WO2022088659A1 - Procédé et appareil de planification de ressources, dispositif électronique, support de stockage et produit-programme - Google Patents

Procédé et appareil de planification de ressources, dispositif électronique, support de stockage et produit-programme Download PDF

Info

Publication number
WO2022088659A1
WO2022088659A1 PCT/CN2021/095292 CN2021095292W WO2022088659A1 WO 2022088659 A1 WO2022088659 A1 WO 2022088659A1 CN 2021095292 W CN2021095292 W CN 2021095292W WO 2022088659 A1 WO2022088659 A1 WO 2022088659A1
Authority
WO
WIPO (PCT)
Prior art keywords
gpus
gpu
resource scheduling
virtual
screening
Prior art date
Application number
PCT/CN2021/095292
Other languages
English (en)
Chinese (zh)
Inventor
霍明明
张炜
陈界
朴元奎
陈宇恒
Original Assignee
北京市商汤科技开发有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 北京市商汤科技开发有限公司 filed Critical 北京市商汤科技开发有限公司
Priority to KR1020217037982A priority Critical patent/KR20220058844A/ko
Publication of WO2022088659A1 publication Critical patent/WO2022088659A1/fr

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F9/00—Arrangements for program control, e.g. control units
    • G06F9/06—Arrangements for program control, e.g. control units using stored programs, i.e. using an internal store of processing equipment to receive or retain programs
    • G06F9/46—Multiprogramming arrangements
    • G06F9/50—Allocation of resources, e.g. of the central processing unit [CPU]
    • G06F9/5005—Allocation of resources, e.g. of the central processing unit [CPU] to service a request
    • G06F9/5027—Allocation of resources, e.g. of the central processing unit [CPU] to service a request the resource being a machine, e.g. CPUs, Servers, Terminals
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F9/00—Arrangements for program control, e.g. control units
    • G06F9/06—Arrangements for program control, e.g. control units using stored programs, i.e. using an internal store of processing equipment to receive or retain programs
    • G06F9/46—Multiprogramming arrangements
    • G06F9/50—Allocation of resources, e.g. of the central processing unit [CPU]
    • G06F9/5061—Partitioning or combining of resources
    • G06F9/5077—Logical partitioning of resources; Management or configuration of virtualized resources
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
    • Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management

Definitions

  • the present application relates to the technical field of artificial intelligence, and in particular, to a resource scheduling method and apparatus, electronic equipment, storage medium and program product.
  • Machine learning can be divided into two categories, one of which is to make computers simulate human learning behaviors to acquire new knowledge or skills, and to reorganize existing knowledge structures to continuously improve their performance; the other is Gain hidden, valid, understandable knowledge from massive amounts of data.
  • the above-mentioned second type of machine learning requires data, algorithms and computing power to realize; among them, computing power needs the support of some computer hardware resources such as Graphics Processing Unit (GPU), so that the computing power can better exert the algorithm and the role of data.
  • GPU Graphics Processing Unit
  • the scheduling device receives a resource scheduling request, it will perform resource scheduling among the GPUs of all these physical machines.
  • the scheduling methods are all random scheduling, which makes it impossible to precisely control the use of resources.
  • Embodiments of the present application provide a resource scheduling method and apparatus, electronic equipment, storage medium and program product, so as to precisely control the use of resources and improve resource scheduling efficiency and resource utilization.
  • an embodiment of the present application provides a resource scheduling method, including: receiving a resource scheduling request for a GPU in a graphics processor GPU cluster, where the resource scheduling request includes grouping information of the GPU to be requested, and the The grouping information is determined according to the task type of the task processing request corresponding to the resource scheduling request; according to the grouping information of the GPU to be requested, all GPUs in the GPU cluster are matched with the grouping information of the GPU to be requested.
  • GPU obtaining a matching result, where the matching result includes at least one target GPU corresponding to the grouping information of the GPU to be requested; and returning the matching result.
  • each GPU includes at least one virtual GPU (Virtual Graphics Processing Unit, vGPU), and the resource scheduling request further includes computing parameters and quantity of vGPU;
  • the grouping information of GPUs after matching the GPUs with the grouping information of the GPUs to be requested among all the GPUs in the GPU cluster, the method further includes: according to the calculation parameters and the number of the vGPUs, in the matching result Screen the vGPUs that satisfy the computing parameters and quantity of the vGPU; return the vGPUs that satisfy the computing parameters and quantity of the vGPU.
  • screening the vGPU that satisfies the resource scheduling request in the matching result includes: screening the matching result that satisfies the vGPU Calculate the vGPU of the parameter to obtain a first screening result; in the first screening result, screen the vGPU resources that meet the requirement of the number of vGPUs.
  • the computing parameters include at least one of the following: computing power and video memory; the matching results are screened for vGPUs that satisfy the computing parameters to obtain a first screening result, including: Obtain the priority corresponding to the computing power and the video memory of each vGPU in each of the target GPUs; if the priority of the computing power is greater than the priority of the video memory, then in each of the target GPUs Screening vGPUs that meet the computing power requirements of the vGPUs requested by the resource scheduling request to obtain a second screening result; screening vGPUs that meet the memory requirements of the vGPUs requested by the resource scheduling in the second screening results, obtaining the first screening results Filter results.
  • the computing parameters include at least one of the following: computing power and video memory; the matching results are screened for vGPUs that satisfy the computing parameters to obtain a first screening result, including: Obtain the priority corresponding to the computing power and the video memory of each vGPU in each of the target GPUs; if the priority of the computing power is less than the priority of the video memory, then in each of the target GPUs Screening vGPUs that meet the video memory requirements of the vGPUs requested by the resource scheduling request to obtain a third screening result; screening vGPUs that meet the computing power requirements of the vGPUs requested by the resource scheduling in the third screening results, obtaining the first screening results Filter results.
  • screening for vGPU resources that meet the requirement of the number of vGPUs includes: if the number of vGPUs in the first screening result is greater than the number of vGPUs The number of vGPU resources required by the resource scheduling request, then in the first screening result, the number of vGPU resources corresponding to the number of vGPU resources required by the resource scheduling request is selected according to the computing parameters in ascending order; If the number of vGPUs in the first screening result is equal to the number of vGPU resources required by the resource scheduling request, return the first screening result; if the number of vGPUs in the first screening result is less than The number of vGPU resources required by the resource scheduling request, then a prompt message that the screening result is empty is returned.
  • the resource scheduling request includes the task type of the task processing request corresponding to the resource scheduling request; vGPUs in different GPUs have tags corresponding to the tags, and the tags corresponding to the vGPUs are determined by the task type of the task processing request corresponding to the resource scheduling request; the method further includes: matching the task processing request corresponding to the resource scheduling request according to the task type of the task processing request corresponding to the resource scheduling request at least one tag corresponding to the task type; the vGPU corresponding to the at least one tag is used as the matching result.
  • an embodiment of the present application provides a resource scheduling apparatus, including: a receiving module configured to receive a resource scheduling request for a GPU in a graphics processing unit GPU cluster, where the resource scheduling request includes grouping information of the GPU to be requested, so The grouping information of the GPUs to be requested is determined according to the task type of the task processing request corresponding to the resource scheduling request; the first matching module is configured to, according to the grouping information of the GPUs to be requested, in all GPUs of the GPU cluster Matching the GPU with the grouping information of the GPU to be requested, and obtaining a matching result, the matching result includes at least one target GPU corresponding to the grouping information of the GPU to be requested; the first returning module is configured to return the matching result.
  • each of the GPUs includes at least one vGPU
  • the resource scheduling request further includes computing parameters and quantities of the vGPUs
  • the apparatus further includes: a screening module configured to The calculation parameters and the number of vGPUs are selected in the matching result to satisfy the calculation parameters and the number of vGPUs of the vGPU; the second return module is configured to return the vGPUs that satisfy the calculation parameters and the number of the vGPUs.
  • the screening module includes: a first screening unit, configured to screen vGPUs that satisfy the calculation parameters in the matching results, to obtain a first screening result; a second screening unit, configured to In the first screening result, the vGPU resources that meet the requirement of the number of vGPUs are screened.
  • the computing parameters include at least one of the following: computing power and video memory; the first screening unit is configured to obtain the computing power of each vGPU in each of the target GPUs The priority corresponding to the video memory; if the priority of the computing power is greater than the priority of the video memory, the vGPU that satisfies the computing power requirement of the vGPU of the resource scheduling request is screened in each of the target GPUs, Obtaining a second screening result; screening vGPUs that meet the video memory requirements of the vGPUs requested by the resource scheduling in the second screening results to obtain the first screening results.
  • the computing parameters include at least one of the following: computing power and video memory; the first screening unit is configured to obtain the computing power of each vGPU in each of the target GPUs The priority corresponding to the video memory; if the priority of the computing power is less than the priority of the video memory, then filter the vGPUs that meet the video memory requirements of the vGPU requested by the resource scheduling request in each of the target GPUs, and obtain The third screening result: screening the vGPU that meets the computing power requirement of the vGPU requested by the resource scheduling in the third screening result to obtain the first screening result.
  • the second screening unit is configured to, if the number of vGPUs in the first screening result is greater than the number of vGPU resources required by the resource scheduling request, perform the filtering in the In the first screening result, a number of vGPU resources corresponding to the number of vGPU resources required by the resource scheduling request are selected according to the computing parameters in ascending order.
  • the second screening unit is configured to, if the number of vGPUs in the first screening result is equal to the number of vGPU resources required by the resource scheduling request, return the The first screening result.
  • the second screening unit is configured to return a screening result if the number of vGPUs in the first screening result is less than the number of vGPU resources required by the resource scheduling request Empty prompt message.
  • the resource scheduling request includes the task type of the task processing request corresponding to the resource scheduling request; vGPUs in different GPUs have tags corresponding to the tags, and the tags corresponding to the vGPUs are The task type of the task processing request corresponding to the resource scheduling request is determined; the apparatus further includes: a second matching module configured to match the resource scheduling request according to the task type of the task processing request corresponding to the resource scheduling request. requesting the corresponding task to process at least one tag corresponding to the task type of the request; and using the vGPU corresponding to the at least one tag as the matching result.
  • an embodiment of the present application provides an electronic device, including: a memory;
  • the computer program is stored in the memory and configured to be executed by the processor to implement the method of the first aspect.
  • an embodiment of the present application provides a computer-readable storage medium on which a computer program is stored, and the computer program is executed by a processor to implement the method described in the first aspect.
  • an embodiment of the present disclosure provides a computer program product, including computer-readable code, and when the computer-readable code is executed in an electronic device, a processor in the electronic device executes the first the method described in the aspect.
  • the resource scheduling method and device, electronic device, storage medium, and program product provided by the embodiments of the present application receive a resource scheduling request for a GPU in a graphics processor GPU cluster, where the resource scheduling request includes grouping information of the GPU to be requested, and the resource scheduling request is to be requested.
  • the grouping information of the requested GPU is determined according to the task type of the task processing request corresponding to the resource scheduling request, and then according to the grouping information of the GPU to be requested, the GPUs with the grouping information of the GPU to be requested are matched among all GPUs in the GPU cluster; The matching result of at least one target GPU corresponding to the grouping information of the GPU to be requested is included.
  • the resource scheduling request includes the grouping information of the GPU to be requested, and the grouping information of the GPU to be requested is determined according to the task type of the task processing request corresponding to the resource scheduling request, when performing GPU resource scheduling, it can be matched according to the grouping information. to the corresponding GPU, so as to achieve more fine-grained resource scheduling and precisely control the use of GPU.
  • FIG. 1 is an application scenario diagram provided by an embodiment of the present application
  • FIG. 2 is a flowchart of a resource scheduling method provided by an embodiment of the present application
  • FIG. 3 is a schematic diagram of grouping GPUs of a physical machine according to an embodiment of the present application.
  • 4A is a schematic diagram of a single online prediction task provided by an embodiment of the present application.
  • 4B is a schematic diagram of multiple online prediction tasks provided by an embodiment of the present application.
  • FIG. 5 is a flowchart of a resource scheduling method provided by another embodiment of the present application.
  • FIG. 6 is a schematic diagram of a vGPU in a physical machine provided by an embodiment of the present application.
  • FIG. 7 is a schematic structural diagram of a resource scheduling apparatus provided by an embodiment of the present application.
  • FIG. 8 is a block diagram of an electronic device provided by an embodiment of the present application.
  • FIG. 1 is an application scenario diagram provided by an embodiment of the present application.
  • the application scenario includes: a user terminal 11 , an AI algorithm device 12 , a scheduling device 13 and a GPU cluster 14 ; the user terminal 11 may at least include electronic devices such as smart phones, Ipads, and personal computers.
  • GPU cluster 14 is a computer cluster that includes a plurality of computer nodes, where each computer node is equipped with one or more GPUs.
  • the user can submit a task processing request through the user terminal 11, such as a model training task, an online prediction task, etc. in an AI scenario, and the task processing request submitted by the user will be sent to the AI algorithm device 12, and the AI algorithm device 12 generates a resource scheduling request according to the task processing request, and sends the resource scheduling request to the scheduling device 13, and the scheduling device 13 further performs resource scheduling in the GPU cluster 14 according to the resource scheduling request, and returns the resource scheduling result to the AI algorithm device.
  • a task processing request through the user terminal 11, such as a model training task, an online prediction task, etc. in an AI scenario
  • the task processing request submitted by the user will be sent to the AI algorithm device 12
  • the AI algorithm device 12 generates a resource scheduling request according to the task processing request, and sends the resource scheduling request to the scheduling device 13, and the scheduling device 13 further performs resource scheduling in the GPU cluster 14 according to the resource scheduling request, and returns the resource scheduling result to the AI algorithm device.
  • the scheduling device 13 performs resource scheduling in the GPU cluster 14 according to the resource scheduling request, that is, allocates the resources required by the task processing request to each GPU in the GPU cluster 14, so that each GPU completes the allocated task and finally realizes the User-submitted tasks handle the processing of requests.
  • the minimum scheduling unit for resources in the prior art is a physical machine.
  • the prior art can only implement scheduling of physical machines.
  • the embodiment of the present application adopts the following technical solutions: the minimum scheduling unit (physical machine) of the GPU cluster 14 is divided into finer granularity, and according to the type of tasks that the GPU cluster 14 needs to process, the GPU cluster 14 All GPUs are tagged, so that when receiving a task processing request from the user, the GPU corresponding to the tag can be screened according to the task type corresponding to the task processing request, so as to achieve more fine-grained resource scheduling and precise control of GPU. use.
  • AI algorithm device 12 may be an independent device or device, or may be a module or component integrated in the user terminal 11 , which is not specifically limited in this embodiment.
  • the embodiments of the present application can be applied to all artificial intelligence scenarios, such as the field of intelligent video analysis.
  • FIG. 2 is a flowchart of a resource scheduling method provided by an embodiment of the present application. As shown in FIG. 2, the resource scheduling method includes the following steps S201 to S203:
  • Step S201 receiving a resource scheduling request for GPUs in the graphics processor GPU cluster 14 .
  • the execution body of this embodiment is the scheduling device 13 shown in FIG. 1 .
  • the scheduling device 13 receives a resource scheduling request from the AI algorithm device 12, the resource scheduling request includes grouping information of the GPU to be requested, and the grouping information of the GPU to be requested is determined according to the task type of the task processing request corresponding to the resource scheduling request.
  • the task type can be divided according to the purpose of the task.
  • the task types include model training and online prediction.
  • the grouping information of the GPU to be requested includes model training grouping information and online prediction grouping information.
  • the user submits a task processing request whose task type is model training to the AI algorithm device 12, and the AI algorithm device 12 generates a resource scheduling request according to the task processing request, and determines the task type to be processed according to the task type corresponding to the task processing request.
  • the grouping information of the requested GPU is the model training grouping information.
  • the grouping information of the GPUs to be requested may be specified by the AI algorithm device 12. If the AI algorithm device 12 does not specify the grouping information of the GPUs to be requested, the default is that all GPUs in the GPU cluster 14 are usable.
  • Step S202 according to the grouping information of the GPU to be requested, match the GPUs with the grouping information of the GPU to be requested among all the GPUs in the GPU cluster 14 to obtain a matching result.
  • the matching result includes at least one target GPU corresponding to the grouping information of the GPU to be requested.
  • the GPU cluster 14 includes multiple physical machines, and each physical machine includes multiple GPUs. In this embodiment, before step S201, all GPUs in the GPU cluster 14 need to be grouped. To group, the usage of the GPU can be determined according to the task type corresponding to the task processing request that needs to be executed by the GPU cluster 14 . The following takes a physical machine as an example to describe the GPU grouping process in detail:
  • FIG. 3 is a schematic diagram of grouping GPUs of a physical machine according to an embodiment of the present application.
  • the physical machine is a physical machine 31 with 9 cards (including a physical machine with 9 GPU cards), which are numbered from 0 to 8 cards, assuming that the user plans Model training and online prediction tasks are carried out on this physical machine at the same time, and it is planned to use 0 to 3 cards for model training and 4 to 8 cards for online prediction, then the grouping information of 0 to 3 cards can be set.
  • the grouping information of 4 cards to 8 cards is set as the online prediction grouping information.
  • the model training grouping information can be marked as label A (Label-A), and the online prediction grouping information can be marked as label B (Label-B).
  • all GPUs in the GPU cluster 14 can be represented as a list, and each GPU corresponds to grouping information. Taking a physical machine including 9 GPU cards as an example, the list of all GPUs is in the form of a list. Table 1 below:
  • GPU card number group information 0 card model training 1 card model training 2 cards model training 3 cards model training 4 cards Online prediction 5 cards Online prediction 6 cards Online prediction 7 cards Online prediction 8 cards Online prediction
  • the GPU grouping information carried in the resource scheduling request is the model training grouping information, and then the GPUs of cards 0 to 3 will be matched.
  • the GPU grouping information is online prediction grouping information, then it will match the GPU of 4 to 8 cards.
  • GPU cluster 14 includes physical machine 1, physical machine 2, and physical machine 3; wherein, physical machine 1 includes GPU0, GPU1, and GPU2; physical machine 2 includes GPU3, GPU4, and GPU5; physical machine 3 includes GPU6, GPU7, and GPU8; Then the GPU1 and GPU2 in the physical machine 1, the GPU5 in the physical machine 2, and the GPU8 in the physical machine 3 can be divided into the same group.
  • each group can be regarded as a resource pool, which can realize logical isolation between resources (GPUs) and resources (GPUs).
  • Step S203 returning a matching result.
  • the matching result includes at least one target GPU corresponding to the grouping information of the GPU to be requested.
  • the matching result may be expressed in the form of a list, and after obtaining the above matching result, the scheduling device 13 generates a GPU list according to the matching result, and returns the GPU list to the AI algorithm device 12 .
  • the form of the GPU list can refer to the following Table 2:
  • the embodiment of the present application receives a resource scheduling request for GPUs in the graphics processor GPU cluster 14, where the resource scheduling request includes grouping information of the GPU to be requested, and the grouping information of the GPU to be requested is based on the task processing request corresponding to the resource scheduling request If the task type is determined, then according to the grouping information of the GPUs to be requested, the GPUs with the grouping information of the GPUs to be requested are matched among all the GPUs in the GPU cluster 14; match results. Since the resource scheduling request includes the grouping information of the GPU to be requested, and the grouping information of the GPU to be requested is determined according to the task type of the task processing request corresponding to the resource scheduling request, when performing GPU resource scheduling, it can be matched according to the grouping information. to the corresponding GPU, so as to achieve more fine-grained resource scheduling and precisely control the use of GPU.
  • the present application can improve the controllability of resource scheduling of AI algorithm applications in vGPU mode. For example, for an 8-card GPU machine, among them, 0 to 3 cards use vGPU mode for resource allocation; 4 to 7 cards use non-vGPU mode for resource allocation. In the prior art, the selection of GPU is random, and it is impossible to control the scheduling of applications in the vGPU mode to cards 0 to 3. However, using the resource scheduling method of the embodiment of the present application, by labeling cards 0 to 3 with vGPU labels, when applying for resources, the scheduling device 13 allocates resources among the GPUs labelled with vGPU, so that it can be very accurate. control the use of resources.
  • the resource scheduling method of the present application can also satisfy the isolation and classified use of GPU resources on a single GPU machine, and maximize the utilization of resources that meet different requirements.
  • the user's resources are tight, and there is only one 8-card GPU machine, but they want to perform model training tasks and online prediction tasks on this 8-card GPU machine at the same time, and they can be well isolated from each other. Influence. In this scenario, it is usually used by static designation, but static designation is time-consuming and labor-intensive.
  • the AI algorithm device 12 informs the scheduling device 13 to use the GPU card resource corresponding to the tag, and selects which GPU card is implemented by the scheduling device 13 without user participation, which improves the usability to a certain extent.
  • the above embodiment introduces the GPU-level resource scheduling process.
  • a task can be implemented by only one GPU card.
  • more GPUs are required. card to meet the concurrent needs of multitasking.
  • a city restricts motor vehicles, and many cameras are set up on the city road to monitor the driving of vehicles on the city road. When a vehicle is detected to violate the traffic restriction rules, the camera will take a picture of the vehicle and send it. Notify the information to the owner, prompting the owner to pay the fine.
  • the camera captures the image it needs to identify the vehicle in the image, then circle the vehicle in the image with a rectangular frame, and then identify the license plate information.
  • an online prediction task needs to be used.
  • Figure 4A if the image captured by the camera includes a car, then there is only one online prediction task, and only one GPU card is needed at this time.
  • Figure 4B the images captured by the camera often include multiple vehicles, and then there are multiple online prediction tasks. If GPU-level resource scheduling is used, the multiple online prediction tasks will be allocated to multiple GPUs, so that GPU resources cannot be fully utilized, resulting in a waste of expensive GPU resources.
  • each GPU can also be divided into smaller scheduling units, that is, using virtual machine technology to virtualize each GPU in Figure 1 to obtain multiple vGPUs, and then assign multiple parallel online prediction tasks to different On vGPU, multiple tasks share the same GPU, thereby improving the resource utilization of a single GPU.
  • the present application can also implement resource scheduling in a GPU sharing scenario, and the implementation manner is as follows:
  • FIG. 5 is a flowchart of a resource scheduling method provided by another embodiment of the present application.
  • the resource scheduling request may further include the computing parameters of the vGPU and the number of vGPUs, where the number of vGPUs is N, and N is a positive integer greater than 0.
  • the resource scheduling method provided by this embodiment includes the following steps:
  • Step S501 according to the calculation parameters of the vGPU and the number of the vGPU, filter the vGPU that satisfies the resource scheduling request in the matching result.
  • this step may be to screen the vGPUs that satisfy the computing parameters and quantity requirements of the vGPUs corresponding to the resource scheduling request in the matching result.
  • FIG. 6 is a schematic diagram of a vGPU in a physical machine according to an embodiment of the present application. As shown in FIG. 6 , each GPU can be further divided into multiple vGPUs (as shown by the circles in FIG. 6 ). It should be noted that each GPU in FIG. 6 includes 3 vGPUs for exemplary illustration only, and does not limit the number of vGPUs.
  • step S501 is executed after the matching result is obtained in step S202.
  • the matching result in this embodiment may be represented by a GPU list, and the matching result may also include computing parameters such as computing power (vcore) and/or video memory (vmemory) of each vGPU of each target GPU, where the computing power of the vGPU is Refers to the computing power of the vGPU.
  • computing power vcore
  • vmemory video memory
  • step S501 may further include the following steps:
  • Step S501a Screen the vGPUs satisfying the calculation parameters in the matching result to obtain the first screening result.
  • the vGPUs that satisfy the computing parameters may be shown in the form of a list, and the vGPU list includes at least one vGPU that satisfies the computing parameters.
  • the computing parameters required by the task processing request submitted by the user include computing power, and the computing power of each vGPU requested by the resource scheduling request is: 3.5, 3.0, 5.2, and 6.1, then the computing power that satisfies the resource scheduling request in Table 3
  • Required vGPUs (first screening results) include: vGPU-2, vGPU-4, vGPU-8, vGPU-9, vGPU-10, vGPU-11, vGPU-12.
  • the first screening result can also be given in the form of a list, and its form is as follows in Table 4:
  • the computing parameters required by the task processing request submitted by the user include video memory, and the video memory of each vGPU requested by the resource scheduling request is: 6GB, 8GB, 8GB, and 6GB respectively; then the vGPUs that satisfy the resource scheduling request include: vGPU-3, vGPU-6, vGPU-7, vGPU-8, vGPU-10, vGPU-11, vGPU-12.
  • vGPUs that satisfy resource scheduling requests include: vGPU-2, vGPU-3, vGPU-4, vGPU-6, vGPU-7, vGPU-8, vGPU-9, vGPU-10, vGPU-11, vGPU- 12.
  • Step S501b in the first screening result, screening vGPU resources that meet the requirement of the number of vGPUs in the resource scheduling request.
  • this step is to filter out N vGPUs in the first screening result.
  • 4 vGPUs may be randomly selected in Table 4.
  • the first 4 vGPUs may also be selected in Table 4 in ascending order of computing power or video memory.
  • the vGPUs that meet the computing power include: vGPU-2, vGPU-4, vGPU-8, vGPU-9, vGPU-10, vGPU-11, vGPU- 12.
  • 4 vGPUs may be randomly selected from the 7 vGPUs, that is, vGPUs that satisfy the computing parameters and quantity of the vGPUs.
  • Step S502 returning to the vGPU that satisfies the resource scheduling request.
  • step S502 may be to return the vGPU that meets the requirements for computing parameters of the vGPU and the requirements for the number of vGPUs to the AI algorithm device 12 .
  • the first time is to filter according to the grouping information.
  • the GPU cluster 14 is very large, many GPUs that are not within the screening range can be filtered out through the grouping information.
  • the second screening process the second time
  • the screening range is narrowed, which can greatly improve the efficiency of resource scheduling.
  • the scheduling device 13 needs to screen all GPUs in the GPU cluster 14 one by one for the GPU resources that can meet the computing parameters and quantity requirements according to the resource scheduling request. If the scale of the GPU cluster 14 is large, the screening range is It will be very large, and the screening time will be very long, making resource scheduling inefficient.
  • the above embodiment introduces the implementation of jointly determining the vGPU according to the computing parameters and the number N. If the computing parameters include computing power and video memory, when the vGPU is jointly determined according to the computing power and video memory, the following two implementations may be included:
  • the first screening is performed in the matching results according to the computing power requested by the resource scheduling request, and then the second screening is performed in the first screening results according to the display content required by the resource scheduling request. filter.
  • the calculation parameters include at least one of the following: computing power and video memory.
  • Step a1 Obtain the computing power of each vGPU in each target GPU and the priority corresponding to the video memory.
  • Step a2 If the priority of computing power is greater than the priority of video memory, screen each of the target GPUs for vGPUs that meet the computing power requirements of the vGPUs requested by resource scheduling to obtain a second screening result.
  • Step a3 Screen the vGPUs that meet the video memory requirements of the vGPUs requested by the resource scheduling in the second screening result to obtain the first screening result.
  • the first screening is performed in the matching result according to the display content requested by the resource scheduling request, and then the second screening is performed in the first screening result according to the computing power required by the resource scheduling request.
  • the computing parameters include at least one of the following: computing power and video memory; and determining the vGPU that satisfies the computing power and video memory in the matching result introduced in step S501a, including:
  • Step b1 Acquire the priority corresponding to the computing power and the video memory of each of the vGPUs in each of the target GPUs.
  • Step b2 If the priority of computing power is lower than the priority of video memory, select vGPUs that meet the video memory requirements of the vGPUs requested by resource scheduling in each of the target GPUs to obtain a third screening result.
  • Step b3 Screen the vGPUs that meet the computing power requirements of the vGPUs requested by the resource scheduling in the third screening result to obtain the first screening result.
  • the number of vGPUs in the first screening result is greater than the number of vGPUs requested by the resource scheduling request. In this case, it is also necessary to filter out the number of vGPUs requested by the resource scheduling request in the first screening result.
  • the number of vGPUs corresponds to the number of vGPUs (N vGPUs are filtered out in the first screening result).
  • the first screening result includes 5 vGPUs. If the number of vGPUs requested by the resource scheduling request is 4, then 4 vGPUs need to be further screened out of the 5 vGPUs, and the scheduling device 13 then selects these 4 vGPUs.
  • the vGPUs are returned to the AI algorithm device 12;
  • the vGPU in the first screening result is directly returned to the scheduling apparatus 13 as the target vGPU.
  • the first screening result includes 5 vGPUs. If the number of vGPUs requested by the resource scheduling request is 5, the 5 vGPUs are directly returned to the AI algorithm device 12 .
  • the scheduling apparatus 13 if the number of vGPUs in the first screening result is less than the number of vGPUs requested by the resource scheduling request, a message that the result is empty is returned to the scheduling apparatus 13 .
  • the first screening result includes 5 vGPUs. If the number of vGPUs requested by the resource scheduling request is 7, then the first screening result cannot meet the requirement of the number of vGPUs requested by the resource scheduling request, representing the GPU cluster. 14 If the resource scheduling request cannot be satisfied, the scheduling device 13 will return a message that the result is empty to the AI algorithm device 12 to notify the AI algorithm device 12 that the GPU cluster 14 cannot meet the resource scheduling request.
  • the first screening results may be sorted according to the order of computing parameters from small to large, and then selected according to the order of computing parameters from small to large
  • the number of vGPU resources required by the resource scheduling request corresponds to the number of vGPU resources, that is, the first N vGPUs are selected in the sorting result.
  • the first screening results may be sorted in ascending order of computing power, and then the top N vGPUs are selected therefrom.
  • the first screening result is as shown in Table 5 below:
  • the number of vGPU resources corresponding to the number of vGPU resources required by the resource scheduling request is selected in descending order of video memory.
  • the computing parameters include video memory
  • the computing parameters include computing power
  • the first screening result to select the number of vGPU resources required by the resource scheduling request in the order of computing power from small to large. The implementation manner of the vGPU resource will not be repeated here.
  • the computing parameters include computing power and video memory
  • the available vGPUs obtained from the first screening are sorted according to the calculation parameters from low to high, and the GPU card (smallest) that can meet the resource requirements is preferentially selected during screening.
  • the GPU card (smallest) that can meet the resource requirements is preferentially selected during screening.
  • the resource scheduling request further includes the task type of the task processing request corresponding to the resource scheduling request, and vGPUs in different GPUs correspond to tags, and the tags corresponding to the vGPUs are tasks corresponding to the resource scheduling request
  • the task type of the processing request is determined; the method of the embodiment of the present application further includes the following method steps:
  • At least one tag corresponding to the task type of the task processing request corresponding to the resource scheduling request is matched; and the vGPU corresponding to the at least one tag is used as the matching result.
  • the labels corresponding to vGPUs in different GPUs are the task types of the task processing requests corresponding to the resource scheduling requests.
  • some of the labels corresponding to 13 vGPUs are model training tasks, and these 13 vGPUs can be distributed on 0 cards.
  • the tags corresponding to the remaining 14 vGPUs are online prediction tasks, then if the task type of the task processing request corresponding to the resource scheduling request is a model training task, the matching result is distributed in 0 Some or all of the 13 vGPUs on any at least two of the 8 cards.
  • FIG. 7 is a schematic structural diagram of a resource scheduling apparatus provided by an embodiment of the present application.
  • the resource scheduling apparatus provided by the embodiment of the present application may execute the processing flow provided by the resource scheduling method embodiment.
  • the resource scheduling apparatus 70 includes: a receiving module 71, a first matching module 72, and a first returning module 73; wherein , the receiving module 71 is configured to receive a resource scheduling request for the GPUs in the graphics processor GPU cluster 14, the resource scheduling request includes the grouping information of the GPU to be requested, and the grouping information of the GPU to be requested is based on the resource scheduling request.
  • the task type of the corresponding task processing request is determined; the first matching module 72 is configured to, according to the grouping information of the GPU to be requested, match all GPUs in the GPU cluster 14 with the grouping information of the GPU to be requested.
  • the GPU obtains a matching result, where the matching result includes at least one target GPU corresponding to the grouping information of the GPU to be requested; the first returning module 73 is configured to return the matching result.
  • each GPU includes at least one vGPU
  • the resource scheduling request further includes the computing parameters and quantity of the vGPU
  • the apparatus further includes: a screening module 74, configured to calculate according to the computing parameters of the vGPU and the number of vGPUs that satisfy the computing parameters and number of the vGPUs in the matching result; the second returning module 75 is configured to return the vGPUs that satisfy the computing parameters and number of the vGPUs.
  • the screening module 74 includes: a first screening unit 741, configured to screen the vGPUs that satisfy the calculation parameters in the matching results to obtain a first screening result; a second screening unit 742. Configure, in the first screening result, to screen for vGPU resources that meet the requirement on the number of vGPUs.
  • the computing parameters include at least one of the following: computing power and video memory; the first screening unit 741 is configured to obtain the computing power of each vGPU in each of the target GPUs The priority corresponding to the power and the video memory; if the priority of the computing power is greater than the priority of the video memory, the vGPU that satisfies the computing power requirement of the vGPU of the resource scheduling request is screened in each of the target GPUs , and obtain a second screening result; in the second screening result, screen the vGPU that meets the video memory requirement of the vGPU requested by the resource scheduling, and obtain the first screening result.
  • the computing parameters include at least one of the following: computing power and video memory; the first screening unit 741 is configured to obtain the computing power of each vGPU in each of the target GPUs The priority corresponding to the power and the video memory; if the priority of the computing power is less than the priority of the video memory, then filter the vGPU that meets the video memory requirements of the vGPU requested by the resource scheduling request in each of the target GPUs, Obtaining a third screening result; screening vGPUs that meet the computing power requirement of the vGPU requested by the resource scheduling in the third screening results, to obtain the first screening result.
  • the second screening unit 742 is configured to, if the number of vGPUs in the first screening result is greater than the number of vGPU resources required by the resource scheduling request, then In the first screening result, select a number of vGPU resources corresponding to the number of vGPU resources required by the resource scheduling request according to the order of computing parameters from small to large; if the number of vGPUs in the first screening result is equal to the The number of the vGPU resources required by the resource scheduling request, the first screening result is returned; if the number of vGPUs in the first screening result is less than the number of vGPU resources required by the resource scheduling request number, the prompt message that the filter result is empty is returned.
  • the resource scheduling request includes the task type of the task processing request corresponding to the resource scheduling request; vGPUs in different GPUs have tags corresponding to the tags, and the tags corresponding to the vGPUs are The task type of the task processing request corresponding to the resource scheduling request is determined; the second matching module 76 is configured to match the task processing request corresponding to the resource scheduling request according to the task type of the task processing request corresponding to the resource scheduling request at least one tag corresponding to the requested task type; and using the vGPU corresponding to the at least one tag as the matching result.
  • the resource scheduling apparatus in the embodiment shown in FIG. 7 can be used to execute the technical solutions of the foregoing method embodiments, and the implementation principles and technical effects thereof are similar, and details are not described herein again.
  • FIG. 8 is a schematic structural diagram of an electronic device provided by an embodiment of the present application.
  • the electronic device provided by the embodiment of the present application can execute the processing flow provided by the resource scheduling method embodiment.
  • the electronic device 80 includes: a memory 81, a processor 82, a computer program, and a communication interface 83; wherein the computer program stores In the memory 81, and configured to be performed by the processor 82, the method steps of the above method embodiments.
  • the electronic device in the embodiment shown in FIG. 8 can be used to implement the technical solutions of the foregoing method embodiments, and the implementation principles and technical effects thereof are similar, and are not repeated here.
  • an embodiment of the present application further provides a computer-readable storage medium on which a computer program is stored, and the computer program is executed by a processor to implement the resource scheduling method described in the foregoing embodiment.
  • the disclosed apparatus and method may be implemented in other manners.
  • the apparatus embodiments described above are only illustrative.
  • the division of the units is only a logical function division. In actual implementation, there may be other division methods.
  • multiple units or components may be combined or Can be integrated into another system, or some features can be ignored, or not implemented.
  • the shown or discussed mutual coupling or direct coupling or communication connection may be through some interfaces, indirect coupling or communication connection of devices or units, and may be in electrical, mechanical or other forms.
  • the units described as separate components may or may not be physically separated, and components displayed as units may or may not be physical units, that is, may be located in one place, or may be distributed to multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution in this embodiment.
  • each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically alone, or two or more units may be integrated into one unit.
  • the above-mentioned integrated unit may be implemented in the form of hardware, or may be implemented in the form of hardware plus software functional units.
  • the above-mentioned integrated units implemented in the form of software functional units can be stored in a computer-readable storage medium.
  • the above-mentioned software function unit is stored in a storage medium, and includes several instructions to cause a computer device (which may be a personal computer, a server, or a network device, etc.) or a processor (processor) to execute the methods described in the various embodiments of the present application. some steps.
  • the aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (Read-Only Memory, ROM), random access memory (Random Access Memory, RAM), magnetic disk or optical disk and other media that can store program codes .
  • the resource scheduling request includes grouping information of the GPU to be requested, and the grouping information of the GPU to be requested is a task processing request corresponding to the resource scheduling request
  • the task type is determined, then according to the grouping information of the GPU to be requested, the GPUs with the grouping information of the GPU to be requested are matched among all the GPUs of the GPU cluster; finally, the grouping information that includes at least one target GPU corresponding to the grouping information of the GPU to be requested is returned. match results.
  • the resource scheduling request includes the grouping information of the GPU to be requested, and the grouping information of the GPU to be requested is determined according to the task type of the task processing request corresponding to the resource scheduling request, when performing GPU resource scheduling, it can be matched according to the grouping information. to the corresponding GPU, so as to achieve more fine-grained resource scheduling and precisely control the use of GPU.

Landscapes

  • Engineering & Computer Science (AREA)
  • Software Systems (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

L'invention concerne un procédé et un appareil de planification de ressources, un dispositif électronique, un support de stockage et un produit-programme. Le procédé consiste : à recevoir une demande de planification de ressources pour une unité de traitement graphique (GPU) dans une grappe d'unités GPU (S201), la demande de planification de ressources comportant des informations de regroupement d'une unité GPU à demander, et les informations de regroupement de l'unité GPU à demander étant déterminées selon un type de tâche d'une demande de traitement de tâche correspondant à la demande de planification de ressources ; en fonction des informations de regroupement de l'unité GPU à demander, à mettre en correspondance, dans toutes les unités GPU de la grappe d'unités GPU, une unité GPU ayant les informations de regroupement de l'unité GPU à demander de sorte à obtenir un résultat de mise en correspondance (S202), le résultat de mise en correspondance comprenant au moins une unité GPU cible correspondant aux informations de regroupement de l'unité GPU à demander ; et à renvoyer le résultat de mise en correspondance (S203).
PCT/CN2021/095292 2020-10-26 2021-05-21 Procédé et appareil de planification de ressources, dispositif électronique, support de stockage et produit-programme WO2022088659A1 (fr)

Priority Applications (1)

Application Number Priority Date Filing Date Title
KR1020217037982A KR20220058844A (ko) 2020-10-26 2021-05-21 리소스 스케줄링 방법 및 장치, 전자 기기, 저장 매체 및 프로그램 제품

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202011158231.7A CN112346859B (zh) 2020-10-26 2020-10-26 资源调度方法及装置、电子设备和存储介质
CN202011158231.7 2020-10-26

Publications (1)

Publication Number Publication Date
WO2022088659A1 true WO2022088659A1 (fr) 2022-05-05

Family

ID=74358745

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2021/095292 WO2022088659A1 (fr) 2020-10-26 2021-05-21 Procédé et appareil de planification de ressources, dispositif électronique, support de stockage et produit-programme

Country Status (3)

Country Link
KR (1) KR20220058844A (fr)
CN (1) CN112346859B (fr)
WO (1) WO2022088659A1 (fr)

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114820279A (zh) * 2022-05-18 2022-07-29 北京百度网讯科技有限公司 基于多gpu的分布式深度学习方法、装置及电子设备
CN115965517A (zh) * 2023-01-09 2023-04-14 摩尔线程智能科技(北京)有限责任公司 图形处理器资源管理方法及装置、电子设备和存储介质
CN115981871A (zh) * 2023-03-17 2023-04-18 苏州万店掌网络科技有限公司 一种gpu资源调度方法、装置、设备及存储介质
CN117687802A (zh) * 2024-02-02 2024-03-12 湖南马栏山视频先进技术研究院有限公司 一种基于云平台的深度学习并行调度方法、装置和云平台
CN117891618A (zh) * 2024-03-15 2024-04-16 杭州宇泛智能科技有限公司 人工智能模型训练平台的资源任务处理方法及装置
CN117891618B (zh) * 2024-03-15 2024-05-24 杭州宇泛智能科技有限公司 人工智能模型训练平台的资源任务处理方法及装置

Families Citing this family (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112346859B (zh) * 2020-10-26 2023-06-16 北京市商汤科技开发有限公司 资源调度方法及装置、电子设备和存储介质
CN113204428B (zh) * 2021-05-28 2023-01-20 北京市商汤科技开发有限公司 资源调度方法、装置、电子设备以及计算机可读存储介质
CN114968272A (zh) * 2022-05-31 2022-08-30 京东方科技集团股份有限公司 一种算法运行方法、装置、设备、存储介质
CN115048216A (zh) * 2022-05-31 2022-09-13 苏州浪潮智能科技有限公司 一种人工智能集群的资源管理调度方法、装置和设备
CN116302568A (zh) * 2023-05-17 2023-06-23 算力互联(北京)科技有限公司 一种算力资源调度方法及系统、调度中心、数据中心
CN116643893B (zh) * 2023-07-27 2023-10-20 合肥中科类脑智能技术有限公司 算力任务的调度方法、调度装置、存储介质及服务器
CN116757915B (zh) * 2023-08-16 2023-11-28 北京蓝耘科技股份有限公司 一种集群gpu资源调度方法
CN117539639A (zh) * 2024-01-05 2024-02-09 北京趋动智能科技有限公司 显存资源调度方法、装置、系统、存储介质及电子设备

Citations (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
EP2620873A1 (fr) * 2012-01-27 2013-07-31 Samsung Electronics Co., Ltd Procédé et appareil d'attribution de ressources de GPU
CN109144710A (zh) * 2017-06-16 2019-01-04 中国移动通信有限公司研究院 资源调度方法、装置及计算机可读存储介质
CN109376011A (zh) * 2018-09-26 2019-02-22 郑州云海信息技术有限公司 虚拟化系统中管理资源的方法和装置
CN109375992A (zh) * 2018-08-17 2019-02-22 华为技术有限公司 一种资源调度方法及装置
CN109634748A (zh) * 2018-12-12 2019-04-16 深圳前海微众银行股份有限公司 集群资源调度方法、装置、设备及计算机可读存储介质
CN110503593A (zh) * 2018-05-18 2019-11-26 微软技术许可有限责任公司 多个图形处理单元的调度
CN110941481A (zh) * 2019-10-22 2020-03-31 华为技术有限公司 资源调度方法、装置及系统
CN112346859A (zh) * 2020-10-26 2021-02-09 北京市商汤科技开发有限公司 资源调度方法及装置、电子设备和存储介质

Family Cites Families (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2017185285A1 (fr) * 2016-04-28 2017-11-02 华为技术有限公司 Procédé et dispositif d'attribution de tâche d'unité de traitement graphique
US10262390B1 (en) * 2017-04-14 2019-04-16 EMC IP Holding Company LLC Managing access to a resource pool of graphics processing units under fine grain control
CN110688218B (zh) * 2019-09-05 2022-11-04 广东浪潮大数据研究有限公司 资源调度方法及装置
CN111158879B (zh) * 2019-12-31 2024-03-22 上海依图网络科技有限公司 一种系统资源的调度方法，装置、机器可读介质和系统

Patent Citations (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
EP2620873A1 (fr) * 2012-01-27 2013-07-31 Samsung Electronics Co., Ltd Procédé et appareil d'attribution de ressources de GPU
CN109144710A (zh) * 2017-06-16 2019-01-04 中国移动通信有限公司研究院 资源调度方法、装置及计算机可读存储介质
CN110503593A (zh) * 2018-05-18 2019-11-26 微软技术许可有限责任公司 多个图形处理单元的调度
CN109375992A (zh) * 2018-08-17 2019-02-22 华为技术有限公司 一种资源调度方法及装置
CN109376011A (zh) * 2018-09-26 2019-02-22 郑州云海信息技术有限公司 虚拟化系统中管理资源的方法和装置
CN109634748A (zh) * 2018-12-12 2019-04-16 深圳前海微众银行股份有限公司 集群资源调度方法、装置、设备及计算机可读存储介质
CN110941481A (zh) * 2019-10-22 2020-03-31 华为技术有限公司 资源调度方法、装置及系统
CN112346859A (zh) * 2020-10-26 2021-02-09 北京市商汤科技开发有限公司 资源调度方法及装置、电子设备和存储介质

Cited By (10)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114820279A (zh) * 2022-05-18 2022-07-29 北京百度网讯科技有限公司 基于多gpu的分布式深度学习方法、装置及电子设备
CN114820279B (zh) * 2022-05-18 2023-03-24 北京百度网讯科技有限公司 基于多gpu的分布式深度学习方法、装置及电子设备
CN115965517A (zh) * 2023-01-09 2023-04-14 摩尔线程智能科技(北京)有限责任公司 图形处理器资源管理方法及装置、电子设备和存储介质
CN115965517B (zh) * 2023-01-09 2023-10-20 摩尔线程智能科技(北京)有限责任公司 图形处理器资源管理方法及装置、电子设备和存储介质
CN115981871A (zh) * 2023-03-17 2023-04-18 苏州万店掌网络科技有限公司 一种gpu资源调度方法、装置、设备及存储介质
CN115981871B (zh) * 2023-03-17 2024-01-26 苏州万店掌网络科技有限公司 一种gpu资源调度方法、装置、设备及存储介质
CN117687802A (zh) * 2024-02-02 2024-03-12 湖南马栏山视频先进技术研究院有限公司 一种基于云平台的深度学习并行调度方法、装置和云平台
CN117687802B (zh) * 2024-02-02 2024-04-30 湖南马栏山视频先进技术研究院有限公司 一种基于云平台的深度学习并行调度方法、装置和云平台
CN117891618A (zh) * 2024-03-15 2024-04-16 杭州宇泛智能科技有限公司 人工智能模型训练平台的资源任务处理方法及装置
CN117891618B (zh) * 2024-03-15 2024-05-24 杭州宇泛智能科技有限公司 人工智能模型训练平台的资源任务处理方法及装置

Also Published As

Publication number Publication date
KR20220058844A (ko) 2022-05-10
CN112346859A (zh) 2021-02-09
CN112346859B (zh) 2023-06-16

Similar Documents

Publication Publication Date Title
WO2022088659A1 (fr) Procédé et appareil de planification de ressources, dispositif électronique, support de stockage et produit-programme
CN109983480B (zh) 使用聚类损失训练神经网络
CN110837410B (zh) 任务调度方法、装置、电子设备及计算机可读存储介质
US10841236B1 (en) Distributed computer task management of interrelated network computing tasks
WO2021083276A1 (fr) Procédé, dispositif et appareil de combinaison d'une fédération horizontale et d'une fédération verticale et support
CN114741207B (zh) 一种基于多维度组合并行的gpu资源调度方法和系统
CN109657793B (zh) 模型训练方法及装置、存储介质及电子设备
CN112416585A (zh) 面向深度学习的gpu资源管理与智能化调度方法
CN113946431B (zh) 一种资源调度方法、系统、介质及计算设备
CN103493088A (zh) 标记权限
CN114416352A (zh) 算力资源分配方法、装置、电子设备及储存介质
CN109598250A (zh) 特征提取方法、装置、电子设备和计算机可读介质
CN115292014A (zh) 图像渲染方法、装置和服务器
CN111124644B (zh) 任务调度资源的确定方法、装置及系统
CN115586961A (zh) 一种ai平台计算资源任务调度方法、装置及介质
CN112148467A (zh) 计算资源的动态分配
CN111813541B (zh) 一种任务调度方法、装置、介质和设备
CN116820714A (zh) 一种算力设备的调度方法、装置、设备和存储介质
US11941519B2 (en) Machine learning training platform
CN109753353A (zh) 虚拟机资源分配方法、装置及电子设备
CN111796934B (zh) 任务下发方法、装置、存储介质和电子设备
CN114816719A (zh) 多任务模型的训练方法及装置
CN112988383A (zh) 一种资源分配方法、装置、设备以及存储介质
CN113657499A (zh) 基于特征选择的权益分配方法、装置、电子设备及介质
CN112148469B (zh) 管理资源的方法、装置及计算机存储介质

Legal Events

Date Code Title Description
ENP Entry into the national phase

Ref document number: 2021569056

Country of ref document: JP

Kind code of ref document: A

121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 21884402

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 21884402

Country of ref document: EP

Kind code of ref document: A1