Task-oriented multi-resource scheduling method and device, equipment and storage medium

By using a task-oriented multi-resource scheduling method, the application is split into a directed acyclic graph of tasks, and resource pre-allocation and path calculation are performed. This solves the problem of uneven resource utilization in traditional network architectures and achieves more efficient resource scheduling and improved user experience.

CN119003162APending Publication Date: 2024-11-22WUHAN POST & TELECOMM RES INST CO LTD
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202411031193.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-30
Publication Date
2024-11-22

AI Technical Summary

Technical Problem

In traditional network architectures, computing resources and connectivity resources are separated, which makes cloud computing resources a bottleneck, edge computing resources cannot be fully utilized, and overall QoS guarantees for applications with multiple resource requirements cannot be achieved. In addition, transmission resources are consumed in large quantities, resulting in a poor user experience.

Method used

By employing a task-oriented multi-resource scheduling method, tasks are split based on the current application's data flow input source, a directed acyclic graph of tasks is generated, resources are pre-allocated and paths are calculated, the estimated total latency is calculated, and computing nodes are determined for resource scheduling.

Benefits of technology

It improves the overall latency and QoS guarantee of applications, makes full use of network resources, enhances user experience and utilization of ubiquitous computing resources, and improves the speed and efficiency of multi-resource scheduling.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119003162A_ABST
    Figure CN119003162A_ABST
Patent Text Reader

Abstract

The invention discloses a task-oriented multi-resource scheduling method and device, equipment and a storage medium, and the method comprises the steps: carrying out the task splitting of a current application according to a data stream input source of the current application, and obtaining a task directed acyclic graph; performing resource pre-allocation on each task according to the task directed acyclic graph to obtain a resource pre-allocation result, and performing path calculation on each task according to the task directed acyclic graph to obtain a path allocation result from each data to each allocated resource; calculating a predicted total time delay of each task according to the resource pre-allocation result and the path allocation result, determining a computing node of each task according to the predicted total time delay, and performing resource scheduling of each task according to the computing node; according to the invention, appropriate resources can be allocated for applications with multi-resource requirements, network resources are fully utilized, the user experience is improved, the utilization rate of ubiquitous computing resources in a network is improved, and the speed and efficiency of task-oriented multi-resource scheduling are improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of network application computing resources, and in particular to a task-oriented multi-resource scheduling method, device, equipment and storage medium. Background Art

[0002] With the development of augmented reality (AR), virtual reality (VR), extended reality (XR), metaverse, artificial intelligence (AI) and even large-scale model applications, network applications have an increasing demand for computing resources. These complex types of applications no longer only require operators to provide network connection resources, but also require the integration of computing resources, storage resources and even data resources on top of high-speed connections.

[0003] For example, a simple VR application needs to transmit the video captured by the front-end camera back to the edge node at high speed, and then perform video analysis. It then needs to perform correlation analysis with the structured information of the person extracted from the database stored in a data center. After compression and rendering, it is sent to the user's mobile phone for display, and the user's latency experience must be very good throughout the process. This application involves not only the transmission of a large amount of video data, but also the analysis of video data, the extraction of data from the database, as well as correlation analysis, rendering processing and other processes.

[0004] The network infrastructure of the future will not only provide network connections. With the development of technologies such as edge computing and computing-network integration, the network infrastructure of the future can be seen as a highly distributed, multi-resource integrated infrastructure that tightly integrates computing resources, connection resources and data resources. Moreover, unlike traditional architectures, computing resources and storage resources are not all concentrated in data centers. The network architecture of the future has the characteristics of ubiquitous computing, ubiquitous intelligence, and ubiquitous connectivity. Therefore, the edge and end of the network can provide computing resources and even data resources. The computing resources or nodes from the end to the edge and then to the cloud center can carry a set of services or computing functions, which of course depends on the capabilities of these nodes themselves.

[0005] Faced with such changes in applications and network architecture, the traditional network application provision model and service quality assurance model face a series of problems. For complex applications with multiple resource requirements, the traditional resource provision model separates connection resources and computing resources. Network operators are only responsible for providing connection resources, while cloud operators are only responsible for providing computing resources. In this case, applications with large data volume computing can easily lead to cloud computing resources becoming a bottleneck, while some edge computing resources in the network cannot be fully utilized. Moreover, since computing and connection are provided separately, it is impossible to achieve overall service quality (QoS) guarantees for applications for users and applications.

[0006] With traditional network architecture and traditional business processing methods, network operators usually provide connection resources and cloud operators provide computing resources to jointly complete such services, and all data processing is completed in the cloud center. In this case, the cloud center will gradually become a bottleneck for computing resources and connection resources. The only solution is to continuously expand the computing resources of the cloud center and the connection resource bandwidth of the cloud center's export. However, the resources on the edge side of the network and distributed outside other cloud centers are not fully utilized. At the same time, a large amount of video resources are transmitted to the cloud center for processing, which will occupy a large amount of transmission resources, and the overall transmission delay is bound to be relatively high, reducing the user experience. Summary of the invention

[0007] The main purpose of the present invention is to provide a task-oriented multi-resource scheduling method, device, equipment and storage medium, aiming to solve the complex applications with multiple resource requirements in the prior art. The traditional resource processing method can easily lead to cloud computing resources becoming a bottleneck for large data volume computing applications, and edge computing resources cannot be fully utilized. The overall QoS guarantee of the application cannot be achieved, and appropriate resources cannot be allocated to applications with multiple resource requirements. It will occupy a large amount of transmission resources, have high latency, and have a poor user experience. Technical problems.

[0008] In a first aspect, the present invention provides a task-oriented multi-resource scheduling method, the task-oriented multi-resource scheduling method comprising the following steps:

[0009] Performing task splitting on the current application according to the data stream input source of the current application to obtain a directed acyclic graph of tasks;

[0010] Pre-allocating resources for each task according to the task directed acyclic graph to obtain a resource pre-allocation result, and calculating paths for each task according to the task directed acyclic graph to obtain a path allocation result from each data to each allocated resource;

[0011] The estimated total delay of each task is calculated according to the resource pre-allocation result and the path allocation result, the computing node of each task is determined according to the estimated total delay, and the resource scheduling of each task is performed according to the computing node.

[0012] Optionally, splitting the current application into tasks according to the data stream input source of the current application to obtain a directed acyclic graph of tasks includes:

[0013] Obtaining a data stream input source of the current application, and identifying the data stream input source according to a source IP address of the data stream input source;

[0014] Obtain required data and required computing resources corresponding to the current application, split the current application into tasks according to the identified data stream input source, the required data and the required computing resources, obtain different tasks, and generate a task directed acyclic graph according to the different tasks.

[0015] Optionally, pre-allocating resources for each task according to the task directed acyclic graph to obtain a resource pre-allocation result, and performing path calculation for each task according to the task directed acyclic graph to obtain a path allocation result from each data to each allocated resource, include:

[0016] Obtaining a data source IP address, a destination IP address, and a computing resource IP address of each task according to the task directed acyclic graph;

[0017] Perform resource allocation according to the data source IP address, the destination IP address and the computing resource IP address to obtain a resource pre-allocation result;

[0018] Path calculation is performed on each task according to the task directed acyclic graph to obtain a path allocation result from each data to each allocated resource.

[0019] Optionally, obtaining a data source IP address, a destination IP address, and a computing resource IP address of each task according to the task directed acyclic graph includes:

[0020] When the data stream input source is an external input source, the IP address of the task data input source corresponding to the current task is used as the data source IP address according to the task directed acyclic graph;

[0021] When the data stream input source is the output data stream generated by the previous task, the IP address of the computing node of the previous task is used as the data source IP address according to the task directed acyclic graph;

[0022] According to the task directed acyclic graph, the initial output IP address of each task is used as the destination IP address;

[0023] When a task has a subsequent dependent task, the computing resource IP address of the subsequent dependent task is calculated, and the computing resource IP address is used as the output IP address of the previous task.

[0024] Optionally, the calculating an estimated total delay of each task according to the resource pre-allocation result and the path allocation result, determining a computing node for each task according to the estimated total delay, and performing resource scheduling for each task according to the computing node includes:

[0025] Obtaining input data volume or computing resources from the resource pre-allocation result;

[0026] When the computing type of the current task is GPU, the value of dividing the input data amount by the number of floating point operations per second (FLOPS) of the GPU is used as the computing delay of the resource to be selected for the current task;

[0027] When the computing type of the current task is CPU, the value of the input data volume divided by DMIPS of the CPU is used as the resource computing delay to be selected for the current task;

[0028] Obtaining the input data volume or the input bandwidth value of the minimum bandwidth part in the input path from the path allocation result, and determining the input path delay according to the input bandwidth value;

[0029] Obtaining the output data volume or the output bandwidth value of the minimum bandwidth part in the output path from the path allocation result, and determining the output path delay according to the output bandwidth value;

[0030] Determine the estimated total delay of each task according to the resource calculation delay to be selected, the input path delay and the output path delay;

[0031] The computing node of each task is determined according to the estimated total delay, and the resource scheduling of each task is performed according to the computing node.

[0032] Optionally, the determining the estimated total delay of each task according to the resource calculation delay to be selected, the input path delay, and the output path delay includes:

[0033] The estimated total delay of each task is determined by calculating the delay of the resource to be selected, the input path delay and the output path delay through the following formula:

[0034] Estimated total delay = R*path delay of input path + W*computation delay of resources to be selected + R*path delay of output path

[0035] Among them, R is the weight of path delay, and W is the weight of calculation delay.

[0036] Optionally, determining a computing node for each task according to the estimated total delay, and performing resource scheduling for each task according to the computing node includes:

[0037] Select matching resources that match each task type from the preset computing resource list;

[0038] Selecting a computing resource with the smallest total delay from the matching resources as a target computing resource according to the estimated total delay;

[0039] Obtain the computing node corresponding to the target computing resource, and perform resource scheduling for each task according to the computing node.

[0040] In a second aspect, to achieve the above-mentioned purpose, the present invention further proposes a task-oriented multi-resource scheduling device, the task-oriented multi-resource scheduling device comprising:

[0041] A task splitting module is used to split the tasks of the current application according to the data stream input source of the current application to obtain a task directed acyclic graph;

[0042] A resource path allocation module is used to pre-allocate resources for each task according to the task directed acyclic graph, obtain resource pre-allocation results, calculate paths for each task according to the task directed acyclic graph, and obtain path allocation results from each data to each allocated resource;

[0043] The resource scheduling module is used to calculate the estimated total delay of each task according to the resource pre-allocation result and the path allocation result, determine the computing node of each task according to the estimated total delay, and perform resource scheduling for each task according to the computing node.

[0044] In the third aspect, to achieve the above-mentioned purpose, the present invention also proposes a task-oriented multi-resource scheduling device, which includes: a memory, a processor, and a task-oriented multi-resource scheduling program stored in the memory and executable on the processor, wherein the task-oriented multi-resource scheduling program is configured to implement the steps of the task-oriented multi-resource scheduling method as described above.

[0045] In a fourth aspect, in order to achieve the above-mentioned purpose, the present invention also proposes a storage medium, on which a task-oriented multi-resource scheduling program is stored, and when the task-oriented multi-resource scheduling program is executed by a processor, the steps of the task-oriented multi-resource scheduling method as described above are implemented.

[0046] The task-oriented multi-resource scheduling method proposed in the present invention performs task splitting on the current application according to the data stream input source of the current application to obtain a task directed acyclic graph; performs resource pre-allocation on each task according to the task directed acyclic graph to obtain a resource pre-allocation result; performs path calculation on each task according to the task directed acyclic graph to obtain a path allocation result from each data to each allocated resource; calculates the estimated total delay of each task according to the resource pre-allocation result and the path allocation result, determines the computing node of each task according to the estimated total delay, and performs resource scheduling for each task according to the computing node; and can allocate appropriate resources to applications with multiple resource requirements, by splitting complex applications into a series of tasks and selecting corresponding transmission paths and computing nodes by comprehensively considering the computing delay and path delay, thereby improving the overall delay and QoS guarantee of the application, making full use of network resources, improving user experience, improving the utilization rate of ubiquitous computing resources in the network, and improving the speed and efficiency of task-oriented multi-resource scheduling. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 A schematic diagram of the device structure of the hardware operating environment involved in the embodiment of the present invention;

[0048] Figure 2 It is a flowchart of the first embodiment of the task-oriented multi-resource scheduling method of the present invention;

[0049] Figure 3 It is a flowchart of a second embodiment of the task-oriented multi-resource scheduling method of the present invention;

[0050] Figure 4 It is a flowchart of a third embodiment of the task-oriented multi-resource scheduling method of the present invention;

[0051] Figure 5 It is a flowchart of a fourth embodiment of the task-oriented multi-resource scheduling method of the present invention;

[0052] Figure 6 It is a flowchart of a fifth embodiment of the task-oriented multi-resource scheduling method of the present invention;

[0053] Figure 7 It is a flowchart of a sixth embodiment of the task-oriented multi-resource scheduling method of the present invention;

[0054] Figure 8 A schematic diagram of the application of multiple resource requirements in the task-oriented multiple resource scheduling method of the present invention;

[0055] Fig. 9 It is a functional module diagram of the first embodiment of the task-oriented multi-resource scheduling device of the present invention.

[0056] The realization of the purpose, functional features and advantages of the present invention will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0057] It should be understood that the specific embodiments described herein are only used to explain the present invention, and are not used to limit the present invention.

[0058] The solution of the embodiment of the present invention is mainly: by splitting the current application into tasks according to the data flow input source of the current application, a task directed acyclic graph is obtained; according to the task directed acyclic graph, resources are pre-allocated to each task to obtain the resource pre-allocation result, and the path calculation is performed on each task according to the task directed acyclic graph to obtain the path allocation result from each data to each allocated resource; according to the resource pre-allocation result and the path allocation result, the estimated total delay of each task is calculated, the computing node of each task is determined according to the estimated total delay, and the resource scheduling of each task is performed according to the computing node; it is possible to allocate appropriate resources to applications with multiple resource requirements, by splitting complex applications into a series of tasks, and Comprehensive consideration of computing delay and path delay to select the corresponding transmission path and computing node improves the overall delay and QoS guarantee of the application, fully utilizes network resources, improves user experience, improves the utilization rate of ubiquitous computing resources in the network, and improves the speed and efficiency of task-oriented multi-resource scheduling. It solves the complex applications with multiple resource requirements in the existing technology. The traditional resource processing method can easily lead to cloud computing resources becoming a bottleneck for large data volume computing applications, and edge computing resources cannot be fully utilized. The overall QoS guarantee of the application cannot be achieved, and appropriate resources cannot be allocated to applications with multiple resource requirements. It will occupy a large amount of transmission resources, have high latency, and have poor user experience. Technical problems.

[0059] Reference Figure 1 , Figure 1 The figure is a schematic diagram of the device structure of the hardware operating environment involved in the embodiment of the present invention.

[0060] like Figure 1As shown, the device may include: a processor 1001, such as a CPU, a communication bus 1002, a user interface 1003, a network interface 1004, and a memory 1005. Among them, the communication bus 1002 is used to realize the connection and communication between these components. The user interface 1003 may include a display screen (Display), an input unit such as a keyboard (Keyboard), and the optional user interface 1003 may also include a standard wired interface and a wireless interface. The network interface 1004 may optionally include a standard wired interface and a wireless interface (such as a Wi-Fi interface). The memory 1005 may be a high-speed RAM memory, or it may be a stable memory (Non-Volatile Memory), such as a disk memory. The memory 1005 may optionally be a storage device independent of the aforementioned processor 1001.

[0061] Those skilled in the art will understand that Figure 1 The device structure shown in the figure does not constitute a limitation of the device, and may include more or less components than shown in the figure, or combine certain components, or arrange the components differently.

[0062] like Figure 1 As shown, the memory 1005 as a storage medium may include an operating device, a network communication module, a user interface module, and a task-oriented multi-resource scheduling program.

[0063] The device of the present invention calls the task-oriented multi-resource scheduling program stored in the memory 1005 through the processor 1001, and performs the following operations:

[0064] Performing task splitting on the current application according to the data stream input source of the current application to obtain a directed acyclic graph of tasks;

[0065] Pre-allocating resources for each task according to the task directed acyclic graph to obtain a resource pre-allocation result, and calculating paths for each task according to the task directed acyclic graph to obtain a path allocation result from each data to each allocated resource;

[0066] The estimated total delay of each task is calculated according to the resource pre-allocation result and the path allocation result, the computing node of each task is determined according to the estimated total delay, and the resource scheduling of each task is performed according to the computing node.

[0067] The device of the present invention calls the task-oriented multi-resource scheduling program stored in the memory 1005 through the processor 1001, and further performs the following operations:

[0068] Obtaining a data stream input source of the current application, and identifying the data stream input source according to a source IP address of the data stream input source;

[0069] Obtain required data and required computing resources corresponding to the current application, split the current application into tasks according to the identified data stream input source, the required data and the required computing resources, obtain different tasks, and generate a task directed acyclic graph according to the different tasks.

[0070] The device of the present invention calls the task-oriented multi-resource scheduling program stored in the memory 1005 through the processor 1001, and further performs the following operations:

[0071] Obtaining a data source IP address, a destination IP address, and a computing resource IP address of each task according to the task directed acyclic graph;

[0072] Perform resource allocation according to the data source IP address, the destination IP address and the computing resource IP address to obtain a resource pre-allocation result;

[0073] Path calculation is performed on each task according to the task directed acyclic graph to obtain a path allocation result from each data to each allocated resource.

[0074] The device of the present invention calls the task-oriented multi-resource scheduling program stored in the memory 1005 through the processor 1001, and further performs the following operations:

[0075] When the data stream input source is an external input source, the IP address of the task data input source corresponding to the current task is used as the data source IP address according to the task directed acyclic graph;

[0076] When the data stream input source is the output data stream generated by the previous task, the IP address of the computing node of the previous task is used as the data source IP address according to the task directed acyclic graph;

[0077] According to the task directed acyclic graph, the initial output IP address of each task is used as the destination IP address;

[0078] When a task has a subsequent dependent task, the computing resource IP address of the subsequent dependent task is calculated, and the computing resource IP address is used as the output IP address of the previous task.

[0079] The device of the present invention calls the task-oriented multi-resource scheduling program stored in the memory 1005 through the processor 1001, and further performs the following operations:

[0080] Obtaining input data volume or computing resources from the resource pre-allocation result;

[0081] When the computing type of the current task is GPU, the value of dividing the input data amount by the number of floating point operations per second (FLOPS) of the GPU is used as the computing delay of the resource to be selected for the current task;

[0082] When the computing type of the current task is CPU, the value of the input data volume divided by DMIPS of the CPU is used as the resource computing delay to be selected for the current task;

[0083] Obtaining the input data volume or the input bandwidth value of the minimum bandwidth part in the input path from the path allocation result, and determining the input path delay according to the input bandwidth value;

[0084] Obtaining the output data volume or the output bandwidth value of the minimum bandwidth part in the output path from the path allocation result, and determining the output path delay according to the output bandwidth value;

[0085] Determine the estimated total delay of each task according to the resource calculation delay to be selected, the input path delay and the output path delay;

[0086] The computing node of each task is determined according to the estimated total delay, and the resource scheduling of each task is performed according to the computing node.

[0087] The device of the present invention calls the task-oriented multi-resource scheduling program stored in the memory 1005 through the processor 1001, and further performs the following operations:

[0088] The estimated total delay of each task is determined by calculating the delay of the resource to be selected, the input path delay and the output path delay through the following formula:

[0089] Estimated total delay = R*path delay of input path + W*computation delay of resources to be selected + R*path delay of output path

[0090] Among them, R is the weight of path delay, and W is the weight of calculation delay.

[0091] The device of the present invention calls the task-oriented multi-resource scheduling program stored in the memory 1005 through the processor 1001, and further performs the following operations:

[0092] Select matching resources that match each task type from the preset computing resource list;

[0093] Selecting a computing resource with the smallest total delay from the matching resources as a target computing resource according to the estimated total delay;

[0094] Obtain the computing node corresponding to the target computing resource, and perform resource scheduling for each task according to the computing node.

[0095] Through the above scheme, this embodiment obtains a directed acyclic graph of tasks by splitting the current application according to the data flow input source of the current application; pre-allocates resources for each task according to the directed acyclic graph of tasks to obtain resource pre-allocation results; calculates paths for each task according to the directed acyclic graph of tasks to obtain path allocation results from each data to each allocated resource; calculates the estimated total delay of each task according to the resource pre-allocation results and the path allocation results, determines the computing nodes of each task according to the estimated total delay, and schedules resources for each task according to the computing nodes; and can allocate appropriate resources to applications with multiple resource requirements, by splitting complex applications into a series of tasks and selecting corresponding transmission paths and computing nodes by comprehensively considering computing delay and path delay, thereby improving the overall delay and QoS guarantee of the application, making full use of network resources, improving user experience, improving the utilization rate of ubiquitous computing resources in the network, and improving the speed and efficiency of task-oriented multi-resource scheduling.

[0096] Based on the above hardware structure, an embodiment of the task-oriented multi-resource scheduling method of the present invention is proposed.

[0097] Reference Figure 2 , Figure 2 It is a flowchart diagram of the first embodiment of the task-oriented multi-resource scheduling method of the present invention.

[0098] In a first embodiment, the task-oriented multi-resource scheduling method comprises the following steps:

[0099] Step S10: split the current application into tasks according to the data stream input source of the current application to obtain a directed acyclic graph of tasks.

[0100] It should be noted that by splitting the tasks according to the data stream input source of the current application and splitting the complex application into a series of task combinations, a task directed acyclic graph can be formed.

[0101] It can be understood that the task here refers to a combination of partial functions of an application that can run independently in the network and requires certain computing resources or network resources to complete; an application can be decomposed into a combination of several tasks; a task includes the computing resources for executing the task, the input data source of the task, and the output stream of the task.

[0102] Step S20: pre-allocate resources for each task according to the task directed acyclic graph to obtain a resource pre-allocation result; perform path calculation for each task according to the task directed acyclic graph to obtain a path allocation result from each data to each allocated resource.

[0103] It should be understood that by allocating computing resources and selecting paths for tasks, that is, pre-allocating resources for each task, the resource pre-allocation result can be obtained, and by performing path calculation for each task, the path allocation result of each data to the corresponding allocated resources can be obtained.

[0104] Step S30: Calculate the estimated total delay of each task according to the resource pre-allocation result and the path allocation result, determine the computing node of each task according to the estimated total delay, and perform resource scheduling for each task according to the computing node.

[0105] It can be understood that the estimated total delay of each task is calculated based on the resource pre-allocation results and the path allocation results, and then the computing node of each task is determined according to the estimated total delay, and the resources of each task are scheduled according to the computing node, which improves the overall user experience of complex applications and improves the utilization rate of ubiquitous computing resources in the network.

[0106] Through the above scheme, this embodiment obtains a directed acyclic graph of tasks by splitting the current application according to the data flow input source of the current application; pre-allocates resources for each task according to the directed acyclic graph of tasks to obtain resource pre-allocation results; calculates paths for each task according to the directed acyclic graph of tasks to obtain path allocation results from each data to each allocated resource; calculates the estimated total delay of each task according to the resource pre-allocation results and the path allocation results, determines the computing nodes of each task according to the estimated total delay, and schedules resources for each task according to the computing nodes; and can allocate appropriate resources to applications with multiple resource requirements, by splitting complex applications into a series of tasks and selecting corresponding transmission paths and computing nodes by comprehensively considering computing delay and path delay, thereby improving the overall delay and QoS guarantee of the application, making full use of network resources, improving user experience, improving the utilization rate of ubiquitous computing resources in the network, and improving the speed and efficiency of task-oriented multi-resource scheduling.

[0107] Further, Figure 3 FIG. 1 is a flow chart of a second embodiment of the task-oriented multi-resource scheduling method of the present invention. Figure 3 As shown, a second embodiment of the task-oriented multi-resource scheduling method of the present invention is proposed based on the first embodiment. In this embodiment, step S10 specifically includes the following steps:

[0108] Step S11: Acquire the data stream input source of the current application, and identify the data stream input source according to the source IP of the data stream input source.

[0109] It should be noted that after obtaining the data stream input source of the current application, the data stream input source can be identified according to the source IP of the data stream input source, and the input data source of the task is identified by the source IP address of the input data source.

[0110] In the specific implementation, in order to facilitate the calculation of routing paths on the network, these important information are all identified by IP addresses. The computing resources that execute the task are identified by the IP address of the computing resource. The IP address is the address when the resource is announced to the outside world on the network, not the internal address. The input data source of the task is identified by the source IP address of the input data source, and the output stream of the task is identified by the output destination IP address.

[0111] Step S12: obtain the required data and required computing resources corresponding to the current application, split the current application into tasks according to the identified data stream input source, the required data and the required computing resources, obtain different tasks, and generate a task directed acyclic graph according to the different tasks.

[0112] It is understandable that different applications correspond to different required data and required computing resources. According to the identified data flow input source, the required data and the required computing resources, the current application can be split into tasks to obtain different tasks, thereby obtaining different task generation task directed acyclic graphs.

[0113] In the specific implementation, task splitting is carried out according to the application's input source and the required data and computing resources. When an application has multiple data input sources, the principle of "split if possible" is adopted. If several data input sources must be processed together, these data input sources are placed in one task and one computing resource for processing, otherwise different data input sources are processed with different tasks as much as possible. This allows for as much parallel processing as possible, improving processing efficiency and performance. If a data input source requires more complex or large-scale processing, it needs to be split into different tasks, and the split tasks should be processed on adjacent nodes as much as possible to reduce transmission overhead.

[0114] This embodiment adopts the above scheme, by obtaining the data stream input source of the current application, identifying the data stream input source according to the source IP of the data stream input source; obtaining the required data and required computing resources corresponding to the current application, splitting the current application into tasks according to the identified data stream input source, the required data and the required computing resources, obtaining different tasks, generating a task directed acyclic graph according to the different tasks, being able to quickly obtain the task directed acyclic graph, being able to allocate appropriate resources to applications with multiple resource requirements, and by splitting complex applications into a series of tasks and comprehensively considering the computing delay and path delay to select the corresponding transmission path and computing node, thereby improving the overall delay and QoS guarantee of the application, making full use of network resources, and improving the speed and efficiency of task-oriented multi-resource scheduling.

[0115] Further, Figure 4 FIG. 1 is a flow chart of a third embodiment of the task-oriented multi-resource scheduling method of the present invention. Figure 4 As shown, a third embodiment of the task-oriented multi-resource scheduling method of the present invention is proposed based on the first embodiment. In this embodiment, step S20 specifically includes the following steps:

[0116] Step S21, obtaining the data source IP address, destination IP address and computing resource IP address of each task according to the task directed acyclic graph.

[0117] It should be noted that each task has an IP address of a data input source, an IP address of an output destination, and a computing resource address of a computing task. According to the task directed acyclic graph, the data source IP address, destination IP address and computing resource IP address of each task can be obtained. The split tasks all have an input data source, so each task has an IP address of a data input source, an IP address of an output destination, and a computing resource address of a computing task.

[0118] Step S22: Allocate resources according to the data source IP address, the destination IP address and the computing resource IP address to obtain a resource pre-allocation result.

[0119] It can be understood that resource allocation is performed according to the data source IP address, the destination IP address and the computing resource IP address to obtain a resource pre-allocation result.

[0120] Step S23: Calculate the path of each task according to the task directed acyclic graph to obtain the path allocation result from each data to each allocated resource.

[0121] It should be understood that, according to the task directed acyclic graph, path calculation can be performed on each task, thereby obtaining a path allocation result from each data to each allocated resource.

[0122] In the specific implementation, the task management function is mainly implemented by the task management module. The main function of the task management function is to realize the decomposition of tasks, the scheduling of tasks and the scheduling of task computing resources. In order to realize the decomposition of tasks, the application must provide the task management module with the application requirements, including the source address, the destination address, and the data calculation required.

[0123] Through the above scheme, this embodiment obtains the data source IP address, destination IP address and computing resource IP address of each task according to the task directed acyclic graph; performs resource allocation according to the data source IP address, the destination IP address and the computing resource IP address to obtain resource pre-allocation results; performs path calculation for each task according to the task directed acyclic graph to obtain path allocation results from each data to each allocated resource, thereby obtaining a task directed acyclic graph, allocating appropriate resources for applications with multiple resource requirements, splitting complex applications into a series of tasks, and selecting corresponding transmission paths and computing nodes by comprehensively considering computing delay and path delay, thereby improving the overall delay and QoS guarantee of the application, making full use of network resources, and improving the speed and efficiency of task-oriented multi-resource scheduling.

[0124] Further, Figure 5 FIG. 4 is a flow chart of a fourth embodiment of a task-oriented multi-resource scheduling method according to the present invention. Figure 5 As shown, a fourth embodiment of the task-oriented multi-resource scheduling method of the present invention is proposed based on the first embodiment. In this embodiment, step S21 specifically includes the following steps:

[0125] Step S211: When the data stream input source is an external input source, the IP address of the task data input source corresponding to the current task is used as the data source IP address according to the task directed acyclic graph.

[0126] It should be noted that if the data input source of the task is a clear external input source, then the IP address of the task data input source is the IP address of the external data input source, that is, when the data flow input source is an external input source, the IP address can be used as the data source IP address according to the task directed acyclic graph.

[0127] Step S212: When the data stream input source is the output data stream generated by the previous task, the IP address of the computing node of the previous task is used as the data source IP address according to the task directed acyclic graph.

[0128] It can be understood that if there is no external data input source, but an output data stream generated by the previous task, then the IP address of the data input source of the task is the IP address of the computing node of the previous task, that is, when the data stream input source is the output data stream generated by the previous task, the IP address of the computing node of the previous task can be used as the data source IP address according to the task directed acyclic graph.

[0129] Step S213: according to the task directed acyclic graph, the initial output IP address of each task is used as the destination IP address.

[0130] It should be understood that the destination IP address is the output IP address of the task, that is, the initial output IP address of the task is determined as the destination IP address of the application.

[0131] Step S214: When a task has a subsequent dependent task, the computing resource IP address of the subsequent dependent task is calculated, and the computing resource IP address is used as the output IP address of the previous task.

[0132] It is understandable that when a task has a subsequent dependent task, after calculating the IP address of the computing resource of the subsequent task, the IP address can be used as the output IP address of the previous task.

[0133] This embodiment adopts the above scheme. When the data stream input source is an external input source, the IP address of the task data input source corresponding to the current task is used as the data source IP address according to the task directed acyclic graph; when the data stream input source is the output data stream generated by the previous task, the IP address of the computing node of the previous task is used as the data source IP address according to the task directed acyclic graph; the initial output IP address of each task is used as the destination IP address according to the task directed acyclic graph; when a task has a subsequent dependent task, the computing resource IP address of the subsequent dependent task is calculated, and the computing resource IP address is used as the output IP address of the previous task. Thus, a task directed acyclic graph can be obtained, and appropriate resources can be allocated to applications with multiple resource requirements. By splitting complex applications into a series of tasks and comprehensively considering the computing delay and path delay to select the corresponding transmission path and computing node, the overall delay and QoS guarantee of the application are improved, network resources are fully utilized, and the speed and efficiency of task-oriented multi-resource scheduling are improved.

[0134] Further, Figure 6 FIG. 5 is a flow chart of a fifth embodiment of a task-oriented multi-resource scheduling method according to the present invention. Figure 6 As shown, a fifth embodiment of the task-oriented multi-resource scheduling method of the present invention is proposed based on the first embodiment. In this embodiment, step S30 specifically includes the following steps:

[0135] Step S31: Obtain input data volume or computing resources from the resource pre-allocation result.

[0136] It should be noted that the corresponding input data volume or computing resources can be obtained from the resource pre-allocation result.

[0137] Step S32: When the calculation type of the current task is GPU, the value of dividing the input data amount by the number of floating-point operations per second (FLOPS) of the GPU is used as the calculation delay of the resource to be selected for the current task.

[0138] It is understandable that when the calculation type of the current task is GPU, the value of the input data volume divided by the floating point operations per second (FLOPS) of the graphics processing unit (GPU) is used as the calculation delay of the resources to be selected for the current task.

[0139] Step S33: When the computing type of the current task is CPU, the value of dividing the input data volume by the CPU's DMIPS is used as the resource calculation delay to be selected for the current task.

[0140] It should be understood that when the calculation type of the current task is central processing unit (CPU), the input data volume divided by the CPU's Dhrystone Million Instructions executed Per Second (DMIPS) value can be used as the calculation delay of the resources to be selected for the current task. DMIPS is mainly used to measure integer computing power.

[0141] In the specific implementation, in order to realize the scheduling of computing resources, computing resources need to be registered with the task management module. The registration content includes: the IP address of the computing resource, the computing power of the computing resource, and the status of the computing resource; the computing power of the computing resource is expressed in FLOPS (whether to distinguish between CPU and GPU). The status of the computing resource mainly reports to the task management body whether there is a task running; when the task management body assigns tasks to computing resources, it will record the matching status of computing resources and tasks; the task management module has a complete list of available computing resources in this domain; when the application requests the network to provide services, it will first provide its own business needs to the network; at this time, the task management module first splits the tasks for the application.

[0142] Step S34: Obtain the input data volume or the input bandwidth value of the minimum bandwidth part in the input path from the path allocation result, and determine the input path delay according to the input bandwidth value.

[0143] It is understandable that the input data volume or the input bandwidth value of the minimum bandwidth part in the input path can be obtained from the path allocation result, and then the input path delay can be determined according to the input bandwidth value.

[0144] Step S35: Obtain the output data volume or the output bandwidth value of the minimum bandwidth part in the output path from the path allocation result, and determine the output path delay according to the output bandwidth value.

[0145] It should be understood that the output data volume or the output bandwidth value of the minimum bandwidth part in the output path can be obtained from the path allocation result, and then the output path delay can be determined according to the output bandwidth value.

[0146] In a specific implementation, the estimation method of the path delay for transmitting input data to the resource to be selected is: input data volume / input path bandwidth; the input path bandwidth selects the bandwidth value of the part with the smallest bandwidth on the input path.

[0147] The calculation delay of the resource to be selected is estimated by: input data volume / FLOPS or DMIPS of the computing resource; if the computing type is GPU, the input data volume is divided by the FLOPS of the GPU, and if it is CPU, the input data volume is divided by the DMIPS of the CPU. Of course, the calculation delay will be affected by many factors in practice, such as cache size, operating system, etc. For the convenience of calculation, a simple method of dividing the input data volume by the FLOPS or DMIPS of the calculation is used here for calculation.

[0148] The path delay of the output path is: output data volume / output path bandwidth; the output path bandwidth value selects the bandwidth value of the minimum bandwidth part on the output path.

[0149] Step S36: Determine the estimated total delay of each task according to the resource calculation delay to be selected, the input path delay, and the output path delay.

[0150] It can be understood that the estimated total delay of each task can be determined according to the calculation delay of the resource to be selected, the input path delay and the output path delay.

[0151] Furthermore, the step S36 specifically includes the following steps:

[0152] The estimated total delay of each task is determined by calculating the delay of the resource to be selected, the input path delay and the output path delay through the following formula:

[0153] Estimated total delay = R*path delay of input path + W*computation delay of resources to be selected + R*path delay of output path

[0154] Among them, R is the weight of path delay, and W is the weight of calculation delay.

[0155] It should be noted that for each input data stream source of the application: according to the computing type and IP address of the data input source of the task, try to select a computing node with a small expected total delay, that is, select a node with a transmission delay and a computing delay as small as possible; the expected total delay of the task is composed of three parts: the path delay of transmitting the input data to the resource to be selected; the computing delay of the resource to be selected; and the path delay from the computing resource to the output destination output path, which is expressed by the formula as above, where R represents the weight of the path delay, and W represents the weight of the computing delay.

[0156] In the specific implementation, it can be configured according to the actual situation; the rules of the configuration strategy are as follows: if the user's calculation is relatively complex and is expected to take a long time, in this case, the calculation delay is the main consideration, then the weight W of the calculation delay can be increased, and R can be reduced; if the calculation is relatively simple and the transmission delay is the main time-consuming factor, the transmission delay can be taken into consideration, R can be increased, and the weight W of the calculation delay can be reduced; the reason for increasing W and R is to avoid the selection bias that occurs when the path delay and the calculation delay are not in the same order of magnitude.

[0157] Step S37: determine the computing node of each task according to the estimated total delay, and perform resource scheduling for each task according to the computing node.

[0158] It should be understood that based on the data of the above three delays, the expected total delay of the task is calculated, and then the computing resource with the smallest expected total delay is selected as the computing node of the task, that is, the computing node of each task can be determined based on the expected total delay, and the resources of each task are scheduled based on the computing node.

[0159] This embodiment adopts the above scheme, by obtaining the input data volume or computing resources from the resource pre-allocation result; when the computing type of the current task is GPU, the value of the input data volume divided by the number of floating-point operations per second FLOPS of the GPU is used as the resource calculation delay to be selected for the current task; when the computing type of the current task is CPU, the value of the input data volume divided by the CPU's 1 million instructions per second DMIPS is used as the resource calculation delay to be selected for the current task; the input bandwidth value of the minimum bandwidth part in the input data volume or the input path is obtained from the path allocation result, and the input path delay is determined according to the input bandwidth value; the output data volume or the bandwidth in the output path is obtained from the path allocation result. The output bandwidth value of the minimum part is determined, and the output path delay is determined according to the output bandwidth value; the estimated total delay of each task is determined according to the calculation delay of the resource to be selected, the input path delay and the output path delay; the computing node of each task is determined according to the estimated total delay, and the resource scheduling of each task is performed according to the computing node; it is possible to allocate appropriate resources for applications with multiple resource requirements, split complex applications into a series of tasks, and select corresponding transmission paths and computing nodes by comprehensively considering the calculation delay and path delay, thereby improving the overall delay and QoS guarantee of the application, making full use of network resources, improving user experience, improving the utilization rate of ubiquitous computing resources in the network, and improving the speed and efficiency of task-oriented multi-resource scheduling.

[0160] Further, Figure 7 FIG. 6 is a flow chart of a sixth embodiment of a task-oriented multi-resource scheduling method according to the present invention. Figure 7 As shown, a sixth embodiment of the task-oriented multi-resource scheduling method of the present invention is proposed based on the first embodiment. In this embodiment, step S37 specifically includes the following steps:

[0161] Step S371: Select matching resources that match each task type from a preset computing resource list.

[0162] It should be noted that for the data input source of each task, computing resources are selected from the computing resource list of the task management module. The selection strategy is to first select computing resources that match the computing type, and then select computing resources with the smallest estimated total delay within the range of matching resources; for example, if the computing request is a GPU computing type, then select the GPU type first in the computing resource list, and if the computing request is a CPU type, then select the CPU type in the computing resource list.

[0163] It is understandable that, because the path delay needs to be calculated, the input path from the input source to the computing resource node and the output path from the computing resource node to the output destination must be calculated for each computing resource list; the calculation method of the input path and the output path can adopt a general routing protocol, such as Open Shortest Path First (OSPF) or Border Gateway Protocol (BGP).

[0164] Step S372: Select a computing resource with the smallest total delay from the matching resources as a target computing resource according to the estimated total delay.

[0165] It should be understood that, according to the estimated total delay, a computing resource with the smallest total delay can be selected from the matching resources as the target computing resource.

[0166] In the specific implementation, if an input source is divided into several tasks, the first task is processed according to the above steps, and the subsequent tasks use the output of the first task as the input source of the current task for estimation, that is, the computing node of the previous task is used as the IP address of the input source of the current task, and the data volume output by the previous task is used as the input data volume of the current task to calculate the expected delay of the current task, and exclude the nodes in the above steps; still according to: expected delay of input source = path delay of input path + computing delay of resources to be selected + path delay of output path. Find the computing node and path with the smallest expected delay.

[0167] Step S373: Obtain the computing node corresponding to the target computing resource, and perform resource scheduling for each task according to the computing node.

[0168] It should be understood that the computing node corresponding to the target computing resource is obtained, and the resource scheduling of each task is performed according to the computing node.

[0169] It can be understood that once a computing resource is selected, the task scheduling module will mark the computing resource and monitor the status of the resource through the heartbeat mechanism; when the computing task is completed, the computing node reports the task completion status; input sources that have no dependencies can be selected in parallel; after the application is split into tasks in this way, a chain of interdependent tasks is formed; and parallel tasks form different chains, and finally the task chains converge to the destination address of the application, and all tasks of the application finally form a directed acyclic graph, and the final destination points to the destination address of the application.

[0170] In the specific implementation, Figure 8 As shown, Figure 8This is a schematic diagram of the application of multiple resource requirements in the task-oriented multi-resource scheduling method of the present invention, see Figure 8 ,The solid arrows represent the directed acyclic graph formed by the tasks, the dotted arrows represent the computing resources allocated to the tasks,and the actual transmission paths.,1.1.1.1, 2.2.2.2, 3.3.3.3, 4.4.4.4, 5.5.5.5, 6.6.6.6 and 10.10.10.10 are IP address examples.,The complex applications are divided into tasks according to the data flow input source of the applications,and then the tasks are selected according to the dependency or parallel relationship of the data flow,to perform data calculation and transmission, making full use of network resources, thereby trying to meet the overall QoS of the users, while improving the utilization of each distributed resource in the network.

[0171] It is understandable that since the video data source has a large amount of data, it can be processed if it is close to the edge side of the data source. This not only reduces the delay in data transmission, but also improves the utilization rate of computing resources on the edge of the network by processing at the edge. At the same time, for data analysis and extraction work in the database, since it can be processed in parallel with the video data source, it is best to place it in different nodes or cloud centers for parallel processing. This obviously improves the overall latency of the application and makes full use of various computing resources distributed in the network.

[0172] It should be understood that this figure is only a simple example for reference, and not all network connections and computing resources are drawn. This embodiment splits tasks based on the input data source of the application, splits complex applications into a combination of a series of tasks, and forms a directed acyclic graph of tasks; then allocates computing resources and selects paths for the tasks; when selecting computing resources, a weighted configurable estimated total delay estimation method is used, which can improve the overall user experience of complex applications and improve the utilization rate of ubiquitous computing resources in the network. Of course, there can also be other settings for increasing or decreasing related units, or other deformations can be made according to actual application conditions, and this embodiment does not limit this.

[0173] This embodiment adopts the above scheme, by selecting matching resources that match each task type from a preset computing resource list; selecting the computing resource with the smallest total delay from the matching resources as the target computing resource according to the expected total delay; obtaining the computing node corresponding to the target computing resource, and performing resource scheduling for each task according to the computing node; it is possible to allocate appropriate resources for applications with multiple resource requirements, make full use of network resources, improve user experience, improve the utilization rate of ubiquitous computing resources in the network, and improve the speed and efficiency of task-oriented multi-resource scheduling.

[0174] Accordingly, the present invention further provides a task-oriented multi-resource scheduling device.

[0175] Reference Fig. 9 , Fig. 9 It is a functional module diagram of the first embodiment of the task-oriented multi-resource scheduling device of the present invention.

[0176] In a first embodiment of the task-oriented multi-resource scheduling device of the present invention, the task-oriented multi-resource scheduling device comprises:

[0177] The task splitting module 10 is used to split the tasks of the current application according to the data stream input source of the current application to obtain a task directed acyclic graph.

[0178] The resource path allocation module 20 is used to pre-allocate resources for each task according to the task directed acyclic graph to obtain resource pre-allocation results, and to calculate paths for each task according to the task directed acyclic graph to obtain path allocation results from each data to each allocated resource.

[0179] The resource scheduling module 30 is used to calculate the estimated total delay of each task according to the resource pre-allocation result and the path allocation result, determine the computing node of each task according to the estimated total delay, and perform resource scheduling for each task according to the computing node.

[0180] The task splitting module 10 is also used to obtain the data stream input source of the current application, identify the data stream input source according to the source IP of the data stream input source; obtain the required data and required computing resources corresponding to the current application, split the tasks of the current application according to the identified data stream input source, the required data and the required computing resources, obtain different tasks, and generate a task directed acyclic graph according to the different tasks.

[0181] The resource path allocation module 20 is also used to obtain the data source IP address, destination IP address and computing resource IP address of each task according to the task directed acyclic graph; perform resource allocation according to the data source IP address, the destination IP address and the computing resource IP address to obtain resource pre-allocation results; perform path calculation for each task according to the task directed acyclic graph to obtain path allocation results from each data to each allocated resource.

[0182] The resource path allocation module 20 is also used to use the IP address of the task data input source corresponding to the current task as the data source IP address according to the task directed acyclic graph when the data flow input source is an external input source; use the IP address of the computing node of the previous task as the data source IP address according to the task directed acyclic graph when the data flow input source is the output data flow generated by the previous task; use the initial output IP address of each task as the destination IP address according to the task directed acyclic graph; and when a task has a subsequent dependent task, calculate the computing resource IP address of the subsequent dependent task, and use the computing resource IP address as the output IP address of the previous task.

[0183] The resource scheduling module 30 is further used to obtain input data volume or computing resources from the resource pre-allocation result; when the computing type of the current task is GPU, the value of the input data volume divided by the number of floating-point operations per second FLOPS of the GPU is used as the computing delay of the resources to be selected for the current task; when the computing type of the current task is CPU, the value of the input data volume divided by the 1 million instructions per second DMIPS of the CPU is used as the computing delay of the resources to be selected for the current task; obtain the input bandwidth value of the input data volume or the minimum bandwidth part in the input path from the path allocation result, and determine the input path delay according to the input bandwidth value; obtain the output bandwidth value of the output data volume or the minimum bandwidth part in the output path from the path allocation result, and determine the output path delay according to the output bandwidth value; determine the estimated total delay of each task according to the computing delay of the resources to be selected, the input path delay and the output path delay; determine the computing node of each task according to the estimated total delay, and perform resource scheduling for each task according to the computing node.

[0184] The resource scheduling module 30 is further used to calculate and determine the estimated total delay of each task according to the resource calculation delay to be selected, the input path delay and the output path delay by the following formula:

[0185] Estimated total delay = R*path delay of input path + W*computation delay of resources to be selected + R*path delay of output path

[0186] Among them, R is the weight of path delay, and W is the weight of calculation delay.

[0187] The resource scheduling module 30 is also used to select matching resources that match each task type from a preset computing resource list; select the computing resource with the smallest total delay from the matching resources as the target computing resource according to the expected total delay; obtain the computing node corresponding to the target computing resource, and perform resource scheduling for each task according to the computing node.

[0188] The steps implemented by each functional module of the task-oriented multi-resource scheduling device may refer to the various embodiments of the task-oriented multi-resource scheduling method of the present invention, and will not be described in detail here.

[0189] In addition, an embodiment of the present invention further provides a storage medium, on which a task-oriented multi-resource scheduling program is stored. When the task-oriented multi-resource scheduling program is executed by a processor, the following operations are implemented:

[0190] Performing task splitting on the current application according to the data stream input source of the current application to obtain a directed acyclic graph of tasks;

[0191] Pre-allocating resources for each task according to the task directed acyclic graph to obtain a resource pre-allocation result, and calculating paths for each task according to the task directed acyclic graph to obtain a path allocation result from each data to each allocated resource;

[0192] The estimated total delay of each task is calculated according to the resource pre-allocation result and the path allocation result, the computing node of each task is determined according to the estimated total delay, and the resource scheduling of each task is performed according to the computing node.

[0193] Furthermore, when the task-oriented multi-resource scheduling program is executed by the processor, the following operations are also implemented:

[0194] Obtaining a data stream input source of the current application, and identifying the data stream input source according to a source IP address of the data stream input source;

[0195] Obtain required data and required computing resources corresponding to the current application, split the current application into tasks according to the identified data stream input source, the required data and the required computing resources, obtain different tasks, and generate a task directed acyclic graph according to the different tasks.

[0196] Furthermore, when the task-oriented multi-resource scheduling program is executed by the processor, the following operations are also implemented:

[0197] Obtaining a data source IP address, a destination IP address, and a computing resource IP address of each task according to the task directed acyclic graph;

[0198] Perform resource allocation according to the data source IP address, the destination IP address and the computing resource IP address to obtain a resource pre-allocation result;

[0199] Path calculation is performed on each task according to the task directed acyclic graph to obtain a path allocation result from each data to each allocated resource.

[0200] Furthermore, when the task-oriented multi-resource scheduling program is executed by the processor, the following operations are also implemented:

[0201] When the data stream input source is an external input source, the IP address of the task data input source corresponding to the current task is used as the data source IP address according to the task directed acyclic graph;

[0202] When the data stream input source is the output data stream generated by the previous task, the IP address of the computing node of the previous task is used as the data source IP address according to the task directed acyclic graph;

[0203] According to the task directed acyclic graph, the initial output IP address of each task is used as the destination IP address;

[0204] When a task has a subsequent dependent task, the computing resource IP address of the subsequent dependent task is calculated, and the computing resource IP address is used as the output IP address of the previous task.

[0205] Furthermore, when the task-oriented multi-resource scheduling program is executed by the processor, the following operations are also implemented:

[0206] Obtaining input data volume or computing resources from the resource pre-allocation result;

[0207] When the computing type of the current task is GPU, the value of dividing the input data amount by the number of floating point operations per second (FLOPS) of the GPU is used as the computing delay of the resource to be selected for the current task;

[0208] When the computing type of the current task is CPU, the value of the input data volume divided by DMIPS of the CPU is used as the resource computing delay to be selected for the current task;

[0209] Obtaining the input data volume or the input bandwidth value of the minimum bandwidth part in the input path from the path allocation result, and determining the input path delay according to the input bandwidth value;

[0210] Obtaining the output data volume or the output bandwidth value of the minimum bandwidth part in the output path from the path allocation result, and determining the output path delay according to the output bandwidth value;

[0211] Determine the estimated total delay of each task according to the resource calculation delay to be selected, the input path delay and the output path delay;

[0212] The computing node of each task is determined according to the estimated total delay, and the resource scheduling of each task is performed according to the computing node.

[0213] Furthermore, when the task-oriented multi-resource scheduling program is executed by the processor, the following operations are also implemented:

[0214] The estimated total delay of each task is determined by calculating the delay of the resource to be selected, the input path delay and the output path delay through the following formula:

[0215] Estimated total delay = R*path delay of input path + W*computation delay of resources to be selected + R*path delay of output path

[0216] Among them, R is the weight of path delay, and W is the weight of calculation delay.

[0217] Furthermore, when the task-oriented multi-resource scheduling program is executed by the processor, the following operations are also implemented:

[0218] Select matching resources that match each task type from the preset computing resource list;

[0219] Selecting a computing resource with the smallest total delay from the matching resources as a target computing resource according to the estimated total delay;

[0220] Obtain the computing node corresponding to the target computing resource, and perform resource scheduling for each task according to the computing node.

[0221] Those skilled in the art will appreciate that all or part of the steps in the above-mentioned implementation method can be accomplished by instructing related hardware through a program, and the program is stored in a storage medium, including a number of instructions for enabling a device (which may be a single-chip microcomputer, chip, etc.) or a processor to execute all or part of the steps of the method described in each implementation method of the present application; and the aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, and other media that can store program codes.

[0222] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the existence of other identical elements in the process, method, article or device including the element.

[0223] The serial numbers of the above embodiments of the present invention are only for description and do not represent the advantages or disadvantages of the embodiments.

[0224] The above are only preferred embodiments of the present invention, and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A task-oriented multi-resource scheduling method, characterized in that: The task-oriented multi-resource scheduling method comprises: Splitting the current application into tasks according to the data stream input source of the current application to obtain a directed acyclic graph of tasks; Pre-allocating resources for each task according to the task directed acyclic graph to obtain a resource pre-allocation result, and calculating paths for each task according to the task directed acyclic graph to obtain a path allocation result from each data to each allocated resource; The estimated total delay of each task is calculated according to the resource pre-allocation result and the path allocation result, the computing node of each task is determined according to the estimated total delay, and the resource scheduling of each task is performed according to the computing node.

2. The task-oriented multi-resource scheduling method according to claim 1, characterized in that: The step of splitting the current application into tasks according to the data stream input source of the current application to obtain a directed acyclic graph of tasks includes: Obtaining a data stream input source of the current application, and identifying the data stream input source according to a source IP address of the data stream input source; Obtain required data and required computing resources corresponding to the current application, split the current application into tasks according to the identified data stream input source, the required data and the required computing resources, obtain different tasks, and generate a task directed acyclic graph according to the different tasks.

3. The task-oriented multi-resource scheduling method according to claim 1, characterized in that: The pre-allocation of resources to each task according to the task directed acyclic graph to obtain a resource pre-allocation result, and the path calculation of each task according to the task directed acyclic graph to obtain a path allocation result from each data to each allocated resource include: Obtaining a data source IP address, a destination IP address, and a computing resource IP address of each task according to the task directed acyclic graph; Perform resource allocation according to the data source IP address, the destination IP address and the computing resource IP address to obtain a resource pre-allocation result; Path calculation is performed on each task according to the task directed acyclic graph to obtain a path allocation result from each data to each allocated resource.

4. The task-oriented multi-resource scheduling method according to claim 3, characterized in that: The obtaining of the data source IP address, the destination IP address and the computing resource IP address of each task according to the task directed acyclic graph includes: When the data stream input source is an external input source, the IP address of the task data input source corresponding to the current task is used as the data source IP address according to the task directed acyclic graph; When the data stream input source is the output data stream generated by the previous task, the IP address of the computing node of the previous task is used as the data source IP address according to the task directed acyclic graph; According to the task directed acyclic graph, the initial output IP address of each task is used as the destination IP address; When a task has a subsequent dependent task, the computing resource IP address of the subsequent dependent task is calculated, and the computing resource IP address is used as the output IP address of the previous task.

5. The task-oriented multi-resource scheduling method according to claim 1, characterized in that: The calculating the estimated total delay of each task according to the resource pre-allocation result and the path allocation result, determining the computing node of each task according to the estimated total delay, and performing resource scheduling of each task according to the computing node includes: Obtaining input data volume or computing resources from the resource pre-allocation result; When the computing type of the current task is GPU, the value of dividing the input data amount by the number of floating point operations per second (FLOPS) of the GPU is used as the computing delay of the resource to be selected for the current task; When the computing type of the current task is CPU, the value of the input data volume divided by DMIPS of the CPU is used as the resource computing delay to be selected for the current task; Obtaining the input data volume or the input bandwidth value of the minimum bandwidth part in the input path from the path allocation result, and determining the input path delay according to the input bandwidth value; Obtaining the output data volume or the output bandwidth value of the minimum bandwidth part in the output path from the path allocation result, and determining the output path delay according to the output bandwidth value; Determine the estimated total delay of each task according to the resource calculation delay to be selected, the input path delay and the output path delay; The computing node of each task is determined according to the estimated total delay, and the resource scheduling of each task is performed according to the computing node.

6. The task-oriented multi-resource scheduling method according to claim 5, characterized in that: The determining the estimated total delay of each task according to the resource calculation delay to be selected, the input path delay and the output path delay includes: The estimated total delay of each task is determined by calculating the delay of the resource to be selected, the input path delay and the output path delay through the following formula: Estimated total delay = R*path delay of input path + W*computation delay of resources to be selected + R*path delay of output path Among them, R is the weight of path delay, and W is the weight of calculation delay.

7. The task-oriented multi-resource scheduling method according to claim 5, characterized in that: The determining of a computing node for each task according to the estimated total delay, and performing resource scheduling for each task according to the computing node, includes: Select matching resources that match each task type from the preset computing resource list; Selecting a computing resource with the smallest total delay from the matching resources as a target computing resource according to the estimated total delay; Obtain the computing node corresponding to the target computing resource, and perform resource scheduling for each task according to the computing node.

8. A task-oriented multi-resource scheduling device, characterized in that: The task-oriented multi-resource scheduling device comprises: A task splitting module is used to split the tasks of the current application according to the data stream input source of the current application to obtain a task directed acyclic graph; A resource path allocation module is used to pre-allocate resources for each task according to the task directed acyclic graph, obtain resource pre-allocation results, calculate paths for each task according to the task directed acyclic graph, and obtain path allocation results from each data to each allocated resource; The resource scheduling module is used to calculate the estimated total delay of each task according to the resource pre-allocation result and the path allocation result, determine the computing node of each task according to the estimated total delay, and perform resource scheduling for each task according to the computing node.

9. A task-oriented multi-resource scheduling device, characterized in that: The task-oriented multi-resource scheduling device includes: a memory, a processor, and a task-oriented multi-resource scheduling program stored in the memory and executable on the processor, wherein the task-oriented multi-resource scheduling program is configured to implement the steps of the task-oriented multi-resource scheduling method as described in any one of claims 1 to 7.

10. A storage medium, characterized in that: The storage medium stores a task-oriented multi-resource scheduling program, which, when executed by a processor, implements the steps of the task-oriented multi-resource scheduling method according to any one of claims 1 to 7.

Citation Information

Cited By

  • Template-based resource rendering method and apparatus, and computing device

    CN119865564A