JP2016218926A - Information processor, information processing method, and computer program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide an information processor for distributing data processing to a plurality of arithmetic devices on a network to cause the plurality of arithmetic devices to perform the data processing while taking transfer times of the data on the network into consideration.SOLUTION: The information processor for distributing data processing to a plurality of arithmetic devices on a network to cause the plurality of arithmetic devices to perform the data processing includes: first acquisition means for acquiring first information on a time of data transfer to each of the plurality of arithmetic devices; second acquisition means for acquiring second information on the respective arithmetic capabilities of the plurality of arithmetic devices; and selection means for selecting an arithmetic device caused to perform the data processing among the plurality of arithmetic devices on the basis of the first information and the second information such that the completion of the data processing becomes smaller than a prescribed value.SELECTED DRAWING: Figure 8

Description

  The present invention relates to an information processing apparatus, an information processing method, and a computer program that use a computing device on a network.

  Parallel computing technology using heterogeneous computing resources such as CPUs, GPUs, and cells improves processor computing power and integrates a number of cores specialized for specific operations represented by cells and GPUs. It has been adopted for various products. Conventionally, in the coding of software in such a heterogeneous computing environment, it has been necessary to implement an original implementation specialized for the processor group to be used.

  As a standard framework for parallel programming in a heterogeneous computing environment, there is OpenCL (Open Computing Language). OpenCL defines a model that abstracts parallel computing devices, memories, and the like, and by using the OpenCL framework, it is possible to support a plurality of processing environments with the same code. This specification is currently being developed and published by the standards organization Kronos Group.

  On the other hand, cloud computing that uses computer processing via a network such as the Internet is gradually expanding as a service. Even in cloud computing, modeling the computing environment is considered to enable both the construction of an application that does not depend on a specific computing environment and the speeding up of the computation.

  Japanese Patent Application Laid-Open No. 2004-151867 discloses a technique for using OpenCL as an interface to cloud computing to speed up application processing. According to this technique, when an application processing request occurs, it is possible to select a device on the cloud based on the characteristics of the computing resource on the cloud, perform distributed processing, and reduce the processing load. The computing resources are specifically a processor and a memory on the cloud.

JP 2011-138506 A

  In order to process operations such as image processing on a network such as a cloud using the technology disclosed in Patent Document 1, it is necessary to transmit and receive data necessary for the processing between operation devices on the cloud. . When this data is large, the data transfer time cannot be ignored with respect to the calculation execution time. For example, it is assumed that there are two computing devices A and B having different computing capabilities on the cloud, and the time for executing an application is 1 second for the computing device A and 2 seconds for the computing device B. Since the data transfer time to the computing device depends on the network environment up to the computing device, the transfer speed is not the same. When the data transfer rate to the computing device A with high computing capability is lower than the data transfer rate to the computing device B with low computing capability, selecting the computing device with the computing capability necessarily shortens the application execution time. It will not be. Assume that the time required to transfer the data necessary for executing the above-described application to the computing device A is 4 seconds, and the time required to transfer to the computing device B is 2 seconds. In this case, the execution time of the application, which is the sum of the calculation execution time and the data transfer time, is 5 seconds for the calculation device A and 4 seconds for the calculation device B, and the execution time in the calculation device B is shorter.

  On the other hand, there are already many OpenCL applications created for the OpenCL environment in devices such as GPUs, rather than intending to use computing devices on the cloud. Therefore, it is desirable that the application can be executed using a computing device on the cloud without changing the existing OpenCL application.

  In OpenCL, an ID that indirectly points to a physical device called a device ID is used as a method for specifying a device that performs an operation. When OpenCL is used as an interface to cloud computing, a device ID is assigned to each node on the cloud. Therefore, device IDs overlap between nodes, and cannot be uniquely identified as they are. Also, in-app applications often do not assume that there are multiple devices as in a cloud environment. For example, the application is created using a fixed device ID such as the device ID of the first identified device. Therefore, in order to select and execute an appropriate device from a plurality of devices on the cloud, the application side must be changed to use the device.

  The present invention has been made in view of the above problems, and provides an information processing apparatus that performs processing of data on a plurality of arithmetic devices on the network in consideration of the transfer time of the data on the network. The purpose is to do.

  In order to solve the above-described problem, the present invention provides an information processing apparatus that distributes data processing to a plurality of arithmetic devices on a network, and relates to a data transfer time to each of the plurality of arithmetic devices. First acquisition means for acquiring one information, second acquisition means for acquiring second information relating to the calculation capability of each of the plurality of calculation devices, the first information and the second information, And selecting means for selecting an arithmetic device for processing the data from the plurality of arithmetic devices so that the completion of the processing of the data becomes smaller than a predetermined value.

  The present invention has been made in view of the above problems, and considers the transfer time of data on the network and performs processing of the data distributed to a plurality of arithmetic devices on the network. Can be provided.

It is a block diagram which shows the node structure of the system which uses OpenCL as an interface for utilizing the calculation resource which exists on a cloud. It is a block diagram which shows the structure of the cloud corresponding OpenCL driver in 1st Embodiment. It is a flowchart which shows the procedure which acquires the resource information and data transfer time of the arithmetic device concerning 1st Embodiment. It is a figure showing an example of the arithmetic device information managed by the arithmetic device information management part concerning 1st Embodiment. It is a flowchart which shows the procedure which selects the arithmetic device suitable for performing the OpenCL application in the arithmetic device selection part concerning 1st Embodiment. It is a figure showing an example of the arithmetic device information after estimation of the total time of the data transfer time and calculation execution time concerning 1st Embodiment. It is a figure showing an example of the arithmetic device information after the node concerning 1st Embodiment was added on the network. It is a block diagram which shows another structure of the OpenCL driver in 2nd Embodiment. It is a flowchart which shows the procedure which performs acquisition of the resource information and data transmission time of a computing device, and device ID assignment concerning 2nd Embodiment. It is a figure showing an example of the arithmetic device information managed by the arithmetic device information management part concerning 2nd Embodiment. It is a figure showing an example of the arithmetic device information after determining the arithmetic device which performs the OpenCL application concerning 2nd Embodiment. It is a figure showing another example of the arithmetic device information after determining the arithmetic device which performs the OpenCL application concerning 2nd Embodiment.

(First embodiment)
The first embodiment will be described below with reference to the drawings and flowcharts. However, the technical scope of the present invention is not limited to this embodiment.

  FIG. 1 is a block diagram illustrating a configuration of a node (information processing apparatus) of a system that uses OpenCL as an interface for using a computation resource existing on the cloud. Each node (information processing apparatus) has a CPU, a recording medium storing various computer programs, and the like, and executes the various computer programs.

  In the figure, reference numeral 101 denotes an OpenCL application execution node, which is a node on the network that executes the OpenCL application.

  Reference numerals 102 to 104 denote OpenCL interface-compatible arithmetic device nodes, which execute OpenCL interface requests required over the network and return the results. In the figure, three nodes are shown as an example, but the number is not limited to this.

  A network 105 is responsible for a communication function between an OpenCL application execution node and an OpenCL interface-compatible arithmetic device node.

  Reference numeral 1011 denotes an OpenCL application, which is an application for parallel arithmetic processing such as image processing. This is equivalent to an application that operates in an OpenCL execution environment closed in a conventional device.

  Reference numeral 1012 denotes a cloud-compatible OpenCL driver that has been changed to use a computing device on the cloud as a computing resource. The API format provided to the OpenCL application is equivalent to the conventional in-device OpenCL environment. Details of the function will be described later.

  A network interface 1013 mediates a communication function between a cloud-compatible OpenCL driver and an OpenCL interface-compatible arithmetic device node on the network.

  Reference numeral 1021 denotes a network interface that mediates communication between the cloud-compatible OpenCL driver on the network and the OpenCL driver of the own node.

  An OpenCL driver 1022 provides an API that conforms to the OpenCL interface to the OpenCL application. It is the same as that in the conventional OpenCL environment in the device.

  An arithmetic device 1023 is a device that executes an arithmetic function based on the OpenCL interface, and is equivalent to that in the conventional in-device OpenCL environment.

  Reference numerals 1011, 1022, and 1023 are equivalent to those in the conventional in-device OpenCL environment, and the features of the present invention are summarized in a configuration in the 1012 cloud-compatible OpenCL driver.

  FIG. 2 is a block diagram showing a configuration of an embodiment of a cloud-compatible OpenCL driver in the present invention.

  In the figure, the OpenCL application 1011 and the network interface 1013 are the same as the blocks with the same symbols in FIG.

  A configuration in the 1012 cloud-compatible OpenCL driver will be described.

  An OpenCL API conversion unit 10121 provides an API equivalent to an OpenCL driver on a conventional in-device OpenCL environment to an OpenCL application. Actual processing such as calculation execution and resource information acquisition is performed by virtualizing the functions defined in the OpenCL interface by using the function of each block described later and the calculation function of the node on the network. It takes on the function of making it appear to be.

  Reference numeral 10122 denotes a resource information acquisition unit that acquires operation resource information of an operation device on a node connected by a network. Information related to computing power such as the type of processors such as GPU and CPU, the number of computing units, the frequency of clocks, and the size of memory is acquired.

  A data transfer rate acquisition unit 10123 acquires a data transfer rate between this node and a node connected to the network.

  Reference numeral 10124 denotes an arithmetic device selection unit which selects an arithmetic device that executes an arithmetic request of the OpenCL application.

  Reference numeral 10125 denotes an arithmetic device information management unit that manages information on the arithmetic devices acquired and determined by the processing units 10122 to 10124, and provides information according to requests from the OpenCL API conversion unit and the respective processing units. To do.

  FIG. 3 is a flowchart showing a procedure for acquiring resource information and data transfer time of the computing device of the node connected to the network in cooperation with the resource acquisition unit and the data transfer time acquisition unit.

  The timing for performing this procedure may be when the OpenCL application execution node is connected to the network, or when a node having a computing device is connected to the network.

  In step S301, a node that has not acquired resource information is searched for among nodes having an OpenCL interface-compatible arithmetic device on the network. If there is no node that has not acquired resource information, the process ends. If there is a node for which resource information has not been acquired, the process proceeds to step S302, and the node is selected as the next processing target.

  In step S302, the resource information of the computing device on the selected node is acquired. In the OpenCL interface, APIs for acquiring usable platform information and device information on the platform are defined. The resource information acquisition unit 10122 accesses the OpenCL driver 1022 over the network via the network interfaces 1013 and 1021, and acquires calculation resource information using this API.

  In step S303, the data transfer rate acquisition unit 10123 acquires the transfer rate of data between the selected node and the OpenCL application execution node.

  After step S303, the process returns to step S301. Thereafter, step S301 to step S303 are repeated by the number of nodes having an OpenCL interface-compatible arithmetic device existing on the network.

  As a result of the above steps, the information on the computing device acquired by the resource acquisition unit and the data transfer time acquisition unit is collected in the computing device information management unit, and a data list that associates the resource information and the data transfer rate is generated. FIG. 4 is an example of computing device information managed by the computing device information management unit. A specific example of the above steps will be described with reference to this figure.

  In step S301, a node having an OpenCL interface-compatible arithmetic device existing on the network is searched, and a node having an IP address of 10.20.30.40 corresponds. As this node identification information, an IP address is recorded as node identification information in a node identification information string 401.

  In step S302, resource information of this node is acquired. In the OpenCL interface, platform information is first acquired. In OpenCL, it is possible to have a plurality of OpenCL execution environments on one node, and information on each platform can be acquired. Each platform can be identified by a platform ID. As platform information, information that two platforms exist in this node is obtained, and the platform IDs are 0 and 1, respectively. This information is recorded in the acquired OpenCL platform ID column 402.

  Next, computing device information is acquired for each platform. In OpenCL, it is possible to have a plurality of computing devices on one platform, and information on each computing device can be acquired. Each computing device can be identified by a device ID. As the computing device information, information that there are two computing devices in the platform with a platform ID of 0 is obtained, and the device IDs are 0 and 1. Further, it is assumed that information indicating that one computing device exists in the platform having the platform ID 1 and the device ID is 0. These pieces of information are recorded in the acquired OpenCL device ID column 403.

  Next, the computing capability information of the computing device is acquired for each computing device. With OpenCL, it is possible to acquire various types of information about each computing device. Here, information necessary for comparing the computing capabilities of the devices is acquired. For example, information such as the type of a processor such as a GPU or a CPU, the number of arithmetic units, the frequency of a clock, the size of a memory, and a device name. A case where processor type information is acquired as an example of computing capacity information is taken as an example. Assume that information called GPU is obtained as processor type information for a device with a platform ID of 0 and a device ID of 0. Further, it is assumed that information of GPU as the processor type information of the device having the platform ID 0 and device ID 1 and CPU as the processor type information of the device having the platform ID 1 and device ID 0 is obtained. These pieces of information are recorded in the calculation capability information sequence 404. The above is the acquisition of the resource information performed in step S302.

  In step S303, the data transfer rate with this node is acquired. As a method for measuring the transfer rate, a method of determining a data transfer rate by generating a temporary OpenCL application with large-capacity data input / output and executing it on a selected node can be considered. Further, the data transfer rate of the inter-node network may be measured by a method that does not depend on the OpenCL interface. It is assumed that information AA is obtained as a data transfer rate with a device having a platform ID of 0 and a device ID of 0. Further, it is assumed that information BB is obtained as the data transfer rate with the device having the platform ID 0 and device ID 1 and CC is obtained as the data transfer rate with the device having the platform ID 1 and device ID 0. The transfer rate obtained here is recorded in the data transfer rate column 405.

  This completes the processing for one node, and the same processing is performed for the next node. The resource information and data transfer rate of the IP address 10.20.40.50 are obtained and recorded in the columns 401 to 405.

  FIG. 5 is a flowchart showing a procedure by which the computing device selection unit 10124 selects a computing device suitable for executing the OpenCL application using the computing device information.

  As a trigger for performing this procedure, it is conceivable that this procedure is performed when the OpenCL application is first executed. As another opportunity, in order to predetermine a computing device suitable for an OpenCL application that may be executed, it may be performed when the system is started up.

  In step S501, data amount information and calculation information transferred from the OpenCL application to be executed are acquired. These pieces of information are information necessary for obtaining the data transfer time and the computation execution time for each computing device in step S503 described later. As the data amount information to be transferred, the data amount input / output to / from the OpenCL application and the data amount of the OpenCL application program itself are acquired.

  Steps S <b> 502 to S <b> 505 are processing that is repeated as many times as the number of previously obtained computing devices. In step S502, device information that has not been subjected to the computing device evaluation processing in steps S503 to S505 is searched for, and if there is no device information, the processing ends. If there is a device that has not been evaluated, the computing device is selected as a processing target, and the process proceeds to step S503.

  In step S503, the total time of the data transfer time and the calculation execution time when the OpenCL application is executed by the selected calculation device is estimated. Here, the data transfer speed and calculation capability information of the device are acquired from the calculation device information management unit 10125, and the data transfer time and calculation execution time are obtained using the data amount information and calculation information transferred by the application acquired in step S501. Estimate the total time of. The total time information is recorded as computing device information managed by the computing device information management unit.

  The data transfer time can be calculated from the data amount information and the data transfer rate.

  Regarding calculation execution time, there is a method of preparing a calculation amount versus calculation execution time conversion table for devices such as GPU and CPU in advance and obtaining the calculation execution time from the calculation amount of the application obtained in step S501. As another method, there is also a method in which the OpenCL application is actually executed by the arithmetic device to obtain the arithmetic execution time.

  In step S504, if the total time estimated in step S503 is smaller than the estimated total time in other computing devices, the process proceeds to step S505. If the total time is not smaller than that of the other device, the process returns to step S502 and proceeds to the evaluation step for the other device.

  In step S505, the device is selected as an OpenCL application execution device. Specifically, information indicating that the device is selected as a device that executes the OpenCL application is transmitted to the arithmetic device information management unit, and the arithmetic device information management unit records the information.

  Thereafter, by repeating steps S502 to S505 as many as the number of computing devices, a device whose total time is reduced is searched, and a computing device that executes the OpenCL application is determined.

  FIG. 6 shows an example of computing device information in which the total time of the data transfer time and computation execution time of a specific OpenCL application is added as a result of executing this procedure. Information on each computing device is added to a column of the total of the data transfer time and the computation execution time of a specific application 606. In this example, since the total time of the device having the node identification information of 10.20.30.40, the platform ID of 0, and the device ID of 1 is small, this device is selected as the computing device that executes the application.

  An example of the operation when the OpenCL application is executed in a state where the arithmetic device information management unit 10125 manages the arithmetic device information as shown in FIG. 6 will be described.

  The OpenCL application requests the OpenCL API conversion unit 10121 to execute arithmetic processing. The OpenCL API conversion unit inquires of the arithmetic device information management unit about device information in order to determine an actual device that performs the operation from the arithmetic devices present on the network. The computing device information management unit performs OpenCL API conversion on the information of the device whose node identification information is 10.20.40.40, whose platform ID is 0, and whose device ID is 0. Pass to the department. The OpenCL API conversion unit causes the corresponding device on this node to execute an operation.

  Next, an example in which a device that executes an OpenCL application is changed by adding a node having an OpenCL interface-compatible computing device to the network will be described.

  The cloud-compatible OpenCL driver selects the resource information described in FIG. 3 and the procedure for obtaining the data transfer time and the arithmetic device described in FIG. 5 for a node having a newly connected OpenCL interface-compatible arithmetic device. Perform the procedure. As a result, it is assumed that the computing device information managed by the computing device information management unit is as shown in FIG. Assume that the newly added node has node identification information of 50.40.30.20, two platforms, and one device each. Considering the devices on the node acquired earlier, the computing device with the minimum total data transfer time and computation execution time (time to complete the process) has node identification information of 50.40.30.20, platform The device has an ID of 0 and a device ID of 0. Therefore, this device is selected as the device that executes the OpenCL application, and the calculation is executed. In the present embodiment, the device that minimizes the time until completion of processing is selected. However, it is sufficient that the device is smaller than a predetermined value set in advance, and it is not always necessary to minimize the device.

  According to the arithmetic processing system in the present embodiment described above, it is possible to select an optimal arithmetic device for executing the OpenCL application from the arithmetic devices existing on the network in consideration of the data transfer time.

(Second Embodiment)
Hereinafter, the second embodiment will be described focusing on differences from the first embodiment. However, the technical scope of the present invention is not limited to this embodiment.

  In the present embodiment, an example will be described in which a unique virtual ID is assigned to an OpenCL interface-compatible computing device on a network, and an optimal computing device is provided to the OpenCL application.

  The configuration of the cloud-compatible OpenCL driver in this embodiment is shown in FIG. A device ID assignment unit 10126 is added to the block shown in FIG. The other blocks are the same as the blocks with the same symbols in FIG.

  A device ID assignment unit 10126 assigns a unique ID to a computing device on a node connected to the network.

  A flowchart showing a procedure in which a resource acquisition unit, a data transfer time acquisition unit, and a device ID allocation unit cooperate to acquire resource information and data transfer time of a computing device of a node connected to the network and assign a unique device ID. Is shown in FIG.

  Steps S301 to S303 are equivalent to the steps having the same reference numerals described in FIG. After step S303, the process proceeds to step S904.

  In step S904, the device ID assignment unit assigns a virtual device ID to the computing device existing in the node. In the OpenCL interface, a device ID that can be uniquely specified in the node is assigned to the computing device acquired in the resource information acquisition step. However, the device ID cannot be uniquely specified between nodes on the network. . Therefore, it is usually uniquely identified by a combination with information for identifying a node such as an IP address. In this step, a virtual device ID that is unique among all the computing devices existing on the network is assigned. As a result, the OpenCL application does not need to know the information of the node where the computing device exists, and can access the device in the same manner as the conventional in-device OpenCL execution environment.

  After step S904, the process returns to step S301. Thereafter, step S301 to step S904 are repeated by the number of nodes having an OpenCL interface-compatible arithmetic device existing on the network.

  As a result of the above steps, the information regarding the computing device acquired and determined by each processing unit is collected in the computing device information management unit, and a data list that associates resource information, data transfer rate, and virtual device ID is generated. FIG. 10 is an example of computing device information managed by the computing device information management unit. A specific example of device ID assignment will be described with reference to FIG.

  The information in columns 401 to 405 is equivalent to the information with the same sign described in FIG.

  In step S904 for the node having node identification information of 10.20.30.40, a virtual device ID that is unique across the nodes is assigned to the computing device existing in that node. Assume that the virtual device ID of a device having a platform ID of 0 and a device ID of 0 is set to 0. Similarly, assume that the virtual device ID of the device with the platform ID 0 and the device ID 1 is set to 1, and the virtual device ID of the device with the platform ID 1 and the device ID 0 is set to 2. The virtual device ID obtained here is recorded in the virtual device ID column 1006 assigned.

  This completes the processing for one node, and the same processing is performed for the other nodes. Regarding the virtual device IDs of the nodes having node identification information of 10.20.40.50 and 50.40.30.20, IDs that are unique among the nodes are assigned and recorded in the column 1006. In this example, numerical values in ascending order from 0 are assigned as virtual device IDs, but any unique value that can be identified may be used.

  Next, FIG. 11 shows an example of computing device information in which the computing device suitable for execution described in FIG. 5 is selected for a plurality of OpenCL applications, and the computing device that executes the OpenCL application is added. .

  Assume that the calculation device selection procedure is executed for the application X, the application Y, and the application Z of the OpenCL application. Assume that a device with a short total time of data transfer time and calculation execution time of the application X is a device having node identification information of 10.20.30.40, a platform ID of 0, and a device ID of 1. Assume that the devices with a small total time of app Y and app Z are both devices with node identification information of 50.40.20.20, platform ID of 0, and device ID of 0. In this case, the application information is recorded in the corresponding device in the OpenCL application sequence executed by the device 1107.

  An example of operation when an OpenCL application developed for an existing in-device OpenCL execution environment is executed in a state where the arithmetic device information management unit 10125 manages the arithmetic device information as shown in FIG. .

  The OpenCL application acquires device information by using an API that acquires device information usable on the platform in order to determine a device to be executed before executing the arithmetic processing. When the OpenCL application calls the API to the OpenCL API conversion unit 10121, the OpenCL API conversion unit acquires device information from the arithmetic device information management unit 10125 and provides it to the application. At this time, a virtual device ID is provided as information corresponding to the OpenCL device ID. When the computing device shown in FIG. 11 exists on the network, six pieces of device information are provided to the application.

  In OpenCL, the application side basically selects on which device the application is executed. Specifically, the application selects one suitable device from the acquired device ID group, and requests the OpenCL driver to execute the operation by specifying the device ID.

  Here, the OpenCL application developed for the in-device OpenCL execution environment often does not assume that there are a plurality of devices. For example, it is assumed that the device ID of the device identified first is selected. . In this example, it is assumed that a device with a virtual device ID of 0 is selected.

  The OpenCL API conversion unit requested to execute the operation with the device ID set to 0 makes an inquiry to the operation device information management unit together with the application information in order to determine a real device existing on the network from the virtual device ID.

  As a first example, it is assumed that the application at this time is other than the applications X, Y, and Z. The computing device information management unit searches whether the OpenCL application in column 1107 in FIG. 11 matches the inquired application. In this case, since there is no match, the node identification information of the device whose virtual device ID is 0 is 10.20.40.40, the information of the device whose platform ID is 0, and device ID is 0 is passed to the OpenCL API conversion unit. The OpenCL API conversion unit causes the corresponding device on this node to execute an operation. In this case, since the computing device suitable for execution has not been specified, the processing is executed as it is with the device designated by the application.

  As a different example, it is assumed that the application is an application X. The computing device information management unit searches whether the OpenCL application in column 1107 in FIG. 11 matches the inquired application, and in this case, it is found that they match. Since it is better to execute the application X on the device with the virtual device ID 1, the information of the device with the node identification information 10.20.40.40, the platform ID 0, and the device ID 1 is sent to the OpenCL API conversion unit. hand over. At the same time, the virtual device ID managed internally is replaced. The computing device information after the replacement is as shown in FIG. The OpenCL API conversion unit causes the corresponding device on this node to execute the calculation, and an optimal calculation device is preferentially set for the application X, and the process can be executed.

  According to the arithmetic processing system in the present embodiment described above, an optimal device can be selected from the arithmetic devices existing on the network in consideration of the data transfer time. Further, it is possible to select and use an optimal OpenCL device on the network without changing the OpenCL application developed as a conventional in-appliance application.

(Other examples)
The present invention supplies a program that realizes one or more functions of the above-described embodiments to a system or apparatus via a network or a storage medium, and one or more processors in the computer of the system or apparatus read and execute the program This process can be realized. It can also be realized by a circuit (for example, ASIC) that realizes one or more functions.

101 OpenCL application execution node 102, 103, 104 OpenCL interface compatible computing device node 105 network 1011 OpenCL application 1012 cloud compatible OpenCL driver 1013, 1021 network interface 1022 OpenCL driver 1023 computing device 10121 OpenCL API conversion unit 10122 resource information acquisition unit 10123 data transfer Speed acquisition unit 10124 Arithmetic device selection unit 10125 Arithmetic device information management unit 10126 Device ID allocation unit

Claims (6)

An information processing apparatus that distributes data processing to a plurality of arithmetic devices on a network,
First acquisition means for acquiring first information on data transfer time to each of the plurality of computing devices;
Second acquisition means for acquiring second information regarding the calculation capability of each of the plurality of calculation devices;
Based on the first information and the second information, a computing device that performs the data processing is selected from the plurality of computing devices such that completion of the data processing is smaller than a predetermined value. And an information processing apparatus. The data processing is an OpenCL application processing,
2. The information processing apparatus according to claim 1, wherein the second acquisition unit acquires second information related to a calculation capability of each of the plurality of calculation devices using an OpenCL interface. Assigning means for assigning a unique virtual device ID on the network to each of the plurality of computing devices;
3. The management apparatus according to claim 2, further comprising management means for managing the virtual device ID and the OpenCL device ID in association with each other and providing the virtual device ID as the OpenCL device ID to the OpenCL application. Information processing device.   3. A setting unit that prioritizes selection by the selection unit when the calculation device specified at the time of execution of the OpenCL application is different from the calculation device selected by the selection unit. 3. The information processing apparatus according to 3. An information processing apparatus that distributes data processing to a plurality of arithmetic devices on a network,
A first acquisition step in which a first acquisition means acquires first information relating to a data transfer time to each of the plurality of computing devices;
A second acquisition step in which a second acquisition means acquires second information on the calculation capability of each of the plurality of calculation devices;
An operation for causing the selection means to process the data based on the first information and the second information so that the completion of the processing of the data is smaller than a predetermined value. And a selection step of selecting a device. Computer
An information processing apparatus that distributes data processing to a plurality of arithmetic devices on a network,
First acquisition means for acquiring first information on data transfer time to each of the plurality of computing devices;
Second acquisition means for acquiring second information regarding the calculation capability of each of the plurality of calculation devices;
Based on the first information and the second information, a computing device that performs the data processing is selected from the plurality of computing devices such that completion of the data processing is smaller than a predetermined value. A computer program for causing the information processing apparatus to function.

Images