Data processing method, data processing device, and program

The method addresses uncertain data arrival timing in pipeline processing by determining server selection and execution timing, ensuring timely and efficient data processing in real-time architectures.

WO2026239454A1PCT designated stage Publication Date: 2026-11-19NT T INC
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
PCT/JP2025/017260
Authority / Receiving Office
WO · WO
Patent Type
Applications
Current Assignee / Owner
Filing Date
2025-05-12
Publication Date
2026-11-19

AI Technical Summary

Technical Problem

Conventional pipeline processing assumes uncertain data arrival timing, which hinders effective integration with real-time architectures, especially in deterministic delay networks.

Method used

A data processing method that determines server selection and execution timing based on processing time and resource availability, dividing processes into stages and controlling tasks on multiple servers to handle data with definite arrival timing.

Benefits of technology

Enables appropriate processing of data with definite arrival times, ensuring no tasks fall behind in periodic processing and maintaining overall processing time within desired limits.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure JP2025017260_19112026_PF_FP_ABST
    Figure JP2025017260_19112026_PF_FP_ABST
Patent Text Reader

Abstract

A data processing system 1 comprises a management server 10 and a plurality of calculation servers 20, and executes, in a stepwise manner on the plurality of calculation servers 20, a plurality of processes on data for which the magnitude of the arrival delay has been determined. The management server 10: determines which calculation server 20 will execute each stage by solving an optimization problem for minimizing the number of calculation servers 20 to be used under a constraint condition formulated on the basis of the stage length when tasks of each stage are executed in each calculation server 20 and the resources of each calculation server 20; determines a timing for executing each task of each stage on the basis of an execution order and the stage length of each stage; and performs control so that each task of each stage is executed at the determined timing on the determined calculation server 20.
Need to check novelty before this filing date? Find Prior Art

Description

Data processing method, data processing apparatus, and program

[0001] The present disclosure relates to a data processing method, a data processing apparatus, and a program.

[0002] A real-time architecture is an architecture that is required to reliably respond within a defined period of time. As one type thereof, there is a system in which processing is executed periodically and managed in time slots such that each piece of processing is completed within a predetermined period of time. In this system, the cycle is determined in advance, and the data acquisition timing is determined according to a pre-designed schedule. Such a real-time architecture is suitable for stream processing that processes data periodically.

[0003] On the other hand, as described in Non-Patent Document 1, along with the change in CPU configuration from single-core to multi-core, multi-threaded pipeline processing has been increasingly adopted.

[0004] "'Enterprise Data Platform Solution' Realizes Real-Time Data Management Utilization of Big Data is the Key to Business Strengthening", NTT Comware Corporation, https: / / www.nttcom.co.jp / comware_plus / solution / edp_1.html

[0005] In an all-photonics network (APN) in which all communication between apparatuses is configured with optical technology, the data arrival timing is deterministic. That the data arrival timing is deterministic means that the data arrival time can be predicted, and is also referred to as deterministic delay. If the data arrival timing is deterministic, stream processing by a real-time architecture in which processing is executed periodically is possible even between distant points.

[0006] Against this background, by combining the real-time architecture and pipeline processing, both real-time performance and increased processing throughput can be achieved.

[0007] However, conventional pipeline processing assumes that the arrival timing (delay) of the data to be processed is uncertain, leaving room for improvement when the arrival timing of the data to be processed is certain.

[0008] This disclosure is made in view of the above, and aims to appropriately perform data processing on data with a definite arrival timing when combining a real-time architecture and pipeline processing.

[0009] A data processing method in one aspect of the present disclosure is a data processing method that performs a plurality of processes in stages on a plurality of servers for which the magnitude of delay arriving from a sensor or edge device to each of a plurality of candidate servers for processing has been determined, wherein a computer determines which server to execute each of the plurality of processes on based on the processing time when each of the plurality of processes is executed on each of the plurality of servers and the amount of resources of each of the plurality of servers, determines the timing for executing each of the plurality of processes based on the execution order of the plurality of processes and the processing time, and controls each of the plurality of processes to be executed on the determined server at the determined timing, wherein the data is video data of an intersection, and each of the plurality of processes is a process for decoding an image from the video data, a process for detecting a vehicle from the decoded image, a process for obtaining information of the detected vehicle, and a process for evaluating the collision risk using the vehicle information.

[0010] According to this disclosure, when a real-time architecture is combined with pipeline processing, it is possible to appropriately process data with a definite arrival time.

[0011] Figure 1 is a diagram showing the configuration of the data processing system. Figure 2 is a diagram showing an example of stage division. Figure 3 is a diagram showing an example of task assignment. Figure 4 is a diagram showing an example of stage division. Figure 5 is a diagram showing an example of task assignment. Figure 6 is a diagram showing an example of data pipeline configuration information. Figure 7 is a diagram showing an example of node configuration information. Figure 8 is a diagram showing an example of stage length information. Figure 9 is a diagram showing an example of task physical placement information. Figure 10 is a diagram showing an example of a task execution plan. Figure 11 is a flowchart showing an example of the processing flow of the data processing system. Figure 12 is a flowchart showing an example of the processing flow for calculating the stage length on each xPU for each stage. Figure 13 is a flowchart showing an example of the processing flow for determining the physical xPU on which to place tasks. Figure 14 is a flowchart showing an example of the processing flow for determining the assignment of tasks to physical xPUs and the execution timing. Figure 15 is a diagram showing an example of the hardware configuration of the management server.

[0012] [Example 1] (System Configuration) Referring to Figure 1, the data processing system of Example 1 will be described. The data processing system 1 shown in Figure 1 (hereinafter referred to as System 1) is a system that photographs an intersection with multiple cameras and calculates the risk of vehicle collisions. For example, System 1 predicts collisions between vehicles turning right and vehicles going straight, and head-on collisions. If the collision risk is high, System 1 provides real-time feedback of the collision risk evaluation results to the intersection. The feedback informs vehicles and pedestrians in the intersection of the collision risk using sound and visual effects.

[0013] System 1 in Figure 1 comprises a management server 10, multiple computing servers 20, a data acquisition server 30, and a camera 40.

[0014] Camera 40 is positioned near the intersection and captures images of the intersection. Camera 40 may be other devices such as sensors, or other devices may be added in addition to camera 40. In the example in Figure 1, ten cameras 40 are positioned at intersection 0 and five cameras are positioned at intersection 1. The number of cameras 40 can be set arbitrarily. System 1 analyzes the images captured by the ten cameras 40 at intersection 0 to calculate the collision risk at intersection 0, and analyzes the images captured by the five cameras 40 at intersection 1 to calculate the collision risk at intersection 1. In System 1, it can be said that intersection application 0, which evaluates the collision risk at intersection 0, and intersection application 1, which evaluates the collision risk at intersection 1, are operating.

[0015] The data acquisition server 30 transmits video data from the camera 40 to the computing server 20 via the network. The network is preferably a deterministic delay network where the timing of data arrival is deterministic. Alternatively, the camera 40 may be directly connected to the network without the data acquisition server 30. Alternatively, the camera 40 and the computing server 20 may be placed in close proximity and connected to each other using shared memory without using a network.

[0016] The computing server 20 is a server equipped with one or more xPUs and executes tasks in each stage of the data pipeline with a fixed delay. xPUs are a general term for processors specialized for specific processing. xPUs include Central Processing Units (CPUs), Graphics Processing Units (GPUs), Tensor Processing Units (TPUs), Field-Programmable Gate Arrays (FPGAs), and Application Specific Integrated Circuits (ASICs). A stage is a processing unit that divides the entire application processing into multiple processes executed in stages. In this embodiment, processing is assigned to the computing server 20 on a stage-by-stage basis. Each stage is executed on a different computing server 20. Each stage may be executed on a different xPU on the same computing server 20. A certain stage may be processed by multiple xPUs (computing servers 20). Also, different stages may be executed on the same xPU. The data acquisition server 30 may execute stages.

[0017] Figure 2 shows an example of the stages in an intersection application. The entire intersection application process is divided into four stages: Stage A, which decodes the video data sent from the camera; Stage B, which detects vehicles from the decoded image; Stage C, which extracts vehicle information by cutting out the vehicles from the decoded image; and Stage D, which generates a digital twin of each vehicle in the intersection and predicts a collision. Since each application is divided into stages, the total number of stages processed by System 1 is the number of processing divisions multiplied by the number of applications.

[0018] Each stage has a task for each piece of data to be processed. In other words, a task is the processing unit for each piece of data in a stage. For example, in the case of intersection application 0, which analyzes video data from 10 cameras 40, one task corresponds to the video data from one camera 40, and each stage will have 10 tasks. A single task may process the data from multiple cameras 40 at once. In addition, in stage D of the intersection application, the processing results of multiple tasks in stage C may be integrated and processed in a single task. Integration and branching of processing may be performed at any stage.

[0019] Tasks are executed on the computing servers 20 assigned by the management server 10 according to the specified timing. Figure 3 shows an example of task assignment. For example, "Data 0 Processing A" in Figure 3 represents one task. Data 0 to 3 represent the video data from each of the four cameras, and Processing A to D represent the processing of stages A to D. xPU0 to 3 are xPUs provided by the computing server 20. The data processed in each stage is read and processed by the xPU executing the next stage. For the data pipeline to function properly, each task must be completed within a predetermined time, and the processing results must be passed to the next xPU with a fixed delay.

[0020] Note that the processing time for each stage may differ, and the tasks in each stage may be executed in parallel. Figure 4 shows an example of stages with different processing times and tasks executed in parallel. In the example in Figure 4, the processing time for stages B and C is longer than that of stages A and D, and the tasks in stages B and C are executed in parallel. Figure 5 shows an example of task allocation. For each data 0 to 3, processing for the next stage begins after processing for the previous stage is completed. Process B is executed in parallel on xPU1_0 to xPU1_2, and process C is executed in parallel on xPU2_0 to xPU2_1.

[0021] The management server 10 comprises a task evaluation unit 11, a task placement determination unit 12, a task assignment unit 13, and a task execution unit 14. Based on data pipeline configuration information and node configuration information, the management server 10 determines the physical placement and execution timing of tasks at each stage of the application, assigns the tasks to the physical xPUs of the computing server 20, and causes the computing server 20 to execute the tasks. In Figure 1, the management server 10 and the computing server 20 are shown as separate devices, but the management server 10 may be part of the computing server 20.

[0022] Data pipeline configuration information is a file that defines the number of streams (tasks) to be processed, the required throughput, and information about each stage into which the application's processing is divided, for each application executed by System 1. The stage information includes the processing content of the stage and the type of xPU used. The stage information may also include the required throughput for each stage. The data pipeline configuration information is created by the application designer and input into the management server 10.

[0023] Figure 6 shows an example of data pipeline configuration information. In the example in Figure 6, the number of streams, required throughput, and stage information are shown for App 0 and App 1. In the example in Figure 6, the processing for App 0 and App 1 consists of Stage 0 and Stage 1. The processing content and the type of xPU used are stored as information for each stage.

[0024] Node configuration information is a file that defines the physical configuration of each computing server 20. Figure 7 shows an example of node configuration information. In the example in Figure 7, information on available xPUs, main memory, and storage is shown for each computing server 20. The available xPU section includes information such as the xPU model number, the number of cores and units, and performance indicators. Note that since tasks can be executed individually on each core, the number of cores can also be said to be the number of xPUs. The main memory and storage information represents the amount of resources held by the computing server 20. Node configuration information may also define arbitrary performance indicators such as NIC bandwidth, memory bandwidth, and I / O bus bandwidth.

[0025] The task evaluation unit 11 evaluates (also called calculates) the stage length (length of processing time) and resource consumption for each stage included in the data pipeline configuration information when the task of each stage is executed on each xPU included in the node configuration information. If the xPU performance of the computing servers 20 is uniform, the stage length is evaluated for each computing server 20 or for each type of xPU provided by the computing servers 20. If the xPU performance of the computing servers 20 is uneven, the stage length is evaluated for each xPU.

[0026] For example, for each application, the task evaluation unit 11 selects one xPU from all xPUs in system 1 for each stage, executes one of the tasks in the stage on the selected xPU, and measures the stage length and resource consumption (memory, storage, etc.). For the other xPUs in system 1, the stage length is calculated based on the difference in performance indicators, using the measurement results of the above xPU as a reference. If the xPU performance of the computing server 20 is uniform, the stage length is calculated for one of the xPUs in the computing server 20, and the other xPUs in the computing server 20 are also considered to have the same stage length. Alternatively, the task evaluation unit 11 may measure the stage length by executing tasks on all xPUs.

[0027] Figure 8 shows an example of stage length information created by the task evaluation unit 11. In the example in Figure 8, for each application, the stage length and resource consumption on the xPU of each computing server 20 are shown for each stage.

[0028] The task allocation determination unit 12 assigns the tasks of each stage to the physical xPUs of each computing server 20.

[0029] For example, if the xPU performance of the computing servers 20 is uniform and tasks belonging to the same stage are processed on the same computing server 20, the task placement determination unit 12 calculates the number of xPUs and resources required for each computing server 20 to execute all tasks in that stage for each computing server 20. Based on the conditions for allocating computing servers 20 between stages of the same application, the conditions for xPU consumption, memory consumption, and storage consumption of each stage on the computing server 20, and the conditions for processing time, the task placement determination unit 12 allocates the stages to the computing servers 20 in a way that minimizes the number of computing servers 20. Specifically, the task placement determination unit 12 allocates the stages to the computing servers 20 in a way that minimizes the number of computing servers 20 based on the following conditions.

[0030] - No duplicate stages of the same application are assigned to each computing server. - For a given computing server, the total xPU consumption of the stages assigned to that server is less than or equal to the xPU capacity of that server. - For a given computing server, the total memory consumption of the stages assigned to that server is less than or equal to the memory capacity of that server. - For a given computing server, the total storage consumption of the stages assigned to that server is less than or equal to the storage capacity of that server. - The "time from start to finish of data processing" for the entire application is less than or equal to the threshold value specified by the user.

[0031] The total consumption of xPUs, memory, and storage can be calculated using the required number of xPUs and the required amount of resources. The time from the start to the end of data processing should take into account the stage length for each server.

[0032] The task placement determination unit 12 may solve the optimization using the above conditions as an integer programming problem. Alternatively, it may use heuristics or metaheuristics such as the Best-Fit algorithm or genetic algorithms.

[0033] Figure 9 shows an example of task physical placement information created by the task placement determination unit 12. In the example in Figure 9, the xPUs assigned to each task in each stage of each application are shown. Since the same xPU of the computing server 20 is assigned to each stage, tasks 0 to 9 in stage 0 of application 0 in Figure 9 are assigned multiple CPU cores of the same computing server 20. Tasks 0 to 9 may be executed on any of the CPU cores.

[0034] The task assignment unit 13 determines the execution timing of each task by referring to the task's physical location information and stage length information. Specifically, the task assignment unit 13 assigns the xPU that will become available earliest from among those allocated to the task, after the processing of the input data for that task in the previous stage is completed. The task assignment unit 13 defines this timing as the task start time, and the task end time as the task start time plus the stage length of the allocated xPU.

[0035] Figure 10 shows an example of a task execution plan created by the task assignment unit 13. In the example in Figure 10, tasks 0 and 1 of stage 0 of application 0 are assigned to different CPU cores with the same start time. Since the stage length of the CPU cores assigned to tasks 0 and 1 is 20ms, the end time is set to the start time plus 20ms. Also, tasks 0 and 1 of stage 1 of application 0 are assigned to the same GPU, and the start time of task 1 is set after the end time of task 0.

[0036] The task execution unit 14 sends the stage (including processing details) and task execution plan to the computing server 20 for each stage, and issues a task execution instruction to the stage. The stages running on the computing server 20 execute each task at the specified time on the specified xPU according to the task execution plan.

[0037] (System Operation) Refer to the flowchart in Figure 11 to explain an example of the processing flow of System 1. Details of each process will be described later.

[0038] In step S1, the task evaluation unit 11 calculates the stage length and resource consumption amount at each xPU for each stage.

[0039] In step S2, the task placement determination unit 12 determines the physical xPU assignment for each task based on the stage length of the task at each xPU.

[0040] In step S3, the task assignment unit 13 determines the execution timing of each task based on the physical xPU assignment of the task and the stage length at the relevant xPU.

[0041] In step S4, the task execution unit 14 executes each task on the calculation server 20 in accordance with the task execution plan.

[0042] With reference to the flowchart of Figure 12, the processing of calculating the stage length at each xPU for each stage by the task evaluation unit 11 will be described.

[0043] The task evaluation unit 11 executes the processes of the following steps S11 to S13 for each stage.

[0044] In step S11, the task evaluation unit 11 selects a reference xPU (hereinafter referred to as the reference xPU) from all xPUs in the system 1, and obtains the stage length when executing the task of the stage on the reference xPU. The task evaluation unit 11 may select the xPU considered to have the highest or lowest performance as the reference xPU, or may select any arbitrary xPU as the reference xPU.

[0045] The task evaluation unit 11 actually executes the task of the stage on the reference xPU to measure the stage length and resource consumption amount. Note that the task evaluation unit 11 may calculate the stage length and resource consumption amount on paper from values such as the number of execution instructions of the stage, the number of executable instructions per second of the xPU, and the used data format.

[0046] The task evaluation unit 11 executes the processes of the following steps S12 and S13 for all the calculation servers 20 in the system 1.

[0047] In step S12, the task evaluation unit 11 sets one of the xPUs included in the calculation server 20 to be processed as the target xPU, and compares the performance indices of the target xPU and the reference xPU. Specifically, the task evaluation unit 11 acquires the performance indices of the target xPU and the reference xPU from the node configuration information, and calculates the performance difference between the xPUs (for example, the magnification of the performance index of the target xPU relative to the performance index of the reference xPU). FLOPS, IPS, or the like can be used as the performance index. The task evaluation unit 11 may acquire the model number of the xPU, and calculate the performance difference by comparing the benchmark scores for each model number.

[0048] In step S13, the task evaluation unit 11 obtains the stage length of the target xPU from the performance difference between the target xPU and the reference xPU and the stage length of the reference xPU. For example, the task evaluation unit 11 multiplies the stage length of the reference xPU by the magnification of the performance index of the target xPU relative to the performance index of the reference xPU to obtain the stage length of the target xPU.

[0049] When it is assumed that the xPU performance in the calculation server 20 is uniform, the task evaluation unit 11 regards other xPUs in the same calculation server 20 as having the same stage length as the target xPU. Note that even for the same calculation server 20, if the xPU types are different, the performances cannot be regarded as the same, so the stage length is obtained for each xPU type.

[0050] When the xPU performance in the calculation server 20 is non-uniform, the task evaluation unit 11 individually obtains the stage length for all xPUs included in the calculation server 20.

[0051] Regarding the processing of steps S12 and S13, the task evaluation unit 11 may actually execute the stage task on the target xPU to measure the stage length.

[0052] The above processing is executed for all calculation servers 20 for all stages.

[0053] With reference to the flowchart of FIG. 13, the processing of determining the physical xPU on which the task is to be arranged by the task arrangement determination unit 12 will be described.

[0054] In step S21, the task placement determination unit 12 calculates the required number of xPUs and the required amount of resources for each stage when the stage is executed on each xPU. In other words, the task placement determination unit 12 calculates the required number of xPUs and the required amount of resources for all combinations of stages and computing servers 20. The required number of xPUs can be calculated by multiplying the throughput required by the stage by the stage length when that stage is allocated to an xPU. The required amount of resources can be calculated by multiplying the required number of xPUs by the resource consumption per task.

[0055] In step S22, the task placement determination unit 12 obtains the number of xPUs of each type and the amount of various resources provided by each computing server 20 from the node configuration information.

[0056] In step S23, the task placement determination unit 12 sets constraints for determining the physical placement of tasks. For example, the task placement determination unit 12 sets constraints such as the aforementioned conditions using the number of xPUs of each type and the amount of various resources available in each computing server 20, the required number of xPUs, the required amount of resources, and the stage length on each xPU.

[0057] In step S24, the task placement determination unit 12 uses a mathematical optimization solver to determine the physical placement for each stage that minimizes the number of computing servers 20 used while satisfying the constraints.

[0058] Referring to the flowchart in Figure 14, the process by which the task allocation unit 13 assigns tasks to physical xPUs and determines the execution timing will be explained.

[0059] The task assignment unit 13 executes the following steps S31 and S32 for all tasks in all stages.

[0060] In step S31, the task assignment unit 13 assigns the task to the xPU that will become available earliest after the processing of the data that will be the input for the task in the previous stage is completed, from among the xPUs to which the task can be assigned.

[0061] In step S32, the task allocation unit 13 reserves the xPU from the task start time to the task end time, with the task end time being the task start time plus the stage length on the allocated xPU.

[0062] (Other Application Examples) In the above Example 1, we described an application for evaluating collision risk in intersections, but significant benefits can be expected from using System 1 in various applications that perform data pipeline processing. Note that the application processing and stage division are just examples, and other processing and divisions are also possible.

[0063] For example, System 1 can be used for short-term power supply and demand adjustment. In short-term power supply and demand adjustment, System 1 is used to collect and analyze power consumption and power supply from factories, data centers, homes, etc., at high frequency and with low latency. Based on the results of the near-future power supply and demand forecast, demand and supply are adjusted in response to adjustment requests from the grid, etc., thereby achieving a stable power supply. The short-term power supply and demand adjustment process can be divided into three stages: Stage A, which receives and analyzes power consumption and power supply data; Stage B, which predicts the amount of power in the near future; and Stage C, which estimates the amount of power to respond to the adjustment requests.

[0064] System 1 can be used for low-latency financial trading, which involves automated trading in financial markets. In low-latency financial trading, System 1 is used to collect various market data that changes moment by moment and to perform real-time analysis. Low-latency financial trading can be divided into four stages: Stage A, which receives and preprocesses data from the market; Stage B, which analyzes the data and infers market trends; Stage C, which makes buy / sell decisions; and Stage D, which automatically generates buy / sell orders and transmits them to the exchange's system.

[0065] System 1 can be used for Virtual Reality (VR) / Augmented Reality (AR), which provides users with interactive experiences in a non-realistic space. In a VR / AR system, the user wears a low-power device, transmits data from the user's device, and uses System 1 to perform deterministic latency computing, enabling low-latency, high-quality rendering. The processing of the VR / AR system can be divided into three stages: Stage A, which preprocesses user data; Stage B, which renders the virtual space based on the user data; and Stage C, which transmits the rendering results.

[0066] System 1 can be used in a robot control system for industrial robots operating in factories and other facilities. By using System 1 to perform advanced analysis of sensor data from the robot, the robot control system can achieve real-time robot operation with high decision-making capabilities. The processing of the robot control system can be divided into four stages: Stage A, which preprocesses the sensor data from the robot; Stage B, which performs inference using the sensor data; Stage C, which makes a decision based on the inference result; and Stage D, which transmits the decision result.

[0067] [Example 2] Next, Example 2 will be described. In Example 2, tasks within a stage are placed on the computing server 20, and it is assumed that the xPU performance within the computing server 20 is uneven. Here, the differences from Example 1 will be explained.

[0068] In Example 1, the stage length on the xPU was calculated for each computing server 20, but in Example 2, the stage length is calculated for each xPU. The stage length of the xPU can be measured or calculated in the same way as in Example 1.

[0069] Furthermore, in Example 1, the required number of xPUs for each stage was calculated, and the computing server 20 was allocated on a stage-by-stage basis. However, in Example 2, the throughput of the xPU when assigning a task to a stage is calculated from the stage length for each xPU, and the xPU is allocated on a task-by-task basis.

[0070] When allocating xPUs on a task-by-task basis, the task allocation determination unit 12 allocates tasks to xPUs in such a way as to minimize the number of computing servers 20, based on the following conditions.

[0071] - For a given stage, the total throughput of the xPUs assigned tasks within that stage exceeds the throughput required by the stage. - For a given xPU, the throughput of the tasks assigned to that xPU does not exceed the xPU's throughput. - For a given xPU, the total memory consumption of the xPUs in the computing server is less than or equal to the memory capacity of that computing server. - For a given xPU, the total storage consumption of the xPUs in the computing server is less than or equal to the storage capacity of that computing server. - The "time from start to finish of data processing" for the entire application is less than or equal to the threshold value specified by the user.

[0072] The first condition integrates the constraints on stage overlap allocation and the number of CPUs in Example 1. Memory consumption and storage consumption are determined using the highest memory or storage consumption values ​​among the tasks assigned to that xPU. The time from the start to the end of data processing takes into account the stage length for each server.

[0073] The task placement determination unit 12 may solve the optimization using the above conditions as an integer programming problem. Alternatively, it may use heuristics or metaheuristics such as the Best-Fit algorithm or genetic algorithms.

[0074] [Effects] The conventional technology, "Energy Efficient VM Scheduling for Cloud Data Centers: Exact Allocation and Migration Algorithms," 2013 13th IEEE / ACM International Symposium on Cluster, Cloud, and Grid Computing, 2013, pp. 671-678, allocates Virtual Machines (VMs) to physical servers by formulating and solving a condition using integer programming that ensures the VM's resource consumption is less than or equal to the physical server's resource consumption. However, the conventional technology does not take into account the performance differences of physical servers, and in definite-delay computing, task execution time changes depending on the performance differences of physical servers. Therefore, if the conventional technology is applied as is, there is a risk that some tasks may not be able to keep up with the periodic processing, or that the overall processing time of the application may not meet the desired time.

[0075] Furthermore, since the performance of actual physical servers is not uniform, the stage length will vary depending on the deployment location, even for the same stage. For example, even if the number of CPU cores is the same, performance will differ depending on the CPU generation. When the stage length changes from server to server, the required number of xPUs changes, but conventional technology does not take server performance into account, so there is a risk that processing will not be completed within the cycle and the system will fail. When the stage length changes from deployment location, the overall processing time will vary depending on the deployment location of the stage or task, making it unpredictable.

[0076] In this disclosure, the stage length for each task on each xPU is determined, the required resources for each computing server 20 are individually analyzed to understand their characteristics, constraints are formulated based on the stage length and the characteristics of each computing server 20, and the computing server 20 on which each task is assigned is determined by finding the solution to an optimization problem that minimizes the number of computing servers 20 to be used.

[0077] As described above, the data processing system 1 comprises a management server 10 and multiple computing servers 20, and performs multiple processes in stages on the multiple computing servers 20 for data with a fixed arrival delay. The management server 10 determines which computing servers 20 will execute each stage by solving an optimization problem that minimizes the number of computing servers 20 to be used, under constraints formulated based on the stage length and resource capacity of each computing server 20 when the tasks of each stage are executed on each computing server 20. It then determines the timing for executing each task in each stage based on the execution order and stage length of each stage, and controls each task in each stage to be executed on the determined computing servers 20 at the determined timing. As a result, in data processing for data with a fixed arrival timing, no tasks will fall behind in the periodic processing, and the overall processing time can be kept within the desired time.

[0078] The management server 10 described above can be a general-purpose computer system, such as the one shown in Figure 15, which includes a central processing unit (CPU) 901, memory 902, storage 903, communication device 904, input device 905, and output device 906. In this computer system, the management server 10 is realized when the CPU 901 executes a predetermined program loaded onto the memory 902. This program can be recorded on a computer-readable non-temporary recording medium such as a magnetic disk, optical disk, or semiconductor memory, or it can be distributed via a network.

[0079] 1 Data Processing System 10 Management Server 11 Task Evaluation Unit 12 Task Placement Determination Unit 13 Task Assignment Unit 14 Task Execution Unit 20 Computing Server 30 Data Acquisition Server 40 Camera

Claims

1. A data processing method in which a computer performs a series of processes on multiple servers in stages on data whose delay has been determined and which arrives at each of multiple candidate servers from a sensor or edge device, the computer determines which server to execute each of the processes on based on the processing time when each of the processes is executed on each of the multiple servers and the amount of resources on each of the multiple servers, determines the timing for executing each of the processes based on the execution order of the processes and the processing time, controls each of the processes to be executed on the determined server at the determined timing, the data being video data of an intersection, and each of the processes being a process to decode an image from the video data, a process to detect a vehicle from the decoded image, a process to obtain information about the detected vehicle, and a process to evaluate the collision risk using the vehicle information.

2. A data processing method according to claim 1, wherein for each of the plurality of processes, the number of processing units required on each of the plurality of servers is determined from the processing time when the process is executed on each of the plurality of servers, and the number of required processing units is used as a constraint condition regarding the processing capacity of the servers.

3. A data processing method according to claim 1, wherein the throughput of each processing unit is determined based on the processing time when each of the plurality of processes is executed in each of the processing units provided in each of the plurality of servers, the constraint on the processing capacity of the servers is that the sum of the throughputs of the processing units to which the processes are assigned exceeds the throughput required by the processes, and each of the plurality of processes consists of at least one task executed for each data, and the server that executes each task is determined for each task.

4. A data processing method according to claim 1, wherein the processing time when each of the plurality of processes is executed on each of the plurality of servers is calculated for each of the plurality of processes by executing the process on one of the plurality of servers to measure a reference processing time, and calculating the processing time on each of the plurality of servers based on a performance indicator between the plurality of servers.

5. A data processing device that performs multiple processes in stages on multiple servers for data whose delay magnitude is determined and which arrives at each of multiple candidate servers from a sensor or edge device, the data processing device determines which server to execute each of the multiple processes on based on the processing time when each of the multiple processes is executed on each of the multiple servers and the amount of resources on each of the multiple servers, determines the timing for executing each of the multiple processes based on the execution order of the multiple processes and the processing time, and controls each of the multiple processes to be executed on the determined server at the determined timing, wherein the data is video data of an intersection, and each of the multiple processes is a process for decoding an image from the video data, a process for detecting a vehicle from the decoded image, a process for obtaining information on the detected vehicle, and a process for evaluating the collision risk using the vehicle information.

6. A program that causes a computer to execute the data processing method described in any one of claims 1 to 4.