the transmission source slave based on the comparison result, and transmitting to the transmission source slave,
The parallel distributed processing method according to claim 1. In the communication step, the master further associates the elapsed time from the reception of the previous measurement result from the slave of the transmission source, and stores it in the aggregation means of the master as the aggregation result,
In the determination step, the master determines a value per unit time of a totaling result of one reception from the transmission source slave in the communication step, among the totaling results stored in the totaling unit, and the communication step. Compare the value per unit time of the total result of one reception from the slave other than the transmission source slave, generate instruction information for the transmission source slave based on the comparison result, and transmit to the transmission source slave To
The parallel distributed processing method according to claim 1. In the determination step, the master calculates an average value of all the aggregation results of the transmission source slave among the aggregation results stored in the aggregation means, and all the aggregation results of the slaves other than the transmission source slave. The instruction value for the transmission source slave based on the comparison result, and transmit to the transmission source slave,
The parallel distributed processing method according to claim 1. In the communication step, the master further associates the elapsed time from the reception of the previous measurement result from the slave of the transmission source, and stores it in the aggregation means of the master as the aggregation result,
In the determining step, the master calculates an average value of values per unit time of all the aggregation results of the transmission source slave among the aggregation results stored in the aggregation means, and the other than the transmission source slave. Comparing the average value of all the aggregate results of the slaves per unit time, generating instruction information for the slave of the transmission source based on the comparison result, and transmitting to the slave of the transmission source,
The parallel distributed processing method according to claim 1. In the determination step, the master performs a statistical process on the aggregation results for all the slaves of the aggregation results received from the slave in the communication step among the aggregation results stored in the aggregation means. Performing instruction information for the transmission source slave based on the statistical processing result, and transmitting to the transmission source slave,
The parallel distributed processing method according to claim 1. In the communication step, the master further associates the elapsed time from the reception of the previous measurement result from the slave of the transmission source, and stores it in the aggregation means of the master as the aggregation result,
In the determination step, for the values related to all of the slaves, the values per unit time of the total result of one reception from the slave in the communication step among the total results stored in the total unit. Performing statistical processing, generating instruction information for the transmission source slave based on the statistical processing result, and transmitting to the transmission source slave,
The parallel distributed processing method according to claim 1. In the determination step, among the total results stored in the total means, an average value for all the slaves of the total results per reception from the slave in the communication step exceeds a predetermined threshold value. In this case, the master generates instruction information for the slave to reduce the load on the master, and transmits the instruction information to the slave.
The parallel distributed processing method according to any one of claims 1 to 7, wherein: A parallel distributed processing system that includes one or more slaves and a master connected to the slaves via a network, and performs parallel distributed processing of tasks on the slaves,
The slave is
A processing unit that executes the task assigned to the slave, measures the execution of the task using a predetermined measurement unit in the parallel distributed processing system, and stores the measurement result in the storage unit of the slave;
Reporting means for transmitting the measurement result stored in the storage means of the slave to the master and receiving instruction information from the master as a response to the transmission;
Control means for controlling the operation of the slave based on the instruction information received by the reporting means;
With
The master
A communication means for receiving a measurement result from the slave, associating the measurement result with information for identifying the slave of the transmission source, and storing the result in the master as a total result;
Based on the counting results stored in the counting means, generating instruction information for the slave, and determining means for transmitting to the slave;
A parallel distributed processing system.

Images