PROCESSING PROCESS

Abstract

A processing procedure comprises the following steps: obtaining job information about a job related to machine learning that is running on one of a plurality of nodes; identifying another node from the plurality of nodes based on the job information that will allow the remaining part of the job to be completed within a learning timeframe; comparing the first cost, which is the cost in a case where the job continues on the one node, with the second cost, which is the cost in a case where the remaining part of the job is executed on the other node; and transferring the job from one node to the other node in a case where the second cost is lower than the first cost.

Description

TECHNICAL AREA

[0001] Exemplary embodiments of the present disclosure relate to technical areas of a processing method and in particular to a processing method that processes a job in relation to machine learning. BACKGROUND

[0002] Services that utilize learned / trained models (i.e., AI (Artificial Intelligence)) generated through machine learning have been proposed. For example, patent literature 1, JP2022-034850A, describes a service that provides information to support safe driving based on the output of a learned model, where the types and installation environments of signs or markings around a vehicle, the vehicle's driving status, the vehicle's position, and the driver's line of sight are input into the learned model.

[0003] In the technique / technology described in patent literature 1, the learned model is used in an in-vehicle device, but the learned model can also be used in a data center with higher processing capacity than that of the in-vehicle device. For example, a service that provides the information described in patent literature 1 to support safe driving requires the real-time processing of a relatively small amount of data from the vehicle. On the other hand, machine learning to create a learning model requires the processing of a large amount of data. That is, the required performance of the data center to be used for providing the aforementioned service differs from the required performance of the data center to be used for performing the machine learning.Furthermore, the costs of developing a learning model often vary depending on the data center used for its development (in other words, where the machine learning is performed). If no action is taken, the costs of developing learning models can increase, which is technically problematic. SUMMARY

[0004] In view of the problems described above, one objective of the present disclosure is to provide a processing method that allows a data center to be selected in such a way as to reduce / control the costs of developing a learning model.

[0005] A processing procedure according to one aspect of the present disclosure comprises the following steps: obtaining job information about a machine learning job being executed on one of several nodes; identifying another of the several nodes based on the job information that will allow the remaining part of the job to be completed within a learning period; comparing the first cost, which is the cost in a case where the job continues on the one node, with the second cost, which is the cost in a case where the remaining part of the job is executed on the other node; and transferring the job from the one node to the other node in a case where the second cost is lower than the first cost. BRIEF DESCRIPTION OF THE DRAWINGS Fig. Figure 1 is a conceptual diagram showing a system according to an exemplary embodiment; Fig. Figure 2 is a block diagram showing a configuration of an information processing device according to the exemplary embodiment; Fig. Figure 3 is a diagram illustrating an example of computing infrastructure information; Fig. Figure 4 is a diagram illustrating an example of job information; Fig. 5 is a diagram showing an example of an image for entering job information; Fig. Figure 6 is a conceptual diagram that illustrates an example of job processing; Fig. Figure 7 is a flowchart illustrating an example of job processing; and Fig. Figure 8 is a diagram that shows an example of an image displaying a result of machine learning. EXAMPLE OF EXECUTION

[0006] A processing method according to an exemplary embodiment is described with reference to the Fig. 1 to Fig. 8 described. (System)

[0007] The system according to the exemplary embodiment is described with reference to Fig. 1 described. In Fig. System 1 comprises data centers DC1, DC2, and DC3, which are interconnected via a network (NW), as well as clouds CL1 and CL2. The number of data centers included in System 1 can be two or fewer, or four or more. The number of clouds included in System 1 can be one, three, or more.

[0008] The locations of data centers DC1, DC2, and DC3 can be arbitrary. For example, data center DC1 could be located in Japan, data center DC2 in the United States, and data center DC3 in Europe. For instance, data center DC1 could be located in Aichi Prefecture, data center DC2 in Kyushu, and data center DC3 in Hokkaido.

[0009] At least one of the data centers DC1, DC2, and DC3 can be a container-type data center. At least part of the power supply for the container-type data center can use used batteries from BEVs (Battery Electric Vehicles).

[0010] At least one of the data centers DC1, DC2, and DC3 can be a data center owned by the company itself (i.e., an on-premises data center). Data centers DC1, DC2, and DC3 can also be provided by another company (i.e., a hosted data center). At least one of the clouds CL1 and CL2 can be a public cloud, sharing an environment built by a cloud service provider with other users. Clouds CL1 and CL2 can include a hosted private cloud, where a cloud environment provided by a cloud service provider is used exclusively by a specific user. Note that the hosted data center and the hosted private cloud can be considered the same concept.

[0011] The data centers DC1, DC2, and DC3, as well as the clouds CL1 and CL2, can be referred to as "nodes." Furthermore, the network NW can be referred to as a "connection." Therefore, System 1 can be described as a computer infrastructure with a multitude of nodes configured to communicate over the network NW.

[0012] System 1 is equipped with a database (DB). The database contains training data intended for machine learning. This training data can be based on commercially available training datasets or on data collected from a variety of vehicles (e.g., connected cars).

[0013] In System 1, machine learning can be performed using at least some of the training data contained in the database DB in at least some of the data centers DC1, DC2 and DC3 and the clouds CL1 and CL2. (Configuration of the information processing device)

[0014] System 1 is equipped with an information processing device 100. The information processing device 100 is described with reference to Fig. 2 described. In Fig. 2 The information processing device 100 is provided with an arithmetic device 110, a storage device 120, a communication device 130, an input device 140 and an output device 150. The arithmetic device 110, the storage device 120, the communication device 130, the input device 140 and the output device 150 can be connected via a data bus 160.

[0015] The information processing device 100 need not necessarily include at least one of the input devices 140 and the output devices 150. In this case, at least one of the input devices 140 and the output devices 150 may be connected to the information processing device 100 via an input / output port of the information processing device 100 (not shown) (i.e., at least one of the input devices 140 and the output devices 150 may be externally attached to the information processing device 100).

[0016] The arithmetic device 110 can contain one or more processors. The processor can be, for example, at least one CPU (central processing unit) or one GPU (graphics processing unit).

[0017] The storage device 120 can contain one or more memory devices. These memory devices can be, for example, at least one RAM (Random Access Memory), one ROM (Read Only Memory), a hard disk device, a magneto-optical disk device, or an SSD (Solid State Drive).

[0018] The communication device 130 can be configured to communicate with a device outside the information processing device 100. The communication device 130 can perform wired or wireless communication.

[0019] The input device 140 is a device configured to receive information input to the information processing device 100 from an external source. The input device 140 may include an operating device (e.g., a keyboard, a mouse, a touch panel, etc.) that can be operated by a user of the information processing device 100. The input device 140 may also include a recording medium reader configured to read information recorded on a recording medium that can be attached to or detached from the information processing device 100, such as a USB (Universal Serial Bus) storage device.In a case where information about the communication device 130 is entered into the information processing device 100 (in other words, in a case where the information processing device 100 obtains information about the communication device 130), the communication device 130 can function as an input device.

[0020] The output device 150 is a device configured to output information to the outside of the information processing device 100. The output device 25 can output visual information such as characters and images, auditory information such as voice / sound, or tactile information such as vibration, similar to the information mentioned above. The output device 25 can include, for example, at least one display, a loudspeaker, a printer, or a vibration motor. The output device 25 can be configured to output information to a recording medium that can be connected to or detached from the information processing device 100, such as a USB storage device. In a case where the information processing device 100 outputs information via the communication device 130, the communication device 130 can act as the output device.

[0021] Storage device 120 is configured to store the desired data. A computer program executed by arithmetic device 110 can be stored in storage device 120. Storage device 120 can temporarily store data that is temporarily used by arithmetic device 110 while the arithmetic device 110 executes the computer program.

[0022] The computer program can be recorded on a computer-readable and non-transient recording medium. In this case, the information processing device 100 can read the computer program from the aforementioned recording medium using a recording medium reader (not shown). Consequently, the computer program can be stored in the storage device 120. The recording medium can be at least one optical disk, a magnetic medium, a magneto-optical disk, a semiconductor memory, or any other medium configured for storing a program.

[0023] The computer program can be obtained from a device outside the information processing device 100 (not shown) via the communication device 130. That is, the information processing device 100 can download the computer program via the communication device 130. Consequently, the computer program can be stored in the storage device 120.

[0024] The arithmetic device 110, together with the storage device 120 in which the computer program is stored, can perform processing to be carried out by the information processing device 100. In other words, the arithmetic device 110, together with the storage device 120 and the computer program stored in the storage device 120, can perform the processing to be carried out by the information processing device 100. For example, the arithmetic device 110, which executes the computer program, can implement logical function blocks in the arithmetic device 110 to carry out the processing to be carried out by the information processing device 100.

[0025] For example, the arithmetic device 110 can include a detection unit 111, a selection unit 112, a determination unit 113, an identification unit 114, a calculation unit 115, and a comparison unit 116 as the functional blocks mentioned above. The arithmetic device 110 can include the detection unit 111, the selection unit 112, the determination unit 113, the identification unit 114, the calculation unit 115, and the comparison unit 116 as physically implemented processing circuits. At least one of the detection unit 111, the selection unit 112, the determination unit 113, the identification unit 114, the calculation unit 115, and the comparison unit 116 can be implemented in a hybrid form consisting of a logical functional block and a physical processing circuit (i.e., hardware).The recording unit 111, the selection unit 112, the determination unit 113, the identification unit 114, the calculation unit 115 and the comparison unit 116 will be described in detail later.

[0026] The storage device 120 stores compute resource information 121 and job information 122. The compute resource information 121 is information about a compute resource available for machine learning. For example, the compute resource information 121 could be information about each of the data centers DC1, DC2, and DC3, and the clouds CL1 and CL2. As in Fig. As shown in Figure 3, the computing resource information (Figure 121) could, for example, include information specifying the computing power, availability, and usage fee of each data center. The data center itself could be represented, for example, by information identifying the data center. Computing power could be represented, for example, by FLOPS (Floating-Point Operations Per Second). Availability could be represented, for example, by the number of available cores.

[0027] The “data center” in Fig. 3 is not limited to data centers DC1, DC2, and DC3, but conceptually also includes clouds CL1 and CL2. As described above, data centers DC1, DC2, and DC3, as well as clouds CL1 and CL2, can be referred to as "nodes." Therefore, the compute resource information 121 can also be referred to as node information. Furthermore, the compute resource information 121 may contain other details besides "data center," "computing power," "availability," and "usage fee."

[0028] Job information 122 refers to information about a task related to machine learning. As in Fig. As shown in Figure 4, the job information 122 could, for example, include information specifying a learning period, a data set, and a data set for each job. The learning period could be, for example, a date indicating a learning period or a timeframe from the present until a learning period. The data set could be, for example, information identifying a data set used for machine learning. The data set could be, for example, information indicating the amount of data in the data set. In addition to "job," "learning period," "data set," and "data set," the job information 122 could also contain other details.

[0029] When the user of information processing device 100 registers a job, a job can be created in Fig. Figure 20, shown in Figure 5, is displayed on a screen, which serves as an example of the output device 150. For example, the user can enter the required information into at least one of a variety of input fields contained in Figure 20 using the input device 140. When the user presses an "OK" button contained in Figure 20 using the input device 140, the information entered by the user is registered in the job information 122.

[0030] For example, information entered into an input field for "Job Name" in Figure 20 can be stored in a "Job" field of Job Information 122. Similarly, information entered into an input field for "Record" in Figure 20 can be stored in a "Record" field of Job Information 122. Similarly, information entered into an input field relating to "Learning Period" in Figure 20 can be stored in a "Learning Period" field of Job Information 122. For example, Information Processing Device 100 can identify the data volume of the record based on the information entered into the input field relating to "Record" in Figure 20. Information Processing Device 100 can then store the identified data volume in a "Data Volume" field of Job Information 122. (Operation of the information processing device)

[0031] The following describes the operation of the Information Processing Device 100. First, the processing procedure is explained in which the Information Processing Device 100 selects a data center to execute a job contained in the job information 122 (i.e., a data center that performs machine learning according to the job). Hereinafter, the term "data center" is not limited to data centers DC1, DC2, and DC3, but also includes clouds CL1 and CL2.

[0032] The procurement unit 111 of the arithmetic device 110 procures the data set and the learning period for a job contained in the job information 122. Based on the procured data set and the procured learning period, the selection unit 112 of the arithmetic device 110 can calculate the computing power required to complete a job within the learning period.

[0033] The selection unit 112 can extract one or more data centers based on the computing resource information 121 that can provide the required computing power (in other words, that can complete a job within the learning timeframe). The selection unit 112 selects from the extracted one or more data centers the data center that will execute a job, thus reducing the costs required for machine learning. The selection unit 112 can select a data center to perform machine learning according to a job. The selection unit 112 can select a plurality of data centers to perform machine learning for a job. If the selection unit 112 selects a plurality of data centers, the determination unit 113 of the arithmetic device 110 determines the training data to be input into each of the plurality of data centers, based on a dataset related to a job.

[0034] The selection unit 112 instructs the selected data center to execute a job via the communication device 130. For example, the selection unit 112 can register a job in a queue related to the selected data center. At this point, the information processing device 100 can transfer the data record related to a job to the selected data center from the database DB, based on the job information 122. If the selection unit 112 selects a plurality of data centers, the information processing device 100 can transfer training data related to a job to each of the plurality of data centers from the database DB, based on a determination result by the determination unit 113. When the selection unit 112 selects the data center that will execute a job, it can register the data center executing the job in the job information 122.

[0035] The data center usage fee varies for each data center. The usage fee is relatively low for on-premises data centers and relatively high for public cloud data centers. The usage fee for hosted data centers is often higher than for on-premises data centers and lower than for public cloud data centers.

[0036] For example, in a case where the extracted data center(s) described above include both on-premises and public cloud data centers, Selection Unit 112 can select the on-premises type to reduce the costs required for machine learning. Similarly, in a case where the extracted data center(s) described above include both hosted and public cloud data centers, Selection Unit 112 can select the hosted type to reduce the costs required for machine learning. In a case where Selection Unit 112 selects a variety of data centers to perform machine learning for a job, Selection Unit 112 can preferentially select the on-premises type to reduce the costs required for machine learning.

[0037] Next, the processing is explained in which the information processing device 100 transfers a remaining part of a job, contained in the job information 122, from one data center that is executing the one job to another data center.

[0038] As described above, the data center usage fee varies for each data center. For example, if another data center with a lower usage fee than the data center running a job is available, transferring the remaining part of that job from one data center to another will reduce the costs required for machine learning.

[0039] A concept for job transfer processing is developed with reference to Fig. 6 described. Fig. 6. It is assumed that data center DC1 is a local data center and cloud CL1 is a public cloud. It is also assumed that the usage fee for data center DC1 is lower than the usage fee for cloud CL1.

[0040] At a time when the information processing device 100 selects a data center to perform job J2, data center DC1 is assumed to be performing job J1. Therefore, it is assumed that the information processing device 100 selects cloud CL1, which is different from data center DC1, as the data center to perform job J2.

[0041] Job J1 will be performed at time t1 in Fig. 6 completed. Data center DC1 may then execute another job. This will occur within a period between time t1 and time t2 in Fig. 6. Information processing device 100 can determine whether or not to transfer job J2r, which is part of job J2 in cloud CL1 after time t2, to data center DC1. For example, information processing device 100 can determine whether data center DC1 will allow job J2r to be completed within the learning period, based on the amount of data used in job J2r, the time required to transfer the data to data center DC1, and the learning period related to job J2. The time tl in Fig. 6 corresponds to an example of the learning period for job J2.

[0042] If it is determined that data center DC1 can complete job J2r within the learning period, information processing device 100 can transfer job J2r to data center DC1. In this case, information processing device 100 can transmit information to cloud CL1 indicating that job J2r should be completed at time t2. Information processing device 100 can register a job J3, equivalent to job J2r, in a queue connected to data center DC1. As a result, data center DC1 can begin completing job J2r at time t3. Fig. 6. Perform machine learning according to job J3.

[0043] An initial value of a parameter of a learning model related to machine learning corresponding to job J3 can be a parameter value at time t2 of a learning model related to machine learning corresponding to job J2. Furthermore, in at least part of a period from time t2 to time t3, in Fig. 6. Transfer the data used in job J2r to data center DC1. This data can be transferred from cloud CL1 to data center DC1 or from database DB to data center DC1. Additionally, CRIU (Checkpoint / Restore in User Space) can be used for the job transfer described above.

[0044] For example, the usage fee for Cloud CL1 could be $10 per 10% of job J2. Similarly, the usage fee for Data Center DC1 could be $5 per 10% of job J2. If Cloud CL1 runs the entire job J2, the cost to run job J2 would be $100.

[0045] For example, at time t2 in Fig. 6. 50% of job J2 is executed. In this case, job J2r corresponds to 50% of job J2. As in Fig. As shown in Figure 6, in a case where job J2r is transferred to data center DC1 as job J3, data center DC1 executes 50% of job J2. That is, cloud CL1 executes 50% of job J2, and data center DC1 executes the remaining 50%. In this case, the cost to execute job J2 is $75. As described above, job transfer enables a reduction in the cost of executing the job (i.e., the cost of machine learning).

[0046] Job transfer processing is described using a flowchart in Fig. 7 described. Fig. 7. Procurement unit 111 of information processing device 100 can procure a job that is executed in a data center with a relatively high usage fee from job information 122 (step S101). In step S101, procurement unit 111 can, for example, procure a job with a low progress rate, such as the job mentioned above. The reason for this is that a significant cost-reducing effect is expected from the job transfer.

[0047] Then, the identification unit 114 of the information processing device 100 can obtain the learning period relating to one of the jobs described above from the job information 122. Based on the job information 122, the identification unit 114 can estimate the amount of data to be used for the remaining part of one of the jobs described above. Based on the estimated amount of data, the identification unit 114 can calculate the data transfer time required to transfer the data. Based on the learning period for that job, the estimated amount of data, and the calculated data transfer time, the identification unit 114 can calculate the computing power required to complete the remaining part of a job within the learning period.The identification unit 114 can identify another data center that can fulfill the calculated computing power (in other words, that can complete the remaining part of a job within the learning period) based on the computing resource information 121 (step S102). That is, one can say that the identification unit 114 identifies another data center based on the time required to execute the remaining part of a job and the data transfer time. The identification unit 114 can identify a variety of data centers as another data center.

[0048] Then, the computation unit 115 of the information processing device 100 can calculate the first cost and the second cost (step S103). The first cost is the cost of continuing a job in a data center that is currently executing the job (e.g., if the entire job is executed in a data center that is currently executing the job). The second cost is the cost of executing the remaining part of a job in a different data center identified by the identification unit 114. The second cost can be the sum of the cost of executing the remaining part of a job in a different data center and the cost of transferring the data used for the remaining part of a job. If the identification unit 114 identifies a plurality of other data centers, the computation unit 115 can calculate the second cost for each of the plurality of other data centers.

[0049] This section describes the costs for transferring the data used for the remaining part of a job. For example, costs are often incurred when extracting data from a public cloud. Therefore, in a case where the data center described above is a public cloud, the data transfer costs could be the sum of the costs for extracting the data used for the remaining part of a job from that one data center and the communication costs for transferring the data to another data center, as described above. In a case where there are no costs for extracting data from a data center, the data transfer costs could be equal to the communication costs for transferring the data used for the remaining part of a job to another data center, as described above.

[0050] Then, the comparator unit 116 of the information processing device 100 can compare the first costs with the second costs. The comparator unit 116 can determine whether the second costs are lower than the first costs or not (step S104). If, in step S106, it is determined that the second costs are lower than the first costs (step S104: Yes), the information processing device 100 can transfer a job from the data center that is currently executing the job to another data center (step S105). If, in step S106, it is determined that the second costs are not lower than the first costs (step S104: No), the one data center can continue to execute the job (i.e., the job cannot be transferred) (step S106).

[0051] When the machine learning corresponding to a job is complete, the information processing device 100 can retrieve result information indicating the job's outcome from the data center. The information processing device 100 can store the result information in the storage device 120. A user of the information processing device 100 can cause the information processing device 100 to display the result information via the input device 140. In this case, the information processing device 100 can display a Fig. 8 Display the illustrated image 30 on a display that serves as an example of the output device 150.

[0052] Job transfers can be performed not just once, but multiple times. For example, a job running in a public cloud can be transferred to a hosted private cloud, and then from the hosted private cloud to an on-premises data center. Job transfers can occur not only between different types of data centers, but also between data centers of the same type. For instance, a job running in a public cloud with a relatively high usage fee can be transferred to a public cloud with a relatively low usage fee.

[0053] For data transfer related to job transfers, a technique / technology such as Linux containers can be used, where the data required to execute a job (e.g., at least one of the applications, libraries, dependencies, and files) are transferred together. As described above, costs can arise when data is extracted from a data center (e.g., a public cloud). When a job running in one data center is transferred to another, some of the data related to that job (e.g., training data) can be deleted from the first data center after the job has finished executing and before the data is transferred to the second data center. This configuration makes it possible to reduce the amount of data extracted from the first data center during the job transfer.This means it is possible to reduce the costs of extracting data from a data center. In this case, the data deleted from one data center can be transferred, for example, from the database DB to another data center. In this case, the data extracted from one data center (i.e., the data transferred from one data center to another) can contain metadata about the deleted data. (Application example)

[0054] A learned / trained model generated by machine learning with the system 1 described above can, for example, be applied to an advanced drive / advanced drive assistance system.

[0055] For example, a baseline model relating to the advanced driver assistance function can be created as a learned model through machine learning using System 1 and commercially available learning datasets contained in the database DB. Furthermore, the baseline model can be fine-tuned through machine learning using System 1 and learning data based on data collected from a large number of vehicles traveling in a specific region, also contained in the database DB. As a result, a learned model of the advanced driver assistance function optimized for the specific region can be generated. Note that low-rank adaptation (LORA) can be used for fine-tuning. (Technical effect)

[0056] AI can be used to create a safer and more comfortable driving environment for vehicles. For example, operational support for peripheral devices such as air conditioning and audio systems, assistance with safer driving, and similar functions can be achieved by running the learned model (i.e., AI) related to the advanced driver assistance function on a device within the vehicle. By running a learning model related to the advanced driver assistance function on a server in a network, in addition to or instead of the in-vehicle device, more enhanced services can be made available to a vehicle user via the communication device installed in the vehicle. A server providing such services must respond to user requests in real time. However, only relatively small amounts of data are input to the server.

[0057] For example, to develop the AI ​​associated with the advanced driver assistance function, a server performing machine learning (similar to a server in the aforementioned data center) must process a large amount of data. However, real-time response is not required if a predetermined development schedule is adhered to. In other words, the processing time does not necessarily need to be short if the development schedule is met. Therefore, the server running the learned model and the server performing the machine learning must have different performance levels.

[0058] As described above, the usage fee often varies depending on the data center. In System 1 according to the present embodiment, the information processing device 100 selects the data center such that the machine learning is completed within the learning period and the costs required for machine learning are reduced. In System 1 according to the present embodiment, the information processing device 100 can also transfer a job running in one data center to another data center, thus reducing the costs required for machine learning. That is, in System 1, the data center is selected to reduce the costs required for machine learning while adhering to a predetermined development plan.Therefore, according to System 1 in the present embodiment, it is possible to select the data center in such a way that the costs for developing a learning model are reduced / controlled.

[0059] System 1 can include an on-premises data center that meets the stable computing needs of the company, and at least one hosted private cloud and one public cloud that meet the remaining computing needs of the company. The Information Processing Device 100 can select the data center to ensure that machine learning is completed within the learning timeframe and that the costs associated with machine learning are reduced. This configuration makes it possible to reduce / control the costs of developing the learning model while simultaneously meeting the company's computing needs. (First modified example)

[0060] The computing resource information 121 may also contain environmental impact information indicating any environmental impact associated with the data center. This environmental impact information may include, for example, an index indicating the level of environmental impact. The environmental impact information may also include information about the type of energy used by the data center. This energy type may include, for example, green energy, renewable energy, fossil energy, and similar sources.

[0061] The computing unit 115 of the information processing device 100 can, for example, calculate the first costs and the second costs based on the environmental impact information contained in the computing resource information 121. In this case, the first costs can include the costs of an initial environmental impact related to the environmental impact of a data center currently performing a job. The second costs can include the costs of a second environmental impact related to the environmental impact of another data center identified by the identification unit 114. The environmental impact costs can, for example, be costs for reducing the environmental impact of the data center.In this case, the environmental impact of a data center using fossil fuels can be higher than that of a data center using renewable energy. This configuration makes it possible to reduce the environmental impact while simultaneously lowering the costs of developing the learning model. (Second modified example)

[0062] The computing resource information 121 may also include environmental impact information, which specifies the environmental impact in relation to the data center, and power supply information about the power supply situation in an area that includes the data center. The power supply information may include, for example, the amount of energy generated by solar power generation, the amount of energy generated by wind power generation, the amount of energy stored in storage batteries, the presence or absence of power suppression, and similar information.

[0063] The computation unit 115 of the information processing device 100 can calculate a first result and a second result instead of first costs and second costs. The first result is a result if a job continues in a data center that is currently running the first job (e.g., if the entire job is executed in a data center that is currently running the first job). The second result is a result if the remaining part of a job is executed in a different data center, identified by the identification unit 114. The first and second results can be calculated based on a monetary score related to usage fees, an environmental score related to environmental impact, and an energy score related to the energy supply situation.The monetary outcome may be lower because the data center usage fee is lower. The environmental impact may be smaller because the data center has less of an environmental footprint. The energy consumption impact may be smaller because the electricity supply capacity in the area where the data center is located is greater.

[0064] For example, the computation unit 115 can calculate the first result as "w1×(first monetary result) + w2×(first environmental result) + w3×(first energy result)". Here, the first monetary result, the first environmental result, and the first energy result refer to a monetary result, an environmental result, and an energy result, respectively, that is assigned to a data center. The computation unit 115 can calculate the second result as "w1×(second monetary result) + w2×(second environmental result) + w3×(second performance result)". Here, the second monetary result, the second environmental result, and the second performance result refer to a monetary result, an environmental result, and a performance result, respectively, that is assigned to a different data center. Furthermore, "w1", "w2", and "w3" are weights. The weight w1 is greater than the weights w2 and w3.A size ratio of the weights w2 and w3 can be determined according to a user guideline.

[0065] The comparator 116 of the information processing device 100 can compare the first result with the second result. If the second result is less than the first result, the information processing device 100 can transfer a job from the data center currently executing the job to another data center. If the second result is not less than the first result, the one data center can continue executing the job (i.e., the job cannot be transferred).

[0066] For example, the amount of energy generated by solar and wind power is easily influenced by weather conditions. If the amount of energy generated by solar or wind power exceeds a certain level of consumption, solar or wind power generation is temporarily halted. This means there are instances where solar and wind power generation cannot be fully utilized. The data center consumes a relatively large amount of electricity. For this reason, it is expected that the possibilities for utilizing solar and wind power will increase if a data center is located in a region with a relatively large power supply capacity.

[0067] By deciding whether or not to transfer the job based on the first and second results, it is possible to reduce environmental impact while simultaneously lowering the costs of developing the learning model. Furthermore, transferring the job to the data center in a region with relatively large power supply capacity can prevent a temporary shutdown of solar or wind power generation.

[0068] Aspects of the present disclosure that arise from the embodiments and modified examples described above are described below.

[0069] A processing method according to one aspect of the present disclosure comprises the following steps: obtaining job information about a job relating to machine learning and running on one of several nodes; identifying another of the several nodes based on the job information that will allow a remaining portion of the job to be completed within a learning timeframe; comparing the first cost, which is the cost in a case where the job continues on the one node, with the second cost, which is the cost in a case where the remaining portion of the job is executed on the other node; and transferring the job from one node to the other node in a case where the second cost is lower than the first cost. In the above embodiment, “the data centers DC1, DC2, and DC3 and the clouds CL1 and CL2” represent an example of “nodes.”

[0070] In one example of the processing procedure, the other node can be identified in the identification step based on the time required to execute the remaining part of the job, calculated based on the job information, and the time required to transfer the data used to execute that remaining part. In this example, the second cost could be the sum of the cost of executing the remaining part of the job and the cost of transferring the data.

[0071] In another example of the processing procedure, the first costs may include initial environmental impact costs relating to an environmental impact on one node, and the second costs may include secondary environmental impact costs relating to an environmental impact on the other node.

[0072] In another example of the processing procedure, in the comparison step, a first result, which is the result if the job continues at one node, can be compared instead of the first cost with a second result, which is the result if the remaining part of the job is carried out at the other node, instead of the second cost. In the transfer step, the job can be transferred from one node to the other in a case where the second result is smaller than the first result. The first result can be calculated based on a monetary result, an environmental impact result, and a power result with respect to one node, and the second result can be calculated based on a monetary result, an environmental impact result, and a power result with respect to the other node.

[0073] In another example of the processing procedure, the processing procedure may further include a step of issuing a report about the job in response to the completion of the job, with the report possibly containing information about the costs.

[0074] In one example of the processing procedure, the multitude of nodes can include at least two on-premises, one private cloud, and one public cloud.

[0075] A system according to one aspect of the present disclosure is a system that controls machine learning of a model in a computer infrastructure containing a plurality of nodes configured to communicate over a network, the system comprising: a procurement unit that procures job information about a job relating to machine learning and executed on one of several nodes; an identification unit that, based on the job information, identifies another node of the multiple nodes that makes it possible to complete a remaining part of the job within a learning period;and a comparison unit that compares first costs, which are the costs in a case where the job continues at one node, with second costs, which are the costs in a case where the remaining part of the job is executed at the other node, transferring the job from one node to the other if the second costs are lower than the first costs.

[0076] The present disclosure is not limited to the examples described above and may be modified, if desired, without departing from the essence or basic concept of the invention, which can be derived from the claims and the entire specification. A processing method incorporating such modifications is also included in the technical concepts of the present disclosure. DESCRIPTION OF REFERENCE NUMBERS

[0077] 1: System, 100: Information processing device, 111: Procurement unit, 112: Selection unit, 113: Destination unit, DB: Database, DC1, DC2, DC3: Data center, CL1, CL2: Cloud, NW: Network QUOTES INCLUDED IN THE DESCRIPTION

[0000] This list of documents cited by the applicant was automatically generated and is included solely for the reader's convenience. The list is not part of the German patent or utility model application. The DPMA accepts no liability for any errors or omissions. Cited patent literature

[0000] JP 2022-034850A

[0002]

Claims

[1] Processing procedure with the steps Obtaining job information about a job related to machine learning that is performed on one of a large number of nodes, Identifying another node from the multitude of nodes that will allow completion of a remaining part of the job within a learning period, based on the job information, Comparing first costs, which are the costs in a case where the job continues at one node, with second costs, which are the costs in a case where the remaining part of the job is performed at the other node, and Transferring the job from one node to another in a case where the second costs are lower than the first costs. [2] Processing method according to claim 1, wherein the other node is identified in the identification step based on a time required to perform the remaining part of the job, calculated based on the job information, and a time required for a transfer of data used to perform the remaining part of the job. [3] Processing method according to claim 1, wherein the second costs are a sum of costs to perform the remaining part of the job and costs to transfer the data. [4] Processing method according to claim 1, wherein The initial costs include the initial environmental impact costs related to an environmental impact on one node, and The second costs include the second environmental impact costs related to an environmental impact on the other node. [5] Processing method according to claim 1, wherein In the comparison step, instead of the first cost, a first result, which is the outcome when the job continues at one node, is compared with a second result, which is the outcome when the remaining part of the job is performed at the other node, instead of the second cost. In the transfer step, the job is transferred from one node to another in one case if the second result is smaller than the first result. the first result is calculated based on a monetary result, an environmental impact result and an energy supply result related to the one node, and The second result is calculated based on a monetary result, an environmental impact result, and an energy supply result related to the other node. [6] Processing method according to claim 1, further comprising a step of issuing a report about the job in response to the completion of the job, wherein the report contains information about costs. [7] Processing method according to claim 1, wherein the plurality of nodes includes at least two of the following: on-premises nodes, private cloud nodes and public cloud nodes.

Images