WO2014020959A1 - 分散処理装置及び分散処理システム並びに分散処理方法 - Google Patents
分散処理装置及び分散処理システム並びに分散処理方法 Download PDFInfo
- Publication number
- WO2014020959A1 WO2014020959A1 PCT/JP2013/062479 JP2013062479W WO2014020959A1 WO 2014020959 A1 WO2014020959 A1 WO 2014020959A1 JP 2013062479 W JP2013062479 W JP 2013062479W WO 2014020959 A1 WO2014020959 A1 WO 2014020959A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- distributed processing
- processes
- collective communication
- communication
- division pattern
- Prior art date
Links
Images
Classifications
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
- H04L67/00—Network arrangements or protocols for supporting network services or applications
- H04L67/01—Protocols
- H04L67/10—Protocols in which an application is distributed across nodes in the network
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F9/00—Arrangements for program control, e.g. control units
- G06F9/06—Arrangements for program control, e.g. control units using stored programs, i.e. using an internal store of processing equipment to receive or retain programs
- G06F9/46—Multiprogramming arrangements
- G06F9/50—Allocation of resources, e.g. of the central processing unit [CPU]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F9/00—Arrangements for program control, e.g. control units
- G06F9/06—Arrangements for program control, e.g. control units using stored programs, i.e. using an internal store of processing equipment to receive or retain programs
- G06F9/46—Multiprogramming arrangements
- G06F9/54—Interprogram communication
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
- H04L67/00—Network arrangements or protocols for supporting network services or applications
- H04L67/01—Protocols
- H04L67/12—Protocols specially adapted for proprietary or special-purpose networking environments, e.g. medical networks, sensor networks, networks in vehicles or remote metering networks
- H04L67/125—Protocols specially adapted for proprietary or special-purpose networking environments, e.g. medical networks, sensor networks, networks in vehicles or remote metering networks involving control of end-device applications over a network
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
- H04L67/00—Network arrangements or protocols for supporting network services or applications
- H04L67/34—Network arrangements or protocols for supporting network services or applications involving the movement of software or configuration parameters
Definitions
- the present invention relates to a distributed processing technique involving collective communication.
- Machine learning and data mining are methods for finding rules and patterns from data, and are used in various situations such as information recommendation, face authentication, voice recognition, and document classification.
- Various methods for such machine learning and data mining have been proposed.
- a model that explains the data is designed, a function that represents the degree of explanation of the data of the model (for example, log likelihood) is generated, and optimization of the model parameters (log likelihood) is performed for the function. Learning is performed by performing maximization.
- the steepest descent method For example, the steepest descent method, the stochastic gradient descent method, and the EM (Expectation Maximization) algorithm are used for the above optimization. Since the optimization processing requires more time as the learning target data becomes larger, parallel and distributed processing is desired when learning is performed with large-scale data.
- DGD Distributed Gradient Descent
- IPM Intelligent Parameter Mixtures
- Non-Patent Document 1 shows an experimental result that the optimization by the IPM method is fast in the structured perceptron and the maximum entropy method. Also, when implementing the DGD method or the IPM method, both methods can use MapReduce (see Non-Patent Document 3 below) which is a distributed processing framework. For this reason, these methods have an advantage that even those who have no knowledge about distributed programming can easily implement them.
- the algorithm of the DGD method and the IPM method includes a process of adding the gradients of all distributed model parameters. In this process, communication occurs in all machines where data is distributed. Therefore, such communication becomes a bottleneck, and in the DGD method and the IPM method, even if the number of machines that perform distributed processing is increased, there is a limit to speeding up the processing. That is, Reduce in MapReduce as a distributed processing framework and Allreduce in MPI (Message Passing Interface) become processing bottlenecks.
- MPI Message Passing Interface
- the present invention has been made in view of the above-described phenomenon, and provides a technique for executing distributed processing with collective communication at high speed.
- the first aspect relates to a distributed processing device that starts at least one of a plurality of processes started by a plurality of distributed processing devices.
- at least one activated process is performed between a plurality of processes in one process group among a plurality of process groups obtained by dividing the plurality of processes by the first division pattern.
- the second aspect relates to a distributed processing method executed by a computer that starts at least one of a plurality of processes started by a plurality of computers.
- the computer performs local collective communication between a plurality of processes in one process group among a plurality of process groups in which the plurality of processes are divided by a first division pattern.
- a second divided pattern different from the first divided pattern used in the local collective communication is selected, and the plurality of processes are divided into the second divided patterns.
- the next local collective communication is performed using a data based on the result of the local collective communication among a plurality of processes in one process group.
- a distributed processing system includes a plurality of distributed processing devices according to the first aspect, an information storage unit that stores at least one of the number of processor cores and connection network information for each distributed processing device, and a connection network
- An execution target determining unit that selects a distributed processing device that activates at least one of the plurality of processes preferentially from at least one of a distributed processing device having a common processor and a distributed processing device having a large number of processor cores;
- a distributed processing management device including an execution instruction unit that transmits an execution instruction for the process to each distributed processing device selected by the target determination unit.
- a program that causes a computer to realize the configuration according to the first aspect described above may be used, or a computer-readable recording medium that records such a program may be used.
- This recording medium includes a non-transitory tangible medium.
- the distributed processing apparatus activates at least one of a plurality of processes activated by the plurality of distributed processing apparatuses.
- the at least one process is a local group among a plurality of processes in one process group of the plurality of process groups in which the plurality of processes are divided by the first division pattern.
- Communication control means for causing the collective communication means to execute the next local collective communication using the data based on the result of the local collective communication among a plurality of processes in one process group of the plurality of process groups.
- the distributed processing method according to the present embodiment is executed by a computer that starts at least one of a plurality of processes started by a plurality of computers.
- the computer performs local collective communication between a plurality of processes in one process group among a plurality of process groups in which the plurality of processes are divided by a first division pattern.
- a second divided pattern different from the first divided pattern used in the local collective communication is selected, and the plurality of processes are divided into the second divided patterns.
- the next local collective communication is performed using a data based on the result of the local collective communication among a plurality of processes in one process group.
- the above division pattern indicates one grouping method for grouping a plurality of processes. Therefore, the second divided pattern different from the first divided pattern means another grouping technique different from the grouping technique of the first divided pattern.
- the plurality of processes described above execute a structured perceptron, a maximum entropy method, and the like in a distributed manner. It is clear from the contents of the following references, for example, that this embodiment works effectively in distributed learning processing using the maximum entropy method. References: Gideon Mann, Ryan McDonald, Mehryar Mohri, "Efficient Large-Scale Distributed Training of Conditional Maximum Entropy Models", NIPS, 2009
- the above-mentioned process means an arbitrary execution unit of software (program), and is not a word used for the purpose of distinguishing from a thread, a task, or the like.
- the above-described collective communication means that predetermined processing is performed on a plurality of processing target data possessed by a plurality of processes, and the obtained data is distributed to the plurality of processes.
- a process for collecting a plurality of processing target data from a plurality of processes and performing a predetermined process on the collected data may be one process or a plurality of processes.
- the predetermined process itself is not limited at all and is, for example, addition.
- local collective communication is performed in each process group formed by dividing a plurality of processes by the first division pattern.
- the collective communication described above is performed only for processes existing in each process group. Therefore, in the local collective communication, communication between processes belonging to different process groups does not occur.
- each data obtained by applying predetermined processing to the multiple processing target data of the multiple processes in each process group is transferred between the multiple processes in each process group.
- data obtained by performing predetermined processing on a plurality of processing target data possessed by a plurality of processes in the process group may be referred to as local processing data. That is, according to the present embodiment, the local processing data reflecting the processing target data of the processes in each process group is shared for each process group by the local collective communication for each process group.
- next local collective communication in order to perform the next local collective communication, a second divided pattern different from the first divided pattern used in the previous local collective communication is selected. Then, the next local collective communication is performed using data based on the result of the previous local collective communication in each process group corresponding to the second division pattern.
- the next local collective communication may be performed using the local processing data obtained by the previous local collective communication as it is, or each process is performed on the local process data obtained by the previous local collective communication. May each be implemented using data obtained by further processing.
- a predetermined process is performed on a plurality of processing target data that reflect the results of the previous local collective communication, which are possessed by a plurality of processes in a new process group.
- the obtained data is distributed to a plurality of processes in the new process group. Since the second division pattern is different from the first division pattern, at the time of the next local collective communication, at least one process that was not included in the collective communication counterpart at the time of the previous local collective communication becomes the collective communication counterpart for each process. . Further, at the time of the next local group communication, each process updates the local processing data reflecting the processing target data of other processes belonging to the same process group at the time of the previous local group communication or the local processing data. Each of which has been processed. Therefore, when the next local collective communication is completed, each process has local processing data reflecting more processing target data than when the previous local collective communication is completed.
- this embodiment by repeating such local collective communication, it is possible to acquire desired data reflecting all the processing target data distributed and arranged in a plurality of processes.
- this embodiment is not limited to repeating only local collective communication in order to obtain desired data.
- FIG. 1 is a diagram conceptually illustrating a configuration example of a distributed processing system in the first embodiment.
- the distributed processing system 1 in the first embodiment includes a master device 10 and a plurality of slave devices 20 (# 1, # 2,..., #N).
- each slave device 20 (# 1, # 2,..., #N) is generically referred to as a slave device 20 unless it is necessary to distinguish each individual device.
- the slave device 20 executes distributed processing with collective communication.
- the distributed processing device and the distributed processing method described above are applied to the slave device 20.
- the slave device 20 can also be called a distributed processing device.
- the master device 10 causes a plurality of slave devices 20 to execute distributed processing in response to a request from the user. Therefore, the master device 10 can also be called a distributed processing management device.
- the master device 10 and the slave device 20 are distinguished from each other. This is because both the master device 10 and the slave device 20 can actually execute the distributed processing program in the distributed processing system 1.
- the slave device 20 is the computer that can actually execute the distributed processing program.
- the master device 10 and the slave device 20 are communicably connected via the communication network 9.
- the communication network 9 is a public network such as the Internet, a WAN (Wide Area Network), a LAN (Local Area Network), a wireless communication network, or the like.
- the communication mode between the master device 10 and the slave device 20 and between the slave devices 20 is not limited.
- the master device 10 and the slave device 20 are so-called computers, and have a CPU (Central Processing Unit) 2, a memory 3, an input / output interface (I / F) 4, and the like that are connected to each other via a bus 5.
- the memory 3 is a RAM (Random Access Memory), a ROM (Read Only Memory), a hard disk, a portable storage medium, or the like.
- the input / output I / F 4 is connected via a communication network 9 to a communication device 7 that communicates with other computers (master device 10 and slave device 20).
- the input / output I / F 4 may be connected to an input unit (not shown) or an output unit (not shown).
- the input unit is a device that receives an input of a user operation such as a keyboard and a mouse.
- the output unit is a device that provides information to a user, such as a display device or a printer.
- each of the master device 10 and the slave device 20 has one CPU 2 is shown, but they may have a plurality of CPUs 2.
- FIG. 2 is a diagram conceptually illustrating a configuration example of the master device 10 in the first embodiment.
- the master device 10 includes a slave information storage unit 11, an initial setting unit 12, a slave control unit 13, an execution instruction unit 14, an execution request acquisition unit 15, and the like.
- Each of these processing units is realized, for example, by executing a program stored in the memory 3 by the CPU 2.
- the program is installed from a portable recording medium such as a CD (Compact Disc) or a memory card or another computer on the network via the input / output I / F 4 and stored in the memory 3.
- a portable recording medium such as a CD (Compact Disc) or a memory card or another computer on the network via the input / output I / F 4 and stored in the memory 3.
- CD Compact Disc
- the slave information storage unit 11 stores various information related to each slave device 20 such that the master device 10 identifies each slave device 20 and communicates with each slave device 20, such as a machine name and an IP address.
- a machine name In the case of using a DNS (Domain Name System) service, only the machine name may be stored.
- DNS Domain Name System
- the initial setting unit 12 performs initial setting of the master device 10.
- the initial setting unit 12 stores information related to the slave device 20 specified by the user in the slave information storage unit 11 after the master device 10 is activated.
- Information regarding the slave device 20 may be input by a user by operating an input unit, for example, or may be acquired from a setting file generated by the user.
- the slave control unit 13 starts and stops the plurality of slave devices 20 based on the information related to the slave device 20 stored in the slave information storage unit 11.
- the execution request acquisition unit 15 acquires an execution request for the distributed processing program. For example, in this execution request, a distributed processing program to be executed by a plurality of slave devices 20 is specified.
- the execution request may include the number of processes realized by the distributed processing program.
- information set in advance may be used.
- the execution request may be input from an input unit (not shown) or may be acquired from another computer.
- the execution instruction unit 14 sends a distributed processing execution instruction to each slave device 20.
- the execution instructing unit 14 determines the number of processes of the slave device 20 that starts the process for distributed processing and the number of processes of each slave device 20 according to the number of processes acquired by the execution request acquiring unit 15, and executes the determination information. Include in instructions.
- FIG. 3 is a diagram conceptually illustrating a configuration example of the slave device 20 in the first embodiment.
- the slave device 20 includes an instruction acquisition unit 21, a program acquisition unit 22, a program execution unit 23, a slave information storage unit 24, and the like.
- Each of these processing units is realized, for example, by executing a program stored in the memory 3 by the CPU 2.
- the program is installed from a portable recording medium such as a CD or a memory card or another computer on the network via the input / output I / F 4 and stored in the memory 3.
- the instruction acquisition unit 21 receives a distributed processing execution instruction from the master device 10.
- This execution instruction includes information on the distributed processing program to be executed, the number of startup processes, and information on the plurality of slave devices 20 participating in the distributed processing.
- the instruction acquisition unit 21 stores information on the slave device 20 in the slave information storage unit 24.
- the slave information storage unit 24 stores information regarding a plurality of slave devices 20 participating in the distributed processing. Specifically, for each slave device 20 participating in the distributed processing, a machine name, an IP address, the number of startup processes, and the like are stored.
- the program acquisition unit 22 acquires the distributed processing program specified by the distributed processing execution instruction received by the instruction acquisition unit 21.
- the program acquisition unit 22 acquires the distributed processing program designated from another computer based on the location information.
- a distributed processing program may be sent from the master device 10 together with the execution instruction.
- the program execution unit 23 executes the distributed processing program acquired by the program acquisition unit 22 by the number of processes specified by the execution instruction acquired by the instruction acquisition unit 21. At this time, the program execution unit 23 uniquely assigns each process ID in order to distinguish each process started on the same slave device 20.
- Each process that performs distributed processing that is started based on the distributed processing program in each slave device 20 is referred to as a process 25. Accordingly, at least one process 25 is generated in each slave device 20 that has received an instruction to execute distributed processing from the master device 10.
- the process 25 includes a collective communication unit 26, a communication control unit 27, and the like. Since the collective communication unit 26 and the communication control unit 27 are realized by executing a distributed processing program, they can also be referred to as collective communication means and communication control means.
- the collective communication unit 26 is a local collective between a plurality of processes in one process group among a plurality of process groups divided in a certain division pattern, like the collective communication unit of the distributed processing apparatus in the above-described embodiment. Communicate.
- the collective communication unit 26 executes Allreduce in MPI, for example, only for a plurality of processes in a certain process group.
- FIG. 4 is a diagram showing an example (addition) of Allreduce.
- Reference numerals N1, N2, N3 and N4 denote processes 25, respectively. That is, FIG. 4 shows an example in which there are four processes N1 to N4 in a certain process group. Each process has a vector of type INT (integer).
- the collective communication unit 26 of the process N1 collects vectors from the other processes N2, N3, and N4, and adds the elements of all the vectors.
- the collective communication unit 26 of the process N1 distributes the vector (6, 7, 9) obtained as a result of the addition to the other processes N2, N3, and N4.
- the communication control unit 27 selects a division pattern for determining a unit (process group) for performing local collective communication in the same manner as the communication control means of the distributed processing device in the above-described embodiment. Specifically, the communication control unit 27 holds information on a plurality of division patterns that can be selected and information on a selection order of the plurality of division patterns, and uses them for the previous local group communication based on the holding information. The division pattern used for the next local collective communication is selected so that the division pattern is different from the divided division pattern.
- FIG. 5 is a diagram illustrating an example of three division patterns related to the four processes 25.
- the process is divided into two process groups each including two processes 25.
- the communication control unit 27 holds information regarding the three division patterns shown in FIG. 5, and after the local collective communication is repeated a predetermined number of times, a plurality of pieces of data reflecting a plurality of target data possessed by the plurality of processes 25 are stored. The information regarding the selection order as shared by each of the processes 25 is held. In the division pattern shown in FIG. 5, the following two selection orders can be maintained.
- Selection order 1 The division pattern A-1 and the division pattern A-2 are alternately and repeatedly selected.
- Selection order 2 The division pattern A-1, the division pattern A-2, and the division pattern A-3 are repeatedly selected in order.
- the flow of data processing when selection order 1 and 2 are executed is as follows.
- the target data possessed by each of the processes N1, N2, N3, and N4 is expressed as D1, D2, D3, and D4
- the operation result data that reflects the target data D1 and the target data D2 is expressed as data D1D2.
- the data D1D2 is shared between the processes N1 and N2 by the local collective communication between the process N1 and the process N2, and the data D3D4 is processed by the local collective communication between the process N3 and the process N4. Shared by N3 and N4.
- data D1D2D3D4 is shared by processes N1 and N4 by local collective communication between process N1 and process N4, and data D1D2D3D4 is processed by local collective communication between process N2 and process N3. Shared by N2 and N3.
- the operation result data (D1D2D3D4) reflecting all the target data of the processes N1 to N4 is transferred to all the processes N1 to N4 by two local collective communications. Shared.
- FIG. 6 is a diagram illustrating an example of three division patterns related to the eight processes 25.
- the process is divided into four process groups each including two processes 25.
- there are 105 division patterns for dividing the eight processes 25 by two processes (8! / (4! 2! 2! 2! 105). Three of these division patterns are shown in FIG.
- the following two selection orders can be taken.
- Selection order 1 The division pattern B-1 and the division pattern B-2 are alternately and repeatedly selected.
- Selection order 2 The division pattern B-2 and the division pattern B-3 are alternately and repeatedly selected.
- the operation result data reflecting all the target data of the processes N1 to N8 is shared by all the processes N1 to N8 by four steps, that is, four times of local collective communication. Even in the selection order in which the division pattern B-1, the division pattern B-2, and the division pattern B-3 are repeatedly selected in order, the operation result data reflecting all the target data of the processes N1 to N8 is reflected in several steps. It can be shared by all processes N1 to N8.
- FIG. 7 is a diagram illustrating an example of three division patterns related to the eight processes 25.
- the process is divided into two process groups each including four processes 25.
- Three of these division patterns are shown in FIG.
- a selection order in which any two division patterns are alternately and repeatedly selected, or a selection order in which the three division patterns are repeatedly selected in order can be executed. According to these selection orders, the calculation result data reflecting all the target data of the processes N1 to N8 can be shared by all the processes N1 to N8 in two steps.
- the communication control unit 27 uses data based on the result of the previous local collective communication between a plurality of processes in one process group among a plurality of process groups divided by the division pattern thus selected.
- the group communication unit 26 is caused to execute the next local group communication.
- FIG. 8 is a diagram illustrating an operation example of the distributed processing system 1 in the first embodiment.
- the master device 10 stores information of each slave device 20 in the slave information storage unit 11 in order to communicate with each slave device 20 (S81).
- the slave information storage unit 11 stores a machine name and an IP address related to each slave device 20 as follows, for example. Node_A, 192.168.0.2 Node_B, 192.168.0.3 Node_C, 192.168.0.4 Node_D, 192.168.0.5 Node_E, 192.168.0.6 Node_F, 192.168.0.7
- the master device 10 activates each slave device 20 based on the information stored in the slave information storage unit 11 (S82). Each slave device 20 that receives the activation request from the master device 10 is activated (S83). According to the above example, the six slave devices 20 from Node_A to Node_F are activated. Each slave device 20 may be independently activated without being requested by the master device 10. In this case, the step (S82) is omitted.
- the master device 10 acquires an execution request for the distributed processing program (S84).
- This execution request includes, for example, the following information.
- the master device 10 determines the number of processes of the slave device 20 that activates the process for distributed processing and each slave device 20, and executes the distributed processing on each slave device 20 determined as the target.
- Each instruction is sent (S85).
- the master device 10 transmits the information of the distributed processing program, the information of the slave device 20 determined as the target of the distributed processing, the number of processes of each slave device 20 and the like to the slave devices 20 together with the execution instruction.
- the slave device 20 When the slave device 20 receives the execution instruction of the distributed processing, the slave device 20 stores the information of each slave device 20 acquired together with the execution instruction in the slave information storage unit 24 (S86).
- the slave device 20 acquires a distributed processing program based on the information acquired together with the execution instruction (S87). For example, when the home directory is shared by NFS (Network File System) or the like, each slave device 20 acquires a distributed processing program (/ home / userA / modified_ipm) by network access.
- NFS Network File System
- Each slave device 20 executes the acquired distributed processing program for the designated number of processes (S88). According to the above example, the slave devices 20 of Node_A and Node_B each start two processes 25, and the other slave devices 20 each start one process 25. Local collective communication is executed in each process 25 activated based on the distributed processing program. The operation of each process 25 will be described later. When each process 25 ends, each slave device 20 sends the result of distributed processing by each process 25 to the master device 10.
- Master device 10 receives the result of the distributed processing from each slave device 20, and outputs the result data (S89).
- the specific form of outputting the result data is not limited, and is realized by, for example, a display unit or another output device (printer or the like) via the input / output I / F 4.
- the master device 10 sends a stop request to each slave device 20 (S90). Receiving this stop request, each slave device 20 stops (S91).
- the stop of each slave device 20 may be a shutdown or a transition to a predetermined operation mode such as a power saving mode.
- each slave device 20 may be independently stopped regardless of a request from the master device 10. In this case, the step (S90) is omitted.
- FIG. 9 is a flowchart showing an operation example of each process 25 in the first embodiment. Here, it is assumed that the process is started in the following form based on the example of FIG. Node_A: Process N1 Node_A: Process N2 Node_B: Process N3 Node_B: Process N4 Node_C: Process N5 Node_D: Process N6 Node_E: Process N7 Node_F: Process N8
- the process 25 When the process 25 is activated, it selects one division pattern from the plurality of division patterns held (S95). Since each process 25 is generated based on the same distributed processing program, the same division pattern is selected. Here, in each process 25, the division patterns C-1 and C-2 shown in FIG. 7 are held, and the selection order in which the division patterns C-1 and C-2 are alternately repeated is held as the selection order. Assuming that According to this assumption, the process 25 first selects the division pattern C-1.
- the process 25 acquires information on the selected division pattern (S96).
- the information acquired here includes information on other processes 25 belonging to the same process group with which each process 25 performs local collective communication and a specific method for realizing the local collective communication.
- the specific realization method of local collective communication indicates the role of each process 25 in local collective communication (data sending side, data receiving side, whether or not to perform a predetermined calculation, etc.) and the like.
- the process N1 acquires a machine name, an IP address, and the like regarding the processes N2 to N4 belonging to the same process group.
- the process N1 collects data from other processes N2 to N4 as a role of local collective communication, performs a predetermined operation on the collected data, and sends the operation result to the other processes N2 to N4. Recognize that they have a role to distribute.
- the process 25 executes local collective communication between a plurality of processes 25 in the same process group corresponding to the division pattern selected in the step (S95) based on the information on the division pattern (S97). According to the example of FIG. 7, local collective communication is performed in processes N1 to N4, and local collective communication is performed in processes N5 to N8.
- the process 25 determines whether or not the execution step (execution count) of the local collective communication has been completed for a predetermined step (predetermined count) (S98).
- the predetermined step indicates the number of times that the calculation result data reflecting all the target data of each process 25 can be shared by all the processes 25. If the process 25 is completed for a predetermined number of steps (S98; YES), the process ends. On the other hand, when the predetermined steps are not completed (S98; NO), the process 25 returns to the step (S95) and continues the process.
- the process 25 selects a division pattern that is different from the division pattern selected last time, based on the selection order information held. According to the example of FIG. 7, the process 25 selects the division pattern C-2. Thus, in the next step, local collective communication is performed in processes N3 to N6, and local collective communication is performed in the remaining processes.
- the communication load associated with collective communication can be spread, and communication associated with collective communication can be prevented from becoming a bottleneck for speeding up distributed processing.
- each process 25 for distributed processing activated by each slave device 20 after a plurality of division patterns and local collective communication are repeated a predetermined number of times, the target data possessed by all processes The order in which multiple division patterns are selected so that data that reflects is shared by all processes is maintained, and local collective communication at each step is performed based on the division pattern selected based on this information. Executed.
- the first embodiment since the number of repetitions of local collective communication is suppressed, it is possible to suppress a decrease in processing speed accompanying an increase in the number of repetitions. Furthermore, according to the first embodiment, a desired distributed processing result can be obtained even if the number of repetitions of local collective communication is suppressed.
- the slave device 20 that starts each process 25 and the number of processes that are started by each slave device 20 are determined by the number of processor cores of each slave device 20 and the connection network.
- the distributed processing system 1 according to the second embodiment will be described with a focus on the content different from the first embodiment. In the following description, the same contents as those in the first embodiment are omitted as appropriate.
- the slave information storage unit 11 further stores the number of processor cores and connection network information for each slave device 20.
- the connection network information is information that can distinguish a network segment or an Ethernet segment (Ethernet is a registered trademark) to which each slave device 20 is connected. Therefore, the same connection network information is set for each slave device 20 belonging to the same network segment.
- the execution instruction unit 14 preferentially executes the distributed processing from the slave device 20 having a large total number of processor cores for each connection network and a large number of processor cores.
- the slave device 20 is selected, and further, the number of processes to be activated by the selected slave device 20 is determined according to the number of processor cores.
- the execution instruction unit 14 can also be called a target determining unit.
- the execution instructing unit 14 transmits to each slave device 20 thus selected an execution instruction including connection network information and the number of processor cores related to the slave devices 20 participating in the distributed processing, and the determined number of processes. .
- FIG. 10 is a diagram conceptually illustrating a configuration example of the slave device 20 in the second embodiment.
- the slave device 20 in the second embodiment further includes a process determination unit 29 in addition to the configuration of the first embodiment. Since the process determination unit 29 is realized by executing a distributed processing program, it can also be called a process determination unit.
- the process determination unit 29 determines which process 25 in the division pattern that the process 25 is initially selected by the communication control unit 27 corresponds to. Based on the information regarding each slave device 20 participating in the distributed processing received from the master device 10 together with the execution instruction, the process determination unit 29 does not generate communication as much as possible in the local collective communication, and the communication latency is high. The above decision is made so that it can be suppressed as much as possible. A specific determination method by the process determination unit 29 will be described later with reference to FIG.
- the master device 10 stores the following information in the slave information storage unit 11. Node_A, 192.168.0.2, number of cores (4), network A Node_B, 192.168.0.3, number of cores (2), network B Node_C, 192.168.0.4, number of cores (2), network C Node_D, 192.168.0.5, number of cores (2), network A Node_E, 192.168.0.6, number of cores (1), network A Node_F, 192.168.0.7, number of cores (2), network A Node_G, 192.168.0.8, number of cores (4), network B
- step (S84) the master device 10 acquires an execution request for the distributed processing program including the following information.
- Distributed processing program / home / userA / modified_ipm Number of startup processes: 8
- step (S85) the master device 10 selects the slave device 20 that is the transmission destination of the distributed processing execution instruction as follows.
- the master device 10 acquires the slave information as described above from the slave information storage unit 11, calculates the total number of processor cores for each connection network, and sorts the connection networks in descending order of the total value. According to the slave information example, the following list is calculated.
- Network A Number of cores (9)
- Network B Number of cores (6)
- Network C Number of cores (2)
- the master device 10 selects a network so as to satisfy the number of startup processes (8) specified by the execution request with priority from the upper network in the list. According to the above example, only the network A that satisfies the number of startup processes (8) is selected. As a result, the slave device 20 connected to the same network having a low communication latency is preferentially selected.
- the master device 10 is connected to the selected network A and has priority from the slave device 20 having a large number of processor cores, and satisfies the number of start processes (8) specified by the execution request. Select. According to the above example, the following three slave devices 20 are selected. As a result, the slave device 20 that does not generate inter-node communication as much as possible is preferentially selected. Node_A, 192.168.0.2, number of cores (4), network A Node_D, 192.168.0.5, number of cores (2), network A Node_F, 192.168.0.7, number of cores (2), network A
- the master device 10 determines the slave device 20 thus selected as the slave device 20 participating in the distributed processing, and further determines the number of processes to be executed by each slave device 20 according to the number of processor cores.
- the number of processors of the slave device 20 of Node_A is determined to be 4
- the number of processors of the slave device 20 of Node_D is determined to 2
- the number of processors of the slave device 20 of Node_F is determined to 2.
- the master device 10 sends a distributed processing execution instruction to each slave device 20 selected in this way together with the number of processes (S85).
- the selection method as described above it is possible to select the slave device 20 that does not cause inter-node communication as much as possible and has a low communication latency as a device participating in the distributed processing.
- FIG. 11 is a flowchart showing an operation example of each process 25 in the second embodiment.
- the division patterns C-1 and C-2 shown in FIG. 7 are held, and as the selection order, the selection order in which the division patterns C-1 and C-2 are alternately repeated is held.
- the process 25 selects one division pattern from the plurality of division patterns held (S111). Based on the above assumption, each process 25 selects a division pattern C-1.
- the process 25 operates as follows.
- the process 25 generates a list of process sets stored in each process group divided by the selected division pattern (S113). According to the example of FIG. 7, the following list of process sets is generated.
- (Process set list) ⁇ Process N1, Process N2, Process N3, Process N4 ⁇ ⁇ Process N5, Process N6, Process N7, Process N8 ⁇
- Process 25 sorts the list of process sets related to each process group in descending order of the number of elements (number of processes) (S114). In the example of FIG. 7, since the number of elements is the same in both process sets (4), the sorting result does not change.
- the process 25 sorts the information of the slave devices 20 participating in the distributed processing in descending order of the number of startup processes designated from the master device 10 (S115). (Slave device list) Node_A, 192.168.0.2, number of processes (4), network A Node_D, 192.168.0.5, number of processes (2), network A Node_F, 192.168.0.7, number of processes (2), network A
- the process 25 preferentially assigns each process included in the upper process set in the process set list to the upper slave apparatus 20 in the slave apparatus list (S116). According to the example of FIG. 7, the assignment shown in FIG. 12 is determined.
- FIG. 12 is a diagram conceptually illustrating an example of process allocation in the second embodiment.
- the process 25 recognizes which process in the division pattern corresponds to the self process according to the allocation result and the information of the slave device 20 in which the self process is activated (S117).
- the process 25 started by the slave device 20 of Node_D recognizes that the own process is the process N5 or N6 in the division pattern C-1 based on the assignment result shown in FIG.
- the process 25 acquires information on the selected division pattern based on the recognition result (S118).
- the process 25 executes local collective communication between a plurality of processes 25 in the same process group corresponding to the division pattern selected in the step (S111) based on the acquired information (S119). In each local collective communication in this division pattern C-1, communication occurs only between the slave device 20 of Node_D and the slave device 20 of Node_F.
- the process 25 does not execute the steps (S113) to (S117). This is because the recognition of which process in the division pattern corresponds to the own process has already been completed. Thereafter, a step (S120) similar to the step (S98) of FIG. 9 in the first embodiment is performed.
- the second embodiment based on the number of processor cores of each slave device 20 and the connection network, the number of slave devices 20 participating in distributed processing and the number of processors activated by the slave devices 20 are determined. Each process 25 is associated with the division pattern selected for the first time. Such determination and association are performed so that local collective communication is performed as much as possible in the multi-core in the same slave device 20 and as much as possible between the slave devices 20 connected to the same network. .
- FIG. 13 is a diagram conceptually showing the configuration of the distributed processing system 1 in the embodiment.
- the distributed processing system 1 further includes a DNS server device 52, an NFS server device 53, and a client device 51 in addition to the devices (the master device 10 and the slave device 20) described in the above embodiments.
- the communication network 9 is realized as a LAN.
- the client device 51, the DNS server device 52, and the NFS server device 53 are so-called computers, like the master device 10 and the slave device 20.
- the master device 10 and the slave device 20 access the file stored in the NFS server device 53 by the client device 51. can do.
- the machine name of the master device 10 is MAT1
- the machine names of the eight slave devices 20 are SLV1 to SLV8.
- the corresponding IP address is obtained by resolving the machine name MAT1 by the DNS service provided by the DNS server device 52.
- machine names SLV1 to SLV8 of the slave device 20 are stored. Also, eight slave devices 20 are activated. In this state, the user operates the client device 51 to send an execution request for the distributed processing program to the master device 10. In this embodiment, the following information is sent together with the execution request.
- Distributed processing program / home / userA / superceptron Number of startup processes: 8
- the distributed processing program is a distributed learning program of the structured perceptron as follows.
- For i 1: Max // The following (1) and (2) are repeated Max times.
- (1) Foreach (x, y) in Data a) Update w if current estimate is wrong (2)
- w PseudoAllredcue (w, i, 4) // Allreduce each 4 units (local collective communication) 4).
- w Allreduce (w) / n
- the master device 10 When the master device 10 receives the above-described execution request from the client device 51, the master device 10 acquires the machine names SLV1 to SLV8 of the slave device 20 that can be activated from the slave information storage unit 11. Since the number of activation processes is specified in the execution request, the master device 10 causes each slave device 20 to activate one process at a time. Specifically, the master device 10 accesses each of the eight slave devices 20, and transmits the location where the distributed processing program exists and the machine name of the slave device 20 in charge of execution (SLV1 to SLV8).
- the slave device 20 stores the machine name of the slave device 20 sent together with the execution instruction from the master device 10 in the slave information storage unit 24. Next, the slave device 20 reads the distributed processing program from / home / userA / superceptron and executes it. As a result, each slave device 20 starts one process for performing distributed learning processing of the structured perceptron.
- SLV1 Process N1
- SLV2 Process N2
- SLV3 Process N3
- SLV4 Process N4
- SLV5 Process N5
- SLV6 Process N6
- SLV7 Process N7
- SLV8 Process N8
- the distributed processing program performs Max iterations.
- PseudoAllredcue is executed for each iteration.
- the execution of PseudoAllreduce in this embodiment corresponds to the collective communication unit 26.
- the communication control unit 27 holds information on the division patterns C-1 and C-2 shown in FIG. 7 and a selection order for repeating them alternately.
- the communication control unit 27 determines whether the division pattern C-1 or the division pattern C-2 in FIG. 7 is based on whether the value of the argument i is an odd number or an even number. For example, when i is 1, the division pattern C-1 is selected.
- the local collective communication is repeated a predetermined number of steps, but the repeat of the local collective communication may be terminated according to the result of the local collective communication (local processing data).
- FIG. 14 is a diagram conceptually illustrating a configuration example of the slave device 20 in the modified example.
- the process 25 further includes an evaluation unit 31 in addition to the configurations of the above-described embodiments. Since the evaluation unit 31 is realized by executing the distributed processing program, it can also be called an evaluation unit.
- the evaluation unit 31 evaluates data based on the result of local collective communication (local processing data).
- the evaluation unit 31 may evaluate the result of local collective communication itself, or may evaluate data obtained by processing the result of local collective communication. For example, when the difference between the result of the previous local group communication and the result of the current local group communication is smaller than a predetermined threshold, the evaluation unit 31 evaluates that the local processing data satisfies a desired level.
- the evaluation unit 31 may have a predetermined evaluation function, and may evaluate that the local processing data satisfies a desired level as a result obtained by giving local processing data to the evaluation function.
- the communication control unit 27 determines whether or not to continue the process according to the evaluation result of the evaluation unit 31. This determination can determine whether or not the next local collective communication is executed by the collective communication unit 26.
- the communication control unit 27 ends the process. Thereby, the process of the collective communication unit 26 is not performed.
- the communication control unit 27 continues the process. Thereby, the next local collective communication is executed by the collective communication unit 26.
- the number of processes in each process group is set to be the same in all division patterns.
- the number of processes in the process group may be different.
- the communication control unit 27 may select a division pattern in which the number of processes in each process group is larger than the initial iteration.
- the communication control unit 27 holds a plurality of division patterns having different numbers of processes in the process group, and uses them for the next local collective communication from among the plurality of division patterns according to the evaluation result of the evaluation unit 31.
- the division pattern to be selected may be selected.
- the communication control unit 27 holds in advance a selection order in which the number of processes in the process group increases in the later iterations, and according to this selection order, the number of processes in each process group is the previous local collective communication. More division patterns may be selected as division patterns for the next local collective communication.
- the result of local collective communication can be quickly reached a desired level, so that the number of repetitions of local collective communication can be further suppressed, and consequently, the speed of distributed processing with collective communication is increased. Can be realized.
- a division pattern used for the next local group communication may be dynamically determined.
- the communication control unit 27 of each process 25 sends the local processing data of each process 25 to the master device 10, and the master device 10 dynamically determines the division pattern based on the collected local processing data.
- the communication control unit 27 of each process 25 may use the determined division pattern for the next local collective communication.
- both the number of processor cores of each slave device 20 and the connection network information are considered, but the slave device 20 that participates in the distributed processing may be selected based on either one.
- the slave information storage unit 11 of the master device 10 may store the number of processor cores or connection network information for each slave device 20.
- the execution instruction unit 14 of the master device 10 may preferentially select the slave devices 20 having the same connection network, or may preferentially select the slave devices 20 as the number of processor cores is large. Even in such a form, since communication is suppressed in each local collective communication or communication latency is controlled to be small, it is possible to realize high speed distributed processing with collective communication. .
Landscapes
- Engineering & Computer Science (AREA)
- Software Systems (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Signal Processing (AREA)
- Computer Networks & Wireless Communication (AREA)
- Health & Medical Sciences (AREA)
- Computing Systems (AREA)
- General Health & Medical Sciences (AREA)
- Medical Informatics (AREA)
- Computer And Data Communications (AREA)
- Data Exchanges In Wide-Area Networks (AREA)
- Multi Processors (AREA)
Abstract
Description
参考文献:Gideon Mann, Ryan McDonald, Mehryar Mohri, "Efficient Large-Scale Distributed Training of Conditional Maxmum Entropy Models", NIPS, 2009
また、上述のプロセスとは、ソフトウェア(プログラム)の任意の実行単位を意味し、スレッド、タスク等と区別する意図で利用される文言ではない。
〔システム構成〕
図1は、第1実施形態における分散処理システムの構成例を概念的に示す図である。第1実施形態における分散処理システム1は、マスタ装置10、複数のスレーブ装置20(#1、#2、・・・、#n)を有する。以降、各スレーブ装置20(#1、#2、・・・、#n)は、特に個々を区別する必要がある場合を除き、スレーブ装置20と総称する。
〈マスタ装置〉
図2は、第1実施形態におけるマスタ装置10の構成例を概念的に示す図である。図2に示されるように、マスタ装置10は、スレーブ情報格納部11、初期設定部12、スレーブ制御部13、実行指示部14、実行要求取得部15等を有する。これら各処理部は、例えば、メモリ3に格納されるプログラムがCPU2により実行されることにより実現される。当該プログラムは、例えば、CD(Compact Disc)、メモリカード等のような可搬型記録媒体やネットワーク上の他のコンピュータから入出力I/F4を介してインストールされ、メモリ3に格納される。
図3は、第1実施形態におけるスレーブ装置20の構成例を概念的に示す図である。図3に示されるように、スレーブ装置20は、指示取得部21、プログラム取得部22、プログラム実行部23、スレーブ情報格納部24等を有する。これら各処理部は、例えば、メモリ3に格納されるプログラムがCPU2により実行されることにより実現される。当該プログラムは、例えば、CD、メモリカード等のような可搬型記録媒体やネットワーク上の他のコンピュータから入出力I/F4を介してインストールされ、メモリ3に格納される。
選択順序2:分割パターンA-1、分割パターンA-2、分割パターンA-3を順に繰り返し選択する。
選択順序2:分割パターンB-2と分割パターンB-3とを交互に繰り返し選択する。
以下、第1実施形態における分散処理方法について図8及び図9を用いて説明する。図8は、第1実施形態における分散処理システム1の動作例を示す図である。
Node_A, 192.168.0.2
Node_B, 192.168.0.3
Node_C, 192.168.0.4
Node_D, 192.168.0.5
Node_E, 192.168.0.6
Node_F, 192.168.0.7
分散処理プログラム /home/userA/modified_ipm
起動プロセス数 8
Node_A, プロセス数=2
Node_B, プロセス数=2
Node_C, プロセス数=1
Node_D, プロセス数=1
Node_E, プロセス数=1
Node_F, プロセス数=1
Node_A:プロセスN1
Node_A:プロセスN2
Node_B:プロセスN3
Node_B:プロセスN4
Node_C:プロセスN5
Node_D:プロセスN6
Node_E:プロセスN7
Node_F:プロセスN8
このように第1実施形態では、上述の実施形態における分散処理装置及び分散処理方法のように、選択された分割パターンに基づいて、全プロセスが複数のプロセスグループに分割された各プロセスグループ内のプロセス間で局所集団通信が行われる。即ち、第1実施形態では、複数の局所集団通信が並列に実行され、各局所集団通信に参加するプロセス数は、全プロセスで集団通信を行うのに比べて、減少する。
第2実施形態では、各プロセス25を起動させるスレーブ装置20及び各スレーブ装置20で起動させるプロセス数が、各スレーブ装置20のプロセッサコア数及び接続ネットワークにより決定される。以下、第2実施形態における分散処理システム1について、第1実施形態と異なる内容を中心説明する。以下の説明では、第1実施形態と同様の内容については適宜省略する。
スレーブ情報格納部11は、各スレーブ装置20に関し、第1実施形態で示された情報に加えて、プロセッサコア数及び接続ネットワーク情報を更にそれぞれ格納する。接続ネットワーク情報は、各スレーブ装置20が接続されるネットワークセグメントやイーサネットセグメント(イーサネットは登録商標)を区別し得る情報である。よって、同一ネットワークセグメントに属する各スレーブ装置20については同じ接続ネットワーク情報が設定される。
以下、第2実施形態における分散処理システム1の動作例について図8及び図11を用いて説明する。
Node_A, 192.168.0.2, コア数(4), ネットワークA
Node_B, 192.168.0.3, コア数(2), ネットワークB
Node_C, 192.168.0.4, コア数(2), ネットワークC
Node_D, 192.168.0.5, コア数(2), ネットワークA
Node_E, 192.168.0.6, コア数(1), ネットワークA
Node_F, 192.168.0.7, コア数(2), ネットワークA
Node_G, 192.168.0.8, コア数(4), ネットワークB
分散処理プログラム: /home/userA/modified_ipm
起動プロセス数: 8
ネットワークA: コア数(9)
ネットワークB: コア数(6)
ネットワークC: コア数(2)
Node_A, 192.168.0.2, コア数(4), ネットワークA
Node_D, 192.168.0.5, コア数(2), ネットワークA
Node_F, 192.168.0.7, コア数(2), ネットワークA
(プロセス集合のリスト)
{プロセスN1,プロセスN2,プロセスN3,プロセスN4}
{プロセスN5,プロセスN6,プロセスN7,プロセスN8}
(スレーブ装置リスト)
Node_A, 192.168.0.2, プロセス数(4), ネットワークA
Node_D, 192.168.0.5, プロセス数(2), ネットワークA
Node_F, 192.168.0.7, プロセス数(2), ネットワークA
このように、第2実施形態では、各スレーブ装置20のプロセッサコア数及び接続ネットワークに基づいて、分散処理に参加するスレーブ装置20及びそのスレーブ装置20で起動されるプロセッサの数が決定される。そして、初回選択された分割パターンに各プロセス25が対応付けられる。このような決定及び対応付けは、局所集団通信が、同じスレーブ装置20内のマルチコアでできる限り実行され、かつ、同一ネットワークに接続されるスレーブ装置20間でできる限り実行されるように実行される。
分散処理プログラム: /home/userA/sperceptron
起動プロセス数: 8
(分散処理プログラム:/home/userA/sperceptron)
1.w=0 //重みベクトル初期化
2.Data={(x1,y1),(x2,y2),・・・(xn,yn)}
//データを読み込む(xは特徴、yはラベル)
3.For i=1:Max //以下の(1)、(2)をMax回繰り返す
(1)Foreach (x,y) in Data
a)現在の推定が間違っていればw更新
(2)w=PseudoAllredcue(w,i,4)
//4台ずつのAllreduce(局所集団通信)
4.w=Allreduce(w)/n
SLV1 : プロセスN1
SLV2 : プロセスN2
SLV3 : プロセスN3
SLV4 : プロセスN4
SLV5 : プロセスN5
SLV6 : プロセスN6
SLV7 : プロセスN7
SLV8 : プロセスN8
上述の各実施形態では、局所集団通信は所定ステップ数繰り返されたが、局所集団通信の結果(局所処理データ)に応じて、局所集団通信の繰り返しが終了されるようにしてもよい。
Claims (15)
- 複数の分散処理装置により起動される複数のプロセスの中の少なくとも1つを起動する分散処理装置において、
前記少なくとも1つのプロセスが、
前記複数のプロセスが第1分割パターンで分割された複数のプロセスグループの中の1つのプロセスグループ内の複数のプロセス間で局所集団通信を行う集団通信手段と、
前記局所集団通信の後、前記局所集団通信において用いられた前記第1分割パターンとは異なる第2分割パターンを選択し、前記複数のプロセスが、該第2分割パターンで分割された、複数のプロセスグループの中の1つのプロセスグループ内の複数のプロセス間で、前記局所集団通信の結果に基づくデータを用いて次の局所集団通信を前記集団通信手段に実行させる通信制御手段と、
を備える分散処理装置。 - 前記局所集団通信の結果に基づくデータを評価する評価手段、
を更に備え、
前記通信制御手段は、前記評価手段の評価結果に応じて、前記集団通信手段に、前記次の局所集団通信を実行させるか否かを判定する、
請求項1に記載の分散処理装置。 - 前記局所集団通信の結果に基づくデータを評価する評価手段、
を更に備え、
前記通信制御手段は、前記評価手段の評価結果に基づいて、前記次の局所集団通信に用いられる分割パターンを選択する、
請求項1又は2に記載の分散処理装置。 - 前記通信制御手段は、各プロセスグループ内のプロセス数が前記第1分割パターンよりも多い分割パターンを前記第2分割パターンとして選択する、
請求項1から3のいずれか1項に記載の分散処理装置。 - 前記通信制御手段は、前記局所集団通信が所定回数繰り返された後、前記複数のプロセスが持つ複数の対象データが反映されたデータが前記複数のプロセスの各々でそれぞれ共有されるような、前記第1分割パターン及び前記第2分割パターンを含む複数の分割パターンの選択順序を保持し、該選択順序に基づいて、前記次の局所集団通信に用いられる分割パターンを選択する、
請求項1から4のいずれか1項に記載の分散処理装置。 - 請求項1から5のいずれか1項に記載の複数の分散処理装置と、
前記各分散処理装置についてプロセッサコア数及び接続ネットワーク情報の少なくとも一方をそれぞれ格納する情報格納部と、
前記接続ネットワークが共通の分散処理装置、及び、前記プロセッサコア数が多い分散処理装置の少なくとも一方から優先的に、前記複数のプロセスの少なくとも1つを起動させる分散処理装置を選択する実行対象決定部と、
前記実行対象決定部により選択された各分散処理装置に、前記プロセスの実行指示をそれぞれ送信する実行指示部と、
を備える分散処理管理装置と、
を含む分散処理システム。 - 請求項1から5のいずれか1項に記載の複数の分散処理装置と、
前記各分散処理装置についてプロセッサコア数及び接続ネットワーク情報をそれぞれ格納する情報格納部と、
接続ネットワーク毎のプロセッサコア数の合計が多くかつプロセッサコア数が多い分散処理装置から優先的に、前記複数のプロセスの少なくとも1つを起動させる分散処理装置を選択し、前記プロセッサコア数に応じて該分散処理装置に起動させるプロセス数を決定する対象決定部と、
前記実行対象決定部により選択された各分散処理装置に、前記実行対象決定部により決定されたプロセス数を含む、前記プロセスの実行指示をそれぞれ送信する実行指示部と、
を備える分散処理管理装置と、
を含む分散処理システム。 - 複数のコンピュータにより起動される複数のプロセスの中の少なくとも1つを起動するコンピュータで実行される分散処理方法において、
前記複数のプロセスが第1分割パターンで分割された複数のプロセスグループの中の1つのプロセスグループ内の複数のプロセス間で局所集団通信を行い、
前記局所集団通信の後、前記局所集団通信において用いられた前記第1分割パターンとは異なる第2分割パターンを選択し、
前記複数のプロセスが前記第2分割パターンで分割された複数のプロセスグループの中の1つのプロセスグループ内の複数のプロセス間で、前記局所集団通信の結果に基づくデータを用いて次の局所集団通信を行う、
ことを含む分散処理方法。 - 前記局所集団通信の結果に基づくデータを評価し、
前記評価結果に応じて、前記次の局所集団通信を実行させるか否かを判定する、
ことを更に含む請求項8に記載の分散処理方法。 - 前記局所集団通信の結果に基づくデータを評価する、
ことを更に含み、
前記第2分割パターンの選択は、前記評価結果に基づいて行われる、
請求項8又は9に記載の分散処理方法。 - 前記第2分割パターンの選択は、各プロセスグループ内のプロセス数が前記第1分割パターンよりも多い分割パターンを前記第2分割パターンとして選択する、
請求項8から10のいずれか1項に記載の分散処理方法。 - 前記第2分割パターンの選択は、前記局所集団通信が所定回数繰り返された後、前記複数のプロセスが持つ複数の対象データが反映されたデータが前記複数のプロセスの各々でそれぞれ共有されるような、前記第1分割パターン及び前記第2分割パターンを含む複数の分割パターンの選択順序に基づいて行われる、
請求項8から11のいずれか1項に記載の分散処理方法。 - 前記コンピュータ又は他のコンピュータが、
前記各コンピュータについてプロセッサコア数及び接続ネットワーク情報の少なくとも一方をそれぞれ格納する情報格納部を参照し、
前記接続ネットワークが共通のコンピュータ、及び、前記プロセッサコア数が多いコンピュータの少なくとも一方から優先的に、前記複数のプロセスの少なくとも1つを起動させるコンピュータとして選択し、
前記選択された各コンピュータに、前記プロセスの実行指示をそれぞれ送信する、
ことを更に含む請求項8から12のいずれか1項に記載の分散処理方法。 - 前記コンピュータ又は他のコンピュータが、
前記各コンピュータについてプロセッサコア数及び接続ネットワーク情報をそれぞれ格納する情報格納部を参照し、
接続ネットワーク毎のプロセッサコア数の合計が多くかつプロセッサコア数が多いコンピュータから優先的に、前記複数のプロセスの少なくとも1つを起動させるコンピュータとして選択し、
前記プロセッサコア数に応じて前記選択された各コンピュータに起動させるプロセス数をそれぞれ決定し、
前記選択された各コンピュータに、前記決定されたプロセス数を含む、前記プロセスの実行指示をそれぞれ送信する、
ことを更に含む請求項8から12のいずれか1項に記載の分散処理方法。 - 請求項8から14のいずれか1項に記載の分散処理方法をコンピュータに実行させるプログラム。
Priority Applications (3)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
JP2014528023A JP6197791B2 (ja) | 2012-07-30 | 2013-04-26 | 分散処理装置及び分散処理システム並びに分散処理方法 |
US14/417,589 US9807152B2 (en) | 2012-07-30 | 2013-04-26 | Distributed processing device and distributed processing system as well as distributed processing method |
EP13826186.2A EP2881862B1 (en) | 2012-07-30 | 2013-04-26 | Distributed processing device and distributed processing system as well as distributed processing method |
Applications Claiming Priority (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
JP2012168698 | 2012-07-30 | ||
JP2012-168698 | 2012-07-30 |
Publications (1)
Publication Number | Publication Date |
---|---|
WO2014020959A1 true WO2014020959A1 (ja) | 2014-02-06 |
Family
ID=50027653
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
PCT/JP2013/062479 WO2014020959A1 (ja) | 2012-07-30 | 2013-04-26 | 分散処理装置及び分散処理システム並びに分散処理方法 |
Country Status (4)
Country | Link |
---|---|
US (1) | US9807152B2 (ja) |
EP (1) | EP2881862B1 (ja) |
JP (1) | JP6197791B2 (ja) |
WO (1) | WO2014020959A1 (ja) |
Cited By (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106164863A (zh) * | 2014-04-03 | 2016-11-23 | 斯特拉托斯卡莱有限公司 | 虚拟中央处理单元的寄存器类型感知的调度 |
JP2017506786A (ja) * | 2014-02-11 | 2017-03-09 | シークオティエント インコーポレイテッド | 個別化されたデジタルコンテンツの自動生成のためのデジタルデータ処理方法および装置 |
JP6227813B1 (ja) * | 2017-01-25 | 2017-11-08 | 株式会社Preferred Networks | 分散深層学習装置及び分散深層学習システム |
JP2018120551A (ja) * | 2017-01-27 | 2018-08-02 | キヤノン株式会社 | データ処理装置、画像再構成装置、データ処理方法、及びプログラム |
US11018896B2 (en) | 2017-04-26 | 2021-05-25 | Fujitsu Limited | Information processing apparatus and information processing method |
Families Citing this family (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JP5850054B2 (ja) * | 2011-08-01 | 2016-02-03 | 日本電気株式会社 | 分散処理管理サーバ、分散システム、及び分散処理管理方法 |
JP2019067084A (ja) * | 2017-09-29 | 2019-04-25 | 富士通株式会社 | 情報処理システム、情報処理装置、及び、転送先決定方法 |
CN111526169B (zh) * | 2019-02-01 | 2022-06-14 | 阿里巴巴集团控股有限公司 | 通过网络发送数据的方法、介质、服务器和计算机设备 |
US11044174B2 (en) * | 2019-08-26 | 2021-06-22 | Citrix Systems, Inc. | Systems and methods for disabling services in a cluster |
CN112486638A (zh) * | 2019-09-11 | 2021-03-12 | 百度时代网络技术(北京)有限公司 | 用于执行处理任务的方法、装置、设备和存储介质 |
CN111723933B (zh) * | 2020-06-03 | 2024-04-16 | 上海商汤智能科技有限公司 | 神经网络模型的训练方法和相关产品 |
Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JPH11110362A (ja) * | 1997-10-01 | 1999-04-23 | Hitachi Ltd | 計算機間データ通信方法 |
JPH11134311A (ja) * | 1997-10-24 | 1999-05-21 | Nec Corp | マルチプロセッサシステムにおける集団通信装置 |
JPH11259441A (ja) * | 1998-03-13 | 1999-09-24 | Hitachi Ltd | 並列計算機における全対全通信方法 |
JP2009193255A (ja) * | 2008-02-13 | 2009-08-27 | Fujitsu Ltd | 集団通信最適化プログラム、集団通信最適化装置および集団通信最適化方法 |
JP2011053876A (ja) * | 2009-09-01 | 2011-03-17 | Fujitsu Ltd | 並列計算システムおよび通信制御プログラム |
JP2012098881A (ja) * | 2010-11-01 | 2012-05-24 | Fujitsu Ltd | 情報処理方法、装置及びプログラム |
-
2013
- 2013-04-26 US US14/417,589 patent/US9807152B2/en active Active
- 2013-04-26 EP EP13826186.2A patent/EP2881862B1/en active Active
- 2013-04-26 WO PCT/JP2013/062479 patent/WO2014020959A1/ja active Application Filing
- 2013-04-26 JP JP2014528023A patent/JP6197791B2/ja active Active
Patent Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JPH11110362A (ja) * | 1997-10-01 | 1999-04-23 | Hitachi Ltd | 計算機間データ通信方法 |
JPH11134311A (ja) * | 1997-10-24 | 1999-05-21 | Nec Corp | マルチプロセッサシステムにおける集団通信装置 |
JPH11259441A (ja) * | 1998-03-13 | 1999-09-24 | Hitachi Ltd | 並列計算機における全対全通信方法 |
JP2009193255A (ja) * | 2008-02-13 | 2009-08-27 | Fujitsu Ltd | 集団通信最適化プログラム、集団通信最適化装置および集団通信最適化方法 |
JP2011053876A (ja) * | 2009-09-01 | 2011-03-17 | Fujitsu Ltd | 並列計算システムおよび通信制御プログラム |
JP2012098881A (ja) * | 2010-11-01 | 2012-05-24 | Fujitsu Ltd | 情報処理方法、装置及びプログラム |
Non-Patent Citations (5)
Title |
---|
GIDEON MANN; RYAN MCDONALD; MEHRYAR MOHRI: "Efficient Large-Scale Distributed Training of Conditional Maxmum Entropy Models", NIPS, 2009 |
JEFFREY DEAN; SANJAY GHEMAWAT: "MapReduce: Simplified Data Processing on Large Clusters", USENIX ASSOCIATION, 6TH SYMPOSIUM ON OPERATING SYSTEMS DESIGN AND IMPLEMENTATION, 2004 |
KEITH B. HALL; SCOTT GILPIN; GIDEON MANN: "MapReduce/Big table for Distributed Optimization", NIPS LCCC WORKSHOP, 2010 |
RYAN MCDONALD; KEITH HALL; GIDEON MANN: "Distributed Training Strategies for the Structured Perceptron", HUMAN LANGUAGE TECHNOLOGIES, THE 2010 ANNUAL CONFERENCE OF THE NORTH AMERICAN CHAPTER OF THE ACL, June 2010 (2010-06-01), pages 456 - 464, XP058270354 |
See also references of EP2881862A4 |
Cited By (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JP2017506786A (ja) * | 2014-02-11 | 2017-03-09 | シークオティエント インコーポレイテッド | 個別化されたデジタルコンテンツの自動生成のためのデジタルデータ処理方法および装置 |
US10559015B2 (en) | 2014-02-11 | 2020-02-11 | Salesforce.Com, Inc. | Digital data processing methods and apparatus for the automated generation of personalized digital content |
CN106164863A (zh) * | 2014-04-03 | 2016-11-23 | 斯特拉托斯卡莱有限公司 | 虚拟中央处理单元的寄存器类型感知的调度 |
JP6227813B1 (ja) * | 2017-01-25 | 2017-11-08 | 株式会社Preferred Networks | 分散深層学習装置及び分散深層学習システム |
JP2018120551A (ja) * | 2017-01-27 | 2018-08-02 | キヤノン株式会社 | データ処理装置、画像再構成装置、データ処理方法、及びプログラム |
US11018896B2 (en) | 2017-04-26 | 2021-05-25 | Fujitsu Limited | Information processing apparatus and information processing method |
Also Published As
Publication number | Publication date |
---|---|
US20150215379A1 (en) | 2015-07-30 |
JP6197791B2 (ja) | 2017-09-20 |
EP2881862A4 (en) | 2016-03-16 |
US9807152B2 (en) | 2017-10-31 |
EP2881862A1 (en) | 2015-06-10 |
EP2881862B1 (en) | 2018-09-26 |
JPWO2014020959A1 (ja) | 2016-07-21 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
JP6197791B2 (ja) | 分散処理装置及び分散処理システム並びに分散処理方法 | |
Liu et al. | Adaptive asynchronous federated learning in resource-constrained edge computing | |
US10764125B2 (en) | Method and device for training model in distributed system | |
US10572290B2 (en) | Method and apparatus for allocating a physical resource to a virtual machine | |
US20140130059A1 (en) | Lattice Computing | |
WO2013080152A1 (en) | Dynamically configurable placement engine | |
WO2011084296A2 (en) | A virtualization and consolidation analysis engine for enterprise data centers | |
CN110308984B (zh) | 一种用于处理地理分布式数据的跨集群计算*** | |
US10108463B2 (en) | System, method of controlling to execute a job, and apparatus | |
WO2016145904A1 (zh) | 一种资源管理方法、装置和*** | |
US9141677B2 (en) | Apparatus and method for arranging query | |
JP6972796B2 (ja) | ソフトウェアサービス実行装置、システム、及び方法 | |
Garala et al. | A performance analysis of load Balancing algorithms in Cloud environment | |
WO2016092856A1 (ja) | 情報処理装置、情報処理システム、タスク処理方法、及び、プログラムを記憶する記憶媒体 | |
US8028291B2 (en) | Method and computer program product for job selection and resource allocation of a massively parallel processor | |
JP2016024612A (ja) | データ処理制御方法、データ処理制御プログラムおよびデータ処理制御装置 | |
US20150365474A1 (en) | Computer-readable recording medium, task assignment method, and task assignment apparatus | |
JP5969315B2 (ja) | データ移行処理システムおよびデータ移行処理方法 | |
Shabeera et al. | Optimising virtual machine allocation in MapReduce cloud for improved data locality | |
US20210390405A1 (en) | Microservice-based training systems in heterogeneous graphic processor unit (gpu) cluster and operating method thereof | |
US10592473B2 (en) | Method for improving energy efficiency of map-reduce system and apparatus thereof | |
Meddeber et al. | Tasks assignment for Grid computing | |
CN108984105B (zh) | 对网络存储设备中的复制任务进行分配的方法和设备 | |
Nzanywayingoma et al. | Task scheduling and virtual resource optimising in Hadoop YARN-based cloud computing environment | |
KR101654969B1 (ko) | 가상화 클러스터 환경에서 네임 노드를 할당하는 방법 및 장치 |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
121 | Ep: the epo has been informed by wipo that ep was designated in this application |
Ref document number: 13826186 Country of ref document: EP Kind code of ref document: A1 |
|
ENP | Entry into the national phase |
Ref document number: 2014528023 Country of ref document: JP Kind code of ref document: A |
|
WWE | Wipo information: entry into national phase |
Ref document number: 2013826186 Country of ref document: EP |
|
WWE | Wipo information: entry into national phase |
Ref document number: 14417589 Country of ref document: US |
|
NENP | Non-entry into the national phase |
Ref country code: DE |