JP2022516486A - Resource management methods and equipment, electronic devices, and recording media - Google Patents

Abstract

An embodiment of the present invention discloses a resource management allocation method and a device, an electronic device, and a recording medium, which method is based on reading a configuration file containing configuration parameters of a virtual GPU of a guest device and said configuration file. In response to generating at least one usable virtual GPU of the guest device and receiving the task to be processed by the target user, the physical GPU is added to the target virtual GPU corresponding to the task to be processed. The at least one available virtual GPU includes the allocation and the target virtual GPU. [Selection diagram] Fig. 2

Description

The present invention relates to the field of computer technology, specifically, resource management methods and devices, electronic devices, and recording media.

Image processing units (Graphics Processing Units, GPUs) are widely used in fields such as scientific computing, big data, and finance. Due to the large number of small and light computing tasks, GPU resources face the challenge of low computing load and high frequency of use. Thus, a single small application task is difficult to take full advantage of the GPU's computing performance, while a large number of small tasks increase the power consumption of the GPU cluster and are overall. Reduced resource utilization.

In order to solve the above problems, GPU virtualization technology has been proposed in the industry. That is, by subdividing each physical GPU resource into a plurality of virtual GPU computing resources, small tasks can be directly processed by the virtual GPU, and the degree of parallelism of the application and the resource utilization efficiency of the GPU cluster are improved.

The embodiments of the present invention provide resource management allocation methods and devices, electronic devices, and recording media.

A first aspect of an embodiment of the invention provides a resource management allocation method, which is based on reading a configuration file containing configuration parameters for a virtual image processing unit (GPU) of a guest device. In response to generating at least one usable virtual GPU of the guest device and receiving the task to be processed by the target user, the physical GPU is added to the target virtual GPU corresponding to the task to be processed. The at least one available virtual GPU includes the allocation and the target virtual GPU.

In one optional embodiment, allocating a physical GPU to the target virtual GPU corresponding to the task to be processed is performed on the target virtual GPU based on the video memory capacity required for the target virtual GPU and the resource status table. The resource state table includes allocating a physical GPU, and includes current state information of each physical GPU in a plurality of physical GPUs.

In one optional embodiment, the current state information of the physical GPU includes the video memory capacity of the physical GPU, the user information of each virtual GPU corresponding to the already occupied video memory of the physical GPU, and the physical GPU. Includes the video memory capacity of each virtual GPU corresponding to the already occupied video memory.

In one optional embodiment, allocating a physical GPU to the target virtual GPU based on the video memory capacity required for the target virtual GPU and the resource status table is the video memory capacity required for the target virtual GPU and the said. By allocating a physical GPU to the target virtual GPU based on the resource state table, the video memory of the plurality of physical GPUs is minimized.

In one optional embodiment, allocating a physical GPU to the target virtual GPU based on the video memory capacity and resource state table required for the target virtual GPU is a plurality of physical GPUs based on the resource state table. According to the arrangement order of, the remaining video memory capacity of each physical GPU is sequentially determined until the determined remaining video memory capacity meets the video memory capacity required for the target virtual GPU, and the remaining video memory capacity is determined. It includes determining a physical GPU that satisfies the video memory capacity required for the target virtual GPU as a physical GPU to be assigned to the target virtual GPU.

In one optional embodiment, the method determines that the remaining video memory capacity of the last arranged physical GPU among the plurality of physical GPUs does not meet the video memory capacity required for the target virtual GPU. Further includes determining that the virtual GPU has no available physical resources.

In one optional embodiment, allocating a physical GPU to the target virtual GPU based on the video memory capacity and resource status table required for the target virtual GPU is the video memory capacity and resources required for the target virtual GPU. By assigning a physical GPU to the target virtual GPU based on the state table, the at least one task of the target user may be allocated to the plurality of physical GPUs as evenly as possible. Includes the task to be processed.

In one optional embodiment, allocating a physical GPU to the target virtual GPU based on the video memory capacity and resource status table required for the target virtual GPU is based on the resource status table of the target user. To determine the number of tasks currently assigned to each physical GPU among the plurality of physical GPUs, and the remaining video memory capacity of each physical GPU, the number of tasks is the smallest, and the remaining video memory. It includes determining a physical GPU whose capacity satisfies the video memory capacity required for the target virtual GPU as a physical GPU to be assigned to the target virtual GPU.

In one optional embodiment, the method comprises the task when the number of tasks is the smallest and the remaining video memory capacity is a plurality of physical GPUs satisfying the video memory capacity required for the target virtual GPU. The front-arranged physical GPU among the plurality of physical GPUs having the smallest number and the remaining video memory capacity satisfying the video memory capacity required for the target virtual GPU is assigned to the target virtual GPU as the physical GPU. Further includes deciding.

In one optional embodiment, the method updates the resource state table based on the physical GPU assigned to the target virtual GPU, or if the task processing state changes, it is based on the change in the task processing state. Further includes updating the resource status table.

In a second aspect, embodiments of the invention provide a resource management allocation method that applies to guest devices, which method reads a host's configuration file, including configuration parameters for the virtual GPU of the guest device. And, when at least one usable virtual GPU of the guest device is generated based on the configuration file, and when a task to be processed by the target user is received, the target virtual GPU corresponding to the task to be processed is used. The resource allocation request is used to request the host to allocate a physical GPU to the target virtual GPU, including sending a resource allocation request to the host based on the at least one available. The virtual GPU includes the target virtual GPU.

In one optional embodiment, the method is based on at least a first resource state table of the guest device before sending a resource allocation request to the host based on the target virtual GPU corresponding to the task to be processed. The first resource further includes determining an idle virtual GPU in one available virtual GPU and determining the target virtual GPU from among the idle virtual GPUs. The state table contains the current state information of each physical GPU among a plurality of physical GPUs.

In one optional embodiment, the method receives a first update instruction for the first resource status table from the host and updates the first resource status table based on the first update instruction. The first update instruction is transmitted when the host detects that the task processing state of the physical GPU has changed.

In one optional embodiment, the at least one physical GPU comprises a first physical GPU with a newly added task, and the first update instruction corresponds to the newly added task of the first physical GPU. At least one of the user information, the information of the first physical GPU, and the video memory capacity of the virtual GPU corresponding to the newly added task is added and / or the at least one physical GPU. Includes a second physical GPU with a newly completed task, and the first update instruction is added with user information corresponding to the newly completed task and at least one of the information of the second physical GPU. Has been done.

In one optional embodiment, the method schedules the virtual GPU to execute the task to be processed, and receives the processing result of the task to be processed from the host. Including further.

In a third aspect, an embodiment of the invention provides a resource management allocation method applied to a host, the method from a guest device to which information about the task to be processed and information about the target virtual GPU are added. It includes receiving a resource allocation request and allocating a target physical GPU to the target virtual GPU based on the information of the target virtual GPU.

In one optional embodiment, assigning a target physical GPU to the target virtual GPU based on the information of the target virtual GPU causes the target virtual GPU to have a physical GPU based on the video memory capacity required for the target virtual GPU. Including assigning.

In one optional embodiment, allocating a physical GPU to the target virtual GPU based on the video memory capacity required for the target virtual GPU includes the video memory capacity required for the target virtual GPU, a resource allocation policy, and the first. The second resource state table includes allocating a physical GPU to the target virtual GPU based on the two resource state tables, and the second resource state table contains current state information of each physical GPU among the plurality of physical GPUs.

In one optional embodiment, allocating a physical GPU to the target virtual GPU based on the video memory capacity required for the target virtual GPU, the resource allocation policy, and the second resource status table is the second. Based on the resource state table, according to the arrangement order of the plurality of physical GPUs, the remaining video memory capacity of each physical GPU is set until the remaining video memory capacity determined meets the video memory capacity required for the target virtual GPU. It includes sequentially determining and determining a physical GPU whose remaining video memory capacity satisfies the video memory capacity required for the target virtual GPU as a physical GPU to be assigned to the target virtual GPU.

In one optional embodiment, if it is determined that the remaining video memory capacity of the last arranged physical GPU among the plurality of physical GPUs does not meet the video memory capacity required for the target virtual GPU, the virtual GPU Further includes determining that no physical resources are available.

In one optional embodiment, allocating a physical GPU to the target virtual GPU based on the video memory capacity required for the target virtual GPU, the resource allocation policy, and the second resource status table is the first. 2 Based on the resource status table, determine the number of tasks already allocated to each physical GPU in the target user's currently multiple physical GPUs, and the remaining video memory capacity of each physical GPU, and the number of tasks. The target user includes determining a physical GPU as the physical GPU to be assigned to the target virtual GPU, which is the least and the remaining video memory capacity satisfies the video memory capacity required for the target virtual GPU. It is a user corresponding to the target virtual GPU.

In one optional embodiment, the method comprises the task when the number of tasks is the smallest and the remaining video memory capacity is a plurality of physical GPUs satisfying the video memory capacity required for the target virtual GPU. The front-arranged physical GPU among the physical GPUs having the smallest number and the remaining video memory capacity satisfying the video memory capacity required for the target virtual GPU is determined as the physical GPU to be assigned to the target virtual GPU. Including that further.

In one optional embodiment, the method updates the second resource state table based on the physical GPU assigned to the target virtual GPU, or changes the task processing state if the task processing state changes. Further includes updating the second resource status table based on.

In one optional embodiment, the method further comprises performing task scheduling based on the task queue of each physical GPU among the plurality of physical GPUs.

In a fourth aspect, an embodiment of the invention provides a resource management system, including a host and a guest device, the host being provided with a plurality of physical GPUs, where the guest device is a virtual machine or a guest device. The guest device includes a container, reads a configuration file containing configuration parameters for the guest device's virtual image processing unit (GPU), and generates at least one available virtual GPU for the guest device based on the configuration file. When a task to be processed by the target user is received, a resource allocation request is sent to the host based on the target virtual GPU corresponding to the task to be processed, and the resource allocation request is physically sent to the target virtual GPU. Used to request the host to allocate a GPU, the at least one available virtual GPU includes the target virtual GPU, and the host receives a resource allocation request from its own guest device. The target physical GPU is assigned to the target virtual GPU based on the information of the target virtual GPU, and the information of the task to be processed and the information of the target virtual GPU are added to the resource allocation request.

In a fifth aspect, an embodiment of the present invention provides a resource management device, which comprises a communication unit and a processing unit, wherein the processing unit uses the communication unit to acquire a configuration file. A task to read the configuration file containing the configuration parameters of the virtual image processing unit (GPU) of the guest device, generate at least one usable virtual GPU of the guest device based on the configuration file, and process the target user. In response to receiving the, the physical GPU is assigned to the target virtual GPU corresponding to the task to be processed, and the at least one available virtual GPU includes the target virtual GPU.

In a sixth aspect, an embodiment of the present invention provides a resource management device, the device includes a communication unit and a processing unit, and the processing unit acquires a host configuration file by using the communication unit. Then, the host's configuration file containing the configuration parameters of the guest device's virtual GPU should be read, and based on the configuration file, at least one usable virtual GPU of the guest device should be generated and processed by the target user. When a task is received, the communication unit is used to send a resource allocation request to the host based on the target virtual GPU corresponding to the task to be processed, and the resource allocation request is sent to the target virtual GPU as a physical GPU. Used to request the host to allocate the at least one available virtual GPU includes the target virtual GPU.

In a seventh aspect, an embodiment of the present invention provides a resource management device, which comprises a communication unit and a processing unit, wherein the processing unit utilizes the communication unit to process a task. The resource allocation request from the guest device to which the information and the information of the target virtual GPU are added is received, and the target physical GPU is assigned to the target virtual GPU based on the information of the target virtual GPU.

In an eighth aspect, an embodiment of the present invention provides a resource management system, the system including a guest device and a host provided with a plurality of physical GPUs, wherein the guest device is an embodiment of the present invention. It is the device of any one of the sixth aspects of the example, and the host is the device of any one of the seventh aspects of the embodiments of the present invention.

In a ninth aspect, an embodiment of the invention provides an electronic device, wherein the electronic device comprises a processor and a memory for storing a computer program, wherein the computer program is executed by the processor described above. , The processor is made to carry out any one of the methods described in the first aspect, the second aspect, or the third aspect of the embodiment of the present invention.

In a tenth aspect, the embodiment of the present invention provides a computer-readable recording medium, the computer-readable recording medium described above stores a computer program, and the computer program described above is a computer of the embodiment of the present invention. Any one of the methods described in the first aspect, the second aspect, or the third aspect is to be performed.

In an eleventh aspect, an embodiment of the present invention provides a computer program product, wherein the computer program product described above includes a non-temporary computer-readable recording medium in which the computer program is stored, and the computer program described above operates. By doing so, the computer is made to carry out any one of the methods described in the first aspect, the second aspect, or the third aspect of the embodiment of the present invention. The computer program product can be a software installation package.

An embodiment of the present invention first reads a configuration file containing the configuration parameters of the virtual GPU of the guest device, and then generates at least one usable virtual GPU of the guest device based on the configuration file. Upon receiving the task to be processed from the customer, the physical GPU is assigned to the target virtual GPU corresponding to the task to be processed, and the at least one available virtual GPU includes the target virtual GPU. In the embodiment of the present invention, the fixed mapping relationship between the virtual GPU and the physical GPU is not built in advance, and after the virtual GPU receives the task to be processed, the corresponding physical GPU is dynamically transferred to the virtual GPU. Because of the allocation, it is possible to flexibly execute management allocation to the physical GPU resource based on the characteristics of the current task, improve the resource utilization rate of the physical GPU, and improve the overall performance of the system.

The drawings herein are incorporated into the specification and form part of the specification, the drawings illustrating embodiments in accordance with the invention, together with the specification the technical solutions of the invention. Used to explain.
It is a schematic diagram of the resource management system provided by the Example of this invention. It is a schematic diagram of the flow of the resource management allocation method disclosed by the Example of this invention. It is an architectural diagram of the resource minimum management allocation policy provided by the embodiment of this invention. FIG. 3 is an architectural diagram of a load balancing allocation policy provided by an embodiment of the present invention. It is a schematic diagram of the flow of the resource management allocation method disclosed by another Example of this invention. It is a schematic diagram of the flow of the resource management allocation method disclosed by still another embodiment of this invention. It is a schematic diagram of the structure of the electronic device provided by the Example of this invention. It is a schematic diagram of the configuration of the electronic device provided by another embodiment of the present invention. It is a schematic diagram of the structure of the electronic device provided by still another embodiment of the present invention. It is a block diagram of the structure of the functional unit of the resource management apparatus which concerns on embodiment of this invention. It is a block diagram of the structure of the functional unit of the resource management apparatus which concerns on another 1 Embodiment of this invention. It is a block diagram of the structure of the functional unit of the resource management apparatus which concerns on still another Embodiment of this invention.

Hereinafter, the technical solution according to the embodiment of the present invention will be clearly and completely described together with the drawings according to the embodiment of the present invention.

The term "and / or" in the present invention merely describes the relationships of related objects and indicates that three relationships may exist. For example, A and / or B includes three cases such that A exists alone, A and B exist at the same time, and B exists alone. Also, the term "at least one" as used herein refers to any one of a plurality of species or any combination of at least two of a plurality of species. For example, including at least one of A, B, C can indicate that it contains any one or more elements selected from a set composed of A, B, and C. The specification of the present invention, the scope of claims, and terms such as "first" and "second" in the above-mentioned drawings are used not to describe a specific order but to distinguish different objects. To. It should be noted that the terms "include" and "provide", and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not limited to the listed steps or units, but includes steps or units that are not optionally listed, or optionally these. Further includes other steps or units specific to the process, method, product, or device.

Reference to "Examples" herein means that the particular features, configurations, or properties described in conjunction with the Examples may be included in at least one Example of the invention. The appearance of the phrase at each position in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment that is mutually exclusive to the other embodiments. One of ordinary skill in the art clearly and implicitly understands that the embodiments described herein can be combined with other embodiments.

The resource management allocation device according to the embodiment of the present invention is a device capable of executing resource management allocation and may be an electronic device. The electronic devices described above include terminal devices, and in a specific implementation, the terminal device described above may be a mobile phone, laptop computer, or tablet computer having a touch-sensitive surface (eg, a touch screen display and / or a touch pad). Including but not limited to other portable devices such as. It should be further understood that in some embodiments, the device may be a desktop computer, server, etc. having a touch-sensitive surface (eg, a touch screen display and / or a touch pad) rather than a portable communication device. Is.

FIG. 1 is a schematic diagram of a resource management system provided by an embodiment of the present invention. The GPU virtualization solution involves a guest device (Guest) 101 and a host (Host) 102. Host 102 refers to the physical machine node of the device with the physical GPU (pGPU) 140, and the user can perform GPU tasks directly on the host. The host further includes a service daemon that is in charge of communication operations such as activation of the physical GPU and data analysis transfer with the guest device. Guest device 101 refers to a virtual machine or container that can use a virtualized GPU. Guest device 101 does not have any GPU device and cannot directly execute GPU tasks. Guest device 101 includes a process for GPU virtualization, which is responsible for the encapsulated GPU dynamic link library and data transfer, and tasks on guest device 101 transparently use the host's physical GPU. To execute the operation.

In the resource management system provided by the embodiments of the present invention, Guest's virtual GPU (vGPU) resource management policy can be selected as follows. That is, Guest's virtual GPU resource management policy is divided into two stages, such as virtual exhibition by configuration option and activation of actual allocation by task, where Guest can be a virtual machine. As shown in FIG. 1, the main allocation process is as follows. First, a virtual machine (Virtual Machine, VM) of a guest device is started, and a plurality of virtual machines such as 110-1 and 110-n may exist here. By reading the configuration file Config 121, the virtual GPU-related parameters currently on the VM, including the number of virtual GPUs, the video memory capacity of each virtual GPU, the duration of use of each virtual GPU, the size of each virtual GPU time slice, etc. To get. Next, the virtual GPU mirroring 122 of the VM is generated by the virtual GPU guest device, and the virtual GPU at this time is not mapped to the physical GPU and is virtually assigned to the user for observation at the software level. There is only virtual GPU information. When the user on the VM starts executing task 123 that uses the GPU resource, the virtual machine sends a request to initialize the GPU information, at which time the VM establishes a connection with the Host and allocates the virtual GPU-physical GPU resource. According to policy 130, the actual association between the virtual GPU 124 and the physical GPU 140 is established. Here, one virtual GPU 124 can be associated with one physical GPU, or can be associated with a plurality of physical GPUs. Host 102 returns the information of the physical GPU 140 to the VM, and returns parameters such as the address ID of the GPU, the device model of the physical GPU, the remaining video memory capacity of the physical GPU, and the start time for requesting the use of the physical GPU. The virtual GPU updates the above information, calls the GPU virtualization process according to the configured parameters, and transfers the data of the task 123 to the Host. The Host service daemon activates the associated physical GPU based on the data to be transferred, processes the task 123, and returns the task processing result to the VM after the task processing is completed. The user on the VM should be aware that the task 123 appears to have been completed on the virtual GPU 124 on the VM, but in reality the task 123 was completed on the physical GPU 140 of the host.

The Guest and Host described above may be provided on the same device, for example, a physical GPU on the host and a guest device installed on the host. It may be installed on different devices, for example, a guest device installed on the user device and each physical GPU node device installed on the network.

FIG. 2 is a schematic diagram of the flow of the resource management allocation method disclosed by the embodiment of the present invention, and as shown in FIG. 2, the resource management allocation method includes the following steps 201 to 203.

At 201, the configuration file containing the configuration parameters of the virtual GPU of the guest device is read.

The configuration file currently contains the related parameters of the virtual GPU of the guest device, and the virtual GPU related parameters are the number of virtual GPUs, the video memory capacity of each virtual GPU, the usage duration of each virtual GPU, and each virtual GPU time slice. It may include, but is not limited to, one or more of sizes and the like.

Optionally, the configuration file may be the system default or may be generated according to user requirements.

At 202, generate at least one available virtual GPU for the guest device based on the configuration file.

In an embodiment of the invention, after reading the configuration file described above, the guest device can generate at least one usable virtual GPU of the guest device based on the configuration file, the virtual GPU described above. , A virtual GPU used at the software level to present and observe to the user. For example, after generating at least one available virtual GPU as described above, the user can, through the user device, the number of virtual GPUs on the guest device, the video memory capacity of each virtual GPU, the duration of use of each virtual GPU. , Parameters such as the size of each virtual GPU time slice can be observed. The virtual GPU at this time is not mapped to the physical GPU.

In 203, in response to receiving a task to be processed by the target user, a physical GPU is assigned to the target virtual GPU corresponding to the task to be processed, and the at least one available virtual GPU is the target virtual GPU. Includes GPU.

In the embodiment of the present invention, after receiving the task to be processed, the guest device executes the process for the task to be processed by allocating the target virtual GPU to the task to be processed described above. Next, the physical GPU is assigned to the target virtual GPU described above based on the specific information of the target virtual GPU described above, that is, a mapping relationship between the virtual GPU and the physical GPU is constructed. After allocating the physical GPU to the target virtual GPU described above, the physical GPU assigned to the target virtual GPU described above processes the task to be processed described above, and feeds back the processing result to the guest device. That is, the task to be processed described above is processed by the physical GPU assigned to the host, and the processing result is fed back to the guest device.

In the embodiment of the present invention, when the configuration file including the parameters of the virtual GPU of the guest device is read and the task to be processed is received, the physical GPU is assigned to the target virtual GPU corresponding to the task to be processed, and the guest device is described. Includes at least one virtual GPU, and the at least one available virtual GPU comprises the target virtual GPU. In the embodiment of the present invention, the mapping relationship between the virtual GPU and the physical GPU is not established before the task is received, and the mapping relationship is established between the virtual GPU and the physical graphic virtual unit after the task is received. can do. Furthermore, management allocation can be flexibly executed for GPU resources according to different resource allocation policies, and the utilization efficiency of the physical GPU can be improved.

In one optional embodiment, allocating the physical GPU to the target virtual GPU corresponding to the task to be processed is the target virtual GPU based on the video memory capacity and the resource status table required for the target virtual GPU. The resource state table includes the current state information of each physical GPU among a plurality of physical GPUs in the host.

The current state information includes the video memory capacity of each physical GPU in the plurality of physical GPUs described above, and the user information of each virtual GPU corresponding to the already occupied video memory of the physical GPU in each physical GPU. And include the video memory capacity of each virtual GPU corresponding to the already occupied video memory of each physical GPU. Alternatively, based on the current state information, the video memory capacity of each physical GPU in the plurality of physical GPUs described above, and each virtual GPU corresponding to the already occupied video memory of the physical GPU in each physical GPU. Information such as the user ID of the user ID and the video memory capacity of each virtual GPU corresponding to the already occupied video memory of the physical GPU can be inferred. Optionally, the current state information described above may further include a task identifier for each virtual GPU in each physical GPU that corresponds to the already occupied video memory of that physical GPU.

Table 1 is a specific embodiment of the resource state table provided by the embodiments of the present invention. Here, pGPU indicates a physical GPU, vGPU indicates a virtual GPU, and ID indicates an address or an identifier.

In a specific implementation, after receiving the task to be processed described above and determining the target virtual GPU, the task to be processed is executed from the resource status table described above based on the video memory capacity of the target virtual GPU. A physical GPU that satisfies the conditions can be determined as the physical GPU to be assigned to the target virtual GPU, for example, as a physical GPU that allocates a physical GPU whose remaining video memory is larger than the video memory of the target virtual GPU described above to the target virtual GPU. Can be decided.

Further, allocating a physical GPU to the target virtual GPU based on the video memory capacity required for the target virtual GPU and the resource status table specifically includes the video memory capacity of the target virtual GPU and the resource status table. , It may include allocating a physical GPU to the target virtual GPU based on the resource allocation policy.

The resource allocation policy described above refers to a mapping rule between a virtual GPU and a physical GPU formulated to achieve a certain effect. For example, it could be a minimum resource allocation policy formulated to achieve a minimum of physical GPU usage, or, for example, to evenly allocate multiple tasks for the same user within each physical GPU. It can be a load balancing allocation policy. As an example, the resource allocation policy may be located on the host and the guest device may store some mapping relationships of the resource allocation policy.

In one optional embodiment, allocating a physical GPU to the target virtual GPU based on the video memory capacity of the target virtual GPU, the resource status table, and the resource allocation policy is based on the resource status table. , The remaining video memory capacity of each physical GPU is sequentially determined and the rest until the determined remaining video memory capacity meets the video memory capacity required for the target virtual GPU according to the arrangement order of the plurality of physical GPUs. A physical GPU whose video memory capacity satisfies the video memory capacity required for the target virtual GPU is determined as a physical GPU to be assigned to the target virtual GPU. When it is determined that the remaining video memory capacity of the last arranged physical GPU among the plurality of physical GPUs does not meet the video memory capacity required for the target virtual GPU, that is, the remaining video memory of each physical GPU. Is smaller than the video memory required for the target virtual GPU, it is determined that the virtual GPU has no available physical resources. In the embodiment of the present invention, the resource allocation policy in such a virtual GPU and physical GPU allocation method is referred to as a resource minimum allocation policy.

The above-mentioned arrangement order of the plurality of physical GPUs may be the order of the ID numbers of the physical GPUs, or may be the order of the arrangements of the plurality of physical GPUs in the above-mentioned resource state table, for example, pGPU1 to pGPUn in Table 1. obtain. The arrangement order of the physical GPUs in Table 1 may be arranged according to the order of the ID numbers of the physical GPUs, or may be arranged according to the position information in the host of the physical GPUs, or the physical GPUs have been added to the host. It may be arranged before and after the time, and the present invention is not limited to this.

In a concrete implementation, after determining the target virtual GPU video memory capacity for the task to be processed above, first determine the remaining video memory capacity of the first physical GPU according to the order of the physical GPUs in the resource state table. Then, determine whether the remaining video memory capacity of the first physical GPU is greater than or equal to the video memory capacity required for the target virtual GPU, and if it is greater than or equal to the video memory capacity required for the target virtual GPU. Determine the first physical GPU as the physical GPU to allocate to the target virtual GPU, otherwise get the remaining video memory capacity of the second physical GPU, then the remaining video memory capacity of the second physical GPU Is more than the video memory capacity required for the target virtual GPU, and if it is more than the video memory capacity required for the target virtual GPU, the second physical GPU is determined as the physical GPU to be assigned to the target virtual GPU. If not, continue to get the remaining video memory capacity of the next one physical GPU and perform a similar decision operation. By analogy, if the remaining video memory capacity of all physical GPUs is all less than the video memory capacity of the target virtual GPU described above, it is determined that there are no physical resources available for the virtual GPU.

FIG. 3 is an architectural diagram of the resource minimum management allocation policy provided by the embodiment of the present invention. Signed shaded rectangles indicate the tasks to be processed, labels within the rectangles indicate the order before and after the tasks are generated, VPTs indicate resource state tables, pGPUs indicate physical GPUs, and so on. vGPU represents a virtual GPU. As can be seen from the architecture diagram, tasks are assigned to physical GPUs arranged in front according to the generated front-back order.

In the embodiment of the present invention, when deciding the physical GPU to be assigned to the target virtual GPU, the utilization rate of each GPU is maximized by allocating the physical GPU arranged in front to the target virtual GPU as a priority, and the GPU Save resources.

In another optional embodiment, allocating a physical GPU to the target virtual GPU based on the video memory capacity of the target virtual GPU, the resource status table, and the resource allocation policy can result in the resource status table. Based on the determination of the number of tasks currently assigned to each physical GPU among the plurality of physical GPUs of the target user and the remaining video memory of each physical GPU, the number of tasks is the smallest and the number of tasks is the smallest. The user ID is a target of the target virtual GPU, including determining a physical GPU whose remaining video memory capacity satisfies the video memory capacity required for the target virtual GPU as a physical GPU to be assigned to the target virtual GPU. The user's identifier. The minimum number of tasks means that the target user has the smallest number of tasks running on the physical GPU, and does not mean that the number of all tasks running on the physical GPU is the smallest. In the embodiment of the present invention, the resource allocation policy in such a virtual GPU and physical GPU allocation method is referred to as a load balancing allocation policy.

In a specific implementation, based on the information in the resource state table, the number of tasks of the target user corresponding to the already occupied video memory of each physical GPU in the plurality of physical GPUs is determined and described above. The distribution of the target user's tasks in the above-mentioned multiple physical GPUs is statistic, and then the physical number of the tasks is the smallest and the remaining video memory capacity meets the video memory capacity required for the target virtual GPU. The GPU can be determined as the physical GPU to be assigned to the target virtual GPU. In this way, by allowing the tasks of the target user described above to be assigned to multiple physical GPUs as evenly as possible, multiple tasks of the same user can be stacked on the same physical GPU, so that multiple tasks of a single user can be performed. It reduces the queuing development of the user and improves the execution efficiency of the user's task.

Further, when the number of tasks is the smallest and the remaining video memory capacity is a plurality of physical GPUs that satisfy the video memory capacity required for the target virtual GPU, the number of tasks is the smallest and the remaining. The front-arranged physical GPU among the physical GPUs whose video memory capacity satisfies the video memory capacity required for the target virtual GPU is determined as the physical GPU to be assigned to the target virtual GPU. The arrangement order of the plurality of physical GPUs may be the order of the ID numbers of the physical GPUs, or may be the order of the arrangements of the plurality of physical GPUs in the above-mentioned resource state table.

FIG. 4 is an architectural diagram of the load balancing allocation policy provided by another embodiment of the present invention. Signed shaded rectangles indicate the tasks to be processed, labels within the rectangles indicate the order before and after the tasks are generated, VPTs indicate resource state tables, pGPUs indicate physical GPUs, and so on. vGPU represents a virtual GPU. As can be seen from the architecture diagram, tasks with the same user ID are evenly assigned to each physical GPU.

In one optional embodiment, after determining the physical GPU of the target virtual GPU described above, the method described further comprises updating the resource state table based on the physical GPU assigned to the target virtual GPU.

In a specific embodiment, after the physical GPU of the target virtual GPU described above is determined, the physical GPU-ID of the target virtual GPU described above, the ID of the target virtual GPU described above, the user ID of the target virtual GPU, and the target virtual GPU Update the above-mentioned resource status table using information such as video memory capacity.

In one optional embodiment, when the execution of the task existing in the plurality of physical GPUs described above is completed, the resource status table is updated based on the task whose execution is completed.

Specifically, the execution of the above-mentioned task in the above-mentioned resource status table is completed by using the physical GPU-ID that has completed the execution of the task and the ID of the virtual GPU corresponding to the task for which the execution has been completed. Information on the physical GPU can be updated.

FIG. 5 is a schematic diagram of the flow of the resource management allocation method disclosed by another embodiment of the present invention, and the method is applied to a guest device. As shown in FIG. 5, the resource management allocation method includes the following steps 501 to 503.

At 501, read the host's configuration file, which contains the configuration parameters of the guest device's virtual GPU.

At 502, create at least one usable virtual GPU for the guest device based on the configuration file.

In the examples of the present invention, the steps 501 and 502 can be referred to in steps 201 and 202 in FIG. 2, and are not described repeatedly here.

In 503, when the task to be processed by the target user is received, the resource allocation request is transmitted to the host based on the target virtual GPU corresponding to the task to be processed, and the resource allocation request is physically sent to the target virtual GPU. Used to request the host to allocate a GPU, the at least one available virtual GPU includes the target virtual GPU.

In an embodiment of the invention, the guest device described above may correspond to the Guest in FIG. 1, and the host described above may correspond to the Host in FIG.

In one example, the guest device can be a virtual machine, which comprises at least one virtual GPU. If the virtual GPU of the virtual machine described above does not have a task to be processed, the virtual GPU of the virtual machine described above does not have a corresponding physical GPU.

After receiving the task to be processed, the virtual machine described above allocates the target virtual GPU corresponding to the task to be processed, and then uses the information of the allocated target virtual GPU such as the video memory capacity of the target virtual GPU. Based on this, by sending a resource allocation request to the above-mentioned host, the host is requested to allocate a physical GPU to the target virtual GPU.

The above-mentioned resource allocation request may be added with the above-mentioned target virtual GPU ID and video memory capacity, and may be further added with the target virtual GPU user ID.

In another example, the guest device can be a container, which comprises at least one virtual GPU. If the virtual GPU of the container has no tasks to process, the virtual GPU of the container does not have a corresponding physical GPU. After receiving the task to be processed, the container allocates the target virtual GPU corresponding to the task to be processed described above, and then based on the information of the allocated target virtual GPU such as the video memory capacity of the target virtual GPU. By sending a resource allocation request to the above-mentioned host, the host is requested to allocate a physical GPU to the target virtual GPU. The above-mentioned resource allocation request may be added with the above-mentioned target virtual GPU ID and video memory capacity, and may be further added with the target virtual GPU user ID.

In an embodiment of the present invention, when a configuration file is read and a task to be processed is received, a resource allocation request is sent to the host based on the target virtual GPU corresponding to the task to be processed, and the configuration file is a guest. The guest device includes at least one virtual GPU, including parameters for the virtual GPU of the device. In the embodiment of the present invention, the virtual GPU and the physical GPU do not have a mapping relationship established before the task is received, and after the task is received, the virtual GPU and the physical GPU are dynamically relative to the virtual GPU and the physical GPU according to the resource allocation policy. Build a mapping relationship with. Therefore, the mapping relationship between the virtual GPU and the physical GPU can be flexibly constructed, and the usage efficiency of the physical GPU can be improved.

In one optional embodiment, the method comprises the number of virtual GPUs in the guest device and the first resource state before sending a resource allocation request to the host based on the target virtual GPU corresponding to the task to be processed. Determining the number of unused virtual GPUs in the guest device based on the table, and if the number of unused virtual GPUs is greater than zero, the target virtual GPU among the unused virtual GPUs. The first resource status table further includes the current usage status information of each physical GPU.

The information contained in the first resource status table described above matches the information contained in the resource status table in FIG. 2, and will not be described repeatedly here.

Optionally, the method further comprises receiving a first update instruction for the first resource state table and updating the first resource state table based on the first update instruction. 1 The update command is transmitted when the host detects that the task processing state of the physical GPU has changed. The change in the task processing state may be due to the addition of a new task to the physical GPU or the completion of the task.

Specifically, when it is detected by the host that a newly added task exists in the physical GPU, the first update instruction includes user information corresponding to the newly added task of the first physical GPU, the first. Information on one physical GPU, at least one of the video memory capacities of the virtual GPU of the newly added task is added. When the host detects that there is a newly completed task in the video memory of the physical GPU, the first update instruction includes the user information corresponding to the newly completed task and the newly completed task. At least one of the information of the corresponding second physical GPU is added. The first physical GPU is a GPU of a newly added task, and the second physical GPU is a GPU corresponding to a newly completed task. The user information may be the user's ID, and the physical GPU information may be the physical GPU's ID.

Optionally, the method further comprises scheduling the virtual GPU to perform the task to be processed and receiving the processing result of the task to be processed from the host.

FIG. 6 is a schematic diagram of the flow of the resource management allocation method disclosed by yet another embodiment of the present invention, and the method is applied to the host. As shown in FIG. 6, the resource management allocation method includes the following steps 601 to 602.

In 601 a resource allocation request from a guest device to which information of a task to be processed and information of a target virtual GPU is added is received, and the target virtual GPU is a virtual GPU included in the guest device.

In an embodiment of the invention, the guest device described above may correspond to the Guest in FIG. 1, and the host described above may correspond to the Host in FIG.

The guest device described above includes at least one virtual GPU. If the virtual GPU of the guest device described above has no task to be processed, the virtual GPU of the guest device described above does not have a corresponding physical GPU.

The information of the target virtual GPU described above may include, but is not limited to, the video memory capacity of the target virtual GPU and the user ID of the target virtual GPU.

In 602, the target physical GPU is assigned to the target virtual GPU based on the information of the target virtual GPU.

In the embodiment of the present invention, allocating the target physical GPU to the target virtual GPU based on the information of the target virtual GPU means allocating the physical GPU to the target virtual GPU based on the video memory capacity of the target virtual GPU. including. The target physical GPU is used to process a task to be processed based on the information of the task to be processed.

Further, allocating the physical GPU to the target virtual GPU based on the video memory capacity of the target virtual GPU is based on the video memory capacity required for the target virtual GPU, the resource allocation policy, and the second resource status table. The second resource state table includes allocating a physical GPU to the target virtual GPU, and includes current state information of each physical GPU among a plurality of physical GPUs.

In the embodiment of the present invention, the specific realization process of allocating the physical GPU to the target virtual GPU based on the video memory capacity of the target virtual GPU described above, the resource allocation policy, and the second resource status table is , The operation corresponding to "allocating the physical GPU to the target virtual GPU based on the video memory capacity required for the target virtual GPU and the resource status table" in FIG. 2 can be referred to. The operation of determining the number of tasks assigned to each physical GPU currently among the plurality of physical GPUs of the target user is the number of tasks assigned to each physical GPU currently assigned to each of the plurality of physical GPUs of the target user. It should be noted that it is a decision and the target user can be understood as a user corresponding to the target virtual GPU. The "second resource state table" in the embodiment of the present invention allocates a physical GPU to the target virtual GPU based on the "video memory capacity required for the target virtual GPU and the resource state table" in FIG. 2 described above. Corresponds to the "resource status table" in. In one example, the second resource state table is the same as the first resource state table, for example, the guest device copies the resource state table on the host locally and the host maintains the second resource state table. Then, when the second resource status table is updated, the guest device is notified to execute the synchronous update for the first resource status table. For example, the first update instruction can be used to notify the guest device to update the first resource status table.

In one optional embodiment, the method updates the second resource state table based on the physical GPU assigned to the target virtual GPU, or when the task processing state changes, the task processing state. It further includes updating the second resource status table based on the change. Updating the second resource status table based on the physical GPU assigned to the target virtual GPU can be the user information corresponding to the newly added task of the assigned physical GPU, the assigned physical GPU information, and the assigned physical GPU information. Includes updating at least one of the video memory capacities of the virtual GPU corresponding to the newly added task. When the task processing status changes, updating the second resource status table based on the change in the task processing status is the user information corresponding to the newly completed task and the information of the physical GPU that completed the task. Includes updating at least one of. The information of the physical GPU includes the ID of the physical GPU, and the user information includes the user ID.

In one optional embodiment, after allocating a physical GPU to the target virtual GPU corresponding to the task to be processed, the method performs task scheduling based on the task queue of each physical GPU in the plurality of physical GPUs. It may further include doing.

The task queue means a task whose physical GPU is already contained in the occupied video memory.

In the embodiment of the present invention, first, a resource allocation request from a guest device is received, information on a task to be processed and a target virtual GPU is added to the resource allocation request, and the target virtual GPU is in the guest device. It is a virtual GPU included in the guest device of the above, and then, based on the information of the target virtual GPU and the resource management policy, the target physical GPU is assigned to the target virtual GPU. In the embodiment of the present invention, a fixed mapping relationship between the virtual GPU and the physical GPU is not built in advance, and after the task is received by the virtual GPU, the corresponding physical GPU is dynamically assigned to the virtual GPU, so that different resources are used. It is possible to flexibly execute management allocation to GPU resources according to the allocation policy, improve the resource utilization rate of the physical GPU, and improve the overall performance of the system.

Similar to the embodiment shown in FIG. 2 above, FIG. 7 is a schematic diagram of the configuration of the electronic device 700 provided by the embodiment of the present invention, and as shown in FIG. 7, the electronic device 700 is , A processor 710, a memory 720, a communication interface 730, and a physical GPU 740. The one or more programs 721 are stored in the memory 720 described above and are executed by the processor 710 described above. The one or more programs 721 include instructions to perform the following steps.

Received a configuration file containing the GPU configuration parameters of the guest device, generated at least one usable virtual GPU for the guest device based on the configuration file, and a task to be processed by the target user. In response to, the target virtual GPU corresponding to the task to be processed is assigned a physical GPU, and the at least one available virtual GPU includes the target virtual GPU.

An embodiment of the present invention reads a configuration file containing the configuration parameters of the virtual GPU of the guest device, generates at least one usable virtual GPU of the guest device based on the configuration file, and receives a task to be processed. If so, a physical GPU is assigned to the target virtual GPU corresponding to the task to be processed, and the at least one available virtual GPU includes the target virtual GPU. In the embodiment of the present invention, the fixed mapping relationship between the virtual GPU and the physical GPU is not built in advance, and the corresponding physical GPU is dynamically assigned to the virtual GPU after the virtual GPU receives the task to be processed. The GPU resources can be flexibly allocated based on the characteristics of the current task, the resource utilization of the physical GPU has been improved, and the overall performance of the GPU system has been improved.

In one optional embodiment, when a physical GPU is assigned to the target virtual GPU corresponding to the task to be processed, the instructions in the program specifically include the video memory capacity and resource status table required for the target virtual GPU. Based on the above, the physical GPU is assigned to the target virtual GPU, and the resource state table includes the current state information of each physical GPU in the plurality of physical GPUs.

In one optional embodiment, the current state information of the physical GPU includes the video memory capacity of the physical GPU, the user information of each virtual GPU corresponding to the already occupied video memory of the physical GPU, and the physical GPU. Includes the video memory capacity of each virtual GPU corresponding to the already occupied video memory.

In one optional embodiment, when allocating a physical GPU to the target virtual GPU based on the video memory capacity and resource status table required for the target virtual GPU, the instructions in the program specifically say the target. By allocating a physical GPU to the target virtual GPU based on the video memory capacity required for the virtual GPU and the resource status table, the video memory of the plurality of physical GPUs is minimized. do.

In one optional embodiment, when allocating a physical GPU to the target virtual GPU based on the video memory capacity and resource status table required for the target virtual GPU, the instructions in the program specifically include the resource. Based on the state table, the remaining video memory capacity of each physical GPU is sequentially determined according to the arrangement order of the plurality of physical GPUs until the determined remaining video memory capacity meets the video memory capacity required for the target virtual GPU. And to determine a physical GPU whose remaining video memory capacity satisfies the video memory capacity required for the target virtual GPU as the physical GPU to be assigned to the target virtual GPU.

In one optional embodiment, the instructions in the program further satisfy the remaining video memory capacity of the last arranged physical GPU among the plurality of physical GPUs to meet the video memory capacity required for the target virtual GPU. If it is determined that there is no physical resource available in the virtual GPU, it is determined that there is no available physical resource.

In one optional embodiment, when allocating a physical GPU to the target virtual GPU based on the video memory capacity and resource status table required for the target virtual GPU, the instructions in the program specifically indicate the target. By allocating a physical GPU to the target virtual GPU based on the video memory capacity required for the virtual GPU and the resource status table, at least one task of the target user should be allocated to the plurality of physical GPUs as evenly as possible. The at least one task includes the task to be processed.

In one optional embodiment, when allocating a physical GPU to the target virtual GPU based on the video memory capacity and resource status table required for the target virtual GPU, the instructions in the program specifically include the resource. Based on the state table, the target user determines the number of tasks currently allocated to each physical GPU among the plurality of physical GPUs, and the remaining video memory capacity of each physical GPU, and the number of tasks. Is executed, and the physical GPU whose remaining video memory capacity satisfies the video memory capacity required for the target virtual GPU is determined as the physical GPU to be assigned to the target virtual GPU.

In one optional embodiment, the instructions in the program further have a plurality of physical GPUs with the least number of tasks and the remaining video memory capacity satisfying the video memory capacity required for the target virtual GPU. If there is, the target virtual GPU is the first-arranged physical GPU among a plurality of physical GPUs having the smallest number of tasks and the remaining video memory capacity satisfying the video memory capacity required for the target virtual GPU. Performs to determine as the physical GPU to assign to.

In one optional embodiment, the instructions in the program further update the resource state table based on the physical GPU assigned to the target virtual GPU, or if the task processing state changes, the task processing. Update the resource status table based on the change in status.

Similar to the embodiment shown in FIG. 5 above, FIG. 8 is a schematic diagram of the configuration of the electronic device 800 provided by another embodiment of the present invention, wherein the electronic device 800 is a guest device. possible. As shown in FIG. 8, the electronic device 800 includes a processor 810, a memory 820, and a communication interface 830. The one or more programs 821 are stored in the memory 820 described above and are executed by the processor 810 described above. The one or more programs 821 include instructions to perform the following steps.

Reading the host's configuration file, including the configuration parameters of the guest device's virtual GPU, generating at least one available virtual GPU for the guest device based on the configuration file, and processing the target user. When a task to be processed is received, the resource allocation request includes sending a resource allocation request to the host based on the target virtual GPU corresponding to the task to be processed, and the resource allocation request causes the physical GPU to the target virtual GPU. The at least one available virtual GPU used to request the host to allocate includes the target virtual GPU.

An embodiment of the present invention reads a configuration file containing the parameters of a virtual GPU of a guest device, generates at least one usable virtual GPU of the guest device based on the configuration file, and receives a task to be processed. If so, a resource allocation request is sent to the host based on the target virtual GPU corresponding to the task to be processed. In the embodiment of the present invention, the fixed mapping relationship between the virtual GPU and the physical GPU is not built in advance, and the physical GPU dynamically sends the resource allocation request to the host after the virtual GPU receives the task. Improves the resource utilization of the GPU and improves the overall performance of the GPU system.

In one optional embodiment, the instructions in the program are further based on the first resource state table before sending the resource allocation request to the host based on the target virtual GPU corresponding to the task to be processed. Determining the idle virtual GPU in at least one available virtual GPU of the guest device and determining the target virtual GPU from the idle virtual GPUs are executed. , The first resource state table contains the current state information of each physical GPU among a plurality of physical GPUs.

In one optional embodiment, the instruction in the program further receives a first update instruction for the first resource status table from the host and the first resource state based on the first update instruction. The table is updated and executed, and the first update instruction is transmitted when it is detected by the host that the task processing state of the physical GPU has changed.

In one optional embodiment, the at least one physical GPU includes a first physical GPU with a newly added task, and the first update instruction is a user corresponding to the newly added task of the first physical GPU. At least one of the information, the information of the first physical GPU, and the video memory capacity of the virtual GPU corresponding to the newly added task is added, and / or the at least one physical GPU is newly added. A second physical GPU having a completed task is included, and at least one of the user information corresponding to the newly completed task and the information of the second physical GPU is added to the first update instruction. ..

In one optional embodiment, the instructions in the program further schedule the virtual GPU to execute the task to be processed, and the processing result of the task to be processed from the host. To receive and to execute.

Similar to the embodiment shown in FIG. 6 described above, FIG. 9 is a schematic diagram of the configuration of the electronic device 900 provided by yet another embodiment of the present invention, wherein the electronic device 900 is a host. possible. As shown in FIG. 9, the electronic device 900 includes a processor 910, a memory 920, a communication interface 930, and a physical GPU 940. The one or more programs 921 are stored in the memory 920 described above and are executed by the processor 910 described above. The one or more programs 921 include instructions for performing the following steps.

Receiving a resource allocation request from a guest device to which information on the task to be processed and information on the target virtual GPU are added, and allocating the target physical GPU to the target virtual GPU based on the information on the target virtual GPU. Including that.

An embodiment of the present invention receives a resource allocation request from a guest device to which information on a task to be processed and information on a target virtual GPU are added, and targets the target virtual GPU based on the information on the target virtual GPU. Allocate a physical GPU. In the embodiment of the present invention, the fixed mapping relationship between the virtual GPU and the physical GPU is not built in advance, and the corresponding physical GPU is dynamically allocated to the virtual GPU after receiving the resource allocation request from the guest device. The GPU resources can be flexibly allocated based on the characteristics of the current task, the resource utilization of the GPU has been improved, and the overall performance of the GPU system has been improved.

In one optional embodiment, when the target physical GPU is assigned to the target virtual GPU based on the information of the target virtual GPU, the instruction in the program specifically determines the video memory capacity required for the target virtual GPU. Based on this, the physical GPU is assigned to the target virtual GPU.

In one optional embodiment, when allocating a physical GPU to the target virtual GPU based on the video memory capacity required for the target virtual GPU, the instructions in the program specifically indicate the video required for the target virtual GPU. Executes allocation of a physical GPU to the target virtual GPU based on the memory capacity, the resource allocation policy, and the second resource status table, and the second resource status table is each of the plurality of physical GPUs. Contains current state information for the physical GPU.

In one optional embodiment, when allocating a physical GPU to the target virtual GPU based on the video memory capacity required for the target virtual GPU, the resource allocation policy, and the second resource status table, in the program. The instruction specifically, based on the second resource state table, according to the arrangement order of the plurality of physical GPUs, until the remaining video memory capacity determined meets the video memory capacity required for the target virtual GPU. The remaining video memory capacity of each physical GPU is sequentially determined, and the physical GPU whose remaining video memory capacity satisfies the video memory capacity required for the target virtual GPU is determined as the physical GPU to be allocated to the target virtual GPU. And execute.

In one optional embodiment, the instructions in the program further satisfy the remaining video memory capacity of the last arranged physical GPU among the plurality of physical GPUs to meet the video memory capacity required for the target virtual GPU. If it is determined that there is no physical resource available in the virtual GPU, it is determined that there is no available physical resource.

In one optional embodiment, in the case of allocating a physical GPU to the target virtual GPU based on the video memory capacity required for the target virtual GPU, the resource allocation policy, and the second resource status table, in the program. Specifically, based on the second resource state table, the number of tasks already allocated to each physical GPU in the current plurality of physical GPUs of the target user, and the remaining video memory of each physical GPU. Determining the capacity and determining the physical GPU with the smallest number of tasks and the remaining video memory capacity satisfying the video memory capacity required for the target virtual GPU as the physical GPU to be allocated to the target virtual GPU. And, the target user is a user corresponding to the target virtual GPU.

In one optional embodiment, the instructions in the program further have a plurality of physical GPUs with the least number of tasks and the remaining video memory capacity satisfying the video memory capacity required for the target virtual GPU. If there is, the front-arranged physical GPU among the physical GPUs having the smallest number of tasks and the remaining video memory capacity satisfying the video memory capacity required for the target virtual GPU is assigned to the target virtual GPU. Performs the determination as a physical GPU.

In one optional embodiment, the instructions in the program further update the second resource state table based on the physical GPU assigned to the target virtual GPU, or if the task processing state changes. The second resource status table is updated based on the change in the task processing status.

In one optional embodiment, the instructions in the program further execute task scheduling based on the task queue of each physical GPU in the plurality of physical GPUs.

FIG. 10 is a block diagram of the configuration of the functional unit of the resource management device 1000 according to the embodiment of the present invention. The resource management device 1000 is applied to electronic devices. The electronic device includes a processing unit 1010 and a communication unit 1020.

The processing unit 1010 uses the communication unit 1020 to acquire a configuration file, reads a configuration file containing configuration parameters of the GPU of the guest device, and can use at least one of the guest devices based on the configuration file. In response to receiving a task to be processed by the target user by creating a virtual GPU, a physical GPU is assigned to the target virtual GPU corresponding to the task to be processed, and the at least one available virtual GPU is Includes the target virtual GPU.

An embodiment of the present invention reads a configuration file containing the parameters of a virtual GPU of a guest device, generates at least one usable virtual GPU of the guest device based on the configuration file, and receives a task to be processed. If so, a physical GPU is assigned to the target virtual GPU corresponding to the task to be processed, and the at least one usable virtual GPU includes the target virtual GPU. In the embodiment of the present invention, the fixed mapping relationship between the virtual GPU and the physical GPU is not built in advance, and the corresponding physical GPU is dynamically assigned to the virtual GPU after the virtual GPU receives the task to be processed. The GPU resources can be flexibly allocated based on the characteristics of the current task, the resource utilization of the GPU has been improved, and the overall performance of the GPU system has been improved.

In one optional embodiment, when a physical GPU is assigned to the target virtual GPU corresponding to the task to be processed, the processing unit 1010 specifically includes a video memory capacity and a resource status table required for the target virtual GPU. The physical GPU is assigned to the target virtual GPU based on the above, and the resource state table includes the current state information of each physical GPU in the plurality of physical GPUs.

In one optional embodiment, the current state information of each physical GPU includes the video memory capacity of the physical GPU, the user information of each virtual GPU corresponding to the already occupied video memory of the physical GPU, and the physical GPU. Includes the video memory capacity of each virtual GPU corresponding to the already occupied video memory of.

In one optional embodiment, when a physical GPU is assigned to the target virtual GPU based on the video memory capacity required for the target virtual GPU and the resource status table, the processing unit 1010 specifically indicates the target virtual GPU. By allocating a physical GPU to the target virtual GPU based on the video memory capacity required for the GPU and the resource status table, the video memory of the plurality of physical GPUs is minimized.

In one optional embodiment, when allocating a physical GPU to the target virtual GPU based on the video memory capacity required for the target virtual GPU and the resource state table, the processing unit 1010 specifically has the resource state. Based on the table, according to the arrangement order of the plurality of physical GPUs, the remaining video memory capacity of each physical GPU is sequentially determined until the determined remaining video memory capacity meets the video memory capacity required for the target virtual GPU. The physical GPU whose remaining video memory capacity satisfies the video memory capacity required for the target virtual GPU is determined as the physical GPU to be assigned to the target virtual GPU.

In one optional embodiment, the processing unit 1010 further does not meet the video memory capacity required for the target virtual GPU in the remaining video memory capacity of the last arranged physical GPU among the plurality of physical GPUs. If it is determined that the virtual GPU has no available physical resources.

In one optional embodiment, when a physical GPU is assigned to the target virtual GPU based on the video memory capacity required for the target virtual GPU and the resource status table, the processing unit 1010 specifically indicates the target virtual GPU. By allocating a physical GPU to the target virtual GPU based on the video memory capacity required for the GPU and the resource status table, at least one task of the target user can be allocated to the plurality of physical GPUs as evenly as possible. The at least one task includes the task to be processed.

In one optional embodiment, when allocating a physical GPU to the target virtual GPU based on the video memory capacity required for the target virtual GPU and the resource state table, the processing unit 1010 specifically has the resource state. Based on the table, the target user's number of tasks currently allocated to each physical GPU among the plurality of physical GPUs, and the remaining video memory capacity of each physical GPU are determined, and the number of tasks is the smallest. And, the physical GPU whose remaining video memory capacity satisfies the video memory capacity required for the target virtual GPU is determined as the physical GPU to be assigned to the target virtual GPU.

In one optional embodiment, the processing unit 1010 further has a plurality of physical GPUs with the least number of tasks and the remaining video memory capacity satisfying the video memory capacity required for the target virtual GPU. In this case, the front-arranged physical GPU among the plurality of physical GPUs having the smallest number of tasks and the remaining video memory capacity satisfying the video memory capacity required for the target virtual GPU is used as the target virtual GPU. Determined as the physical GPU to allocate.

In one optional embodiment, the processing unit 1010 further updates the resource status table based on the physical GPU assigned to the target virtual GPU, or when the task processing status changes, the task processing status is changed. The resource status table is updated based on the change.

FIG. 11 is a block diagram of the configuration of the functional unit of the resource management device 1100 according to another embodiment of the present invention. Applied to the resource management device 1100 electronic device, the electronic device includes a processing unit 1110 and a communication unit 1120.

The processing unit 1110 uses the communication unit 1120 to acquire a host configuration file, reads the host configuration file including the configuration parameters of the virtual GPU of the guest device, and the guest is based on the configuration file. When at least one usable virtual GPU of the device is generated and a task to be processed by the target user is received, the communication unit 1120 is used to allocate resources based on the target virtual GPU corresponding to the task to be processed. The request is sent to the host, the resource allocation request is used to request the host to allocate a physical GPU to the target virtual GPU, and the at least one available virtual GPU is the target virtual GPU. including.

An embodiment of the present invention reads a configuration file containing the parameters of a virtual GPU of a guest device, generates at least one usable virtual GPU of the guest device based on the configuration file, and receives a task to be processed. In this case, the communication unit is used to send a resource allocation request to the host based on the target virtual GPU corresponding to the task to be processed. In the embodiment of the present invention, the fixed mapping relationship between the virtual GPU and the physical GPU is not built in advance, and the resource allocation request is sent to the host after the virtual GPU receives the task. Therefore, the resource utilization rate of the physical GPU. And improved the overall performance of the GPU system.

In one optional embodiment, the processing unit 1110 further bases on the first resource state table before sending a resource allocation request to the host based on the target virtual GPU corresponding to the task to be processed. The virtual GPU that is idle in at least one available virtual GPU of the guest device is determined, the target virtual GPU is determined from the virtual GPU that is idle, and the first resource state table is Contains the current state information of each physical GPU in a plurality of physical GPUs.

In one optional embodiment, the processing unit 1110 further receives a first update instruction for the first resource status table from the host and updates the first resource status table based on the first update instruction. However, the first update command is transmitted when the host detects that the task processing state of at least one physical GPU has changed.

In one optional embodiment, the first update instruction includes user information corresponding to the newly added task of the first physical GPU, information of the first physical GPU, and virtual GPU corresponding to the newly added task. At least one of the video memory capacity, the user information corresponding to the completed task, and the information of the second physical GPU corresponding to the completed task is added, and the first physical GPU is newly added. It is the GPU of the added task, and the second physical GPU is the GPU that has completed the task.

In one optional embodiment, the processing unit 1110 further schedules the virtual GPU to execute the task to be processed, and receives the processing result of the task to be processed from the host.

FIG. 12 is a block diagram of the configuration of the functional unit of the resource management device 1200 according to still another embodiment of the present invention. The resource management device 1200 is applied to electronic devices. The electronic device includes a processing unit 1210 and a communication unit 1220, and the processing unit 1210 is a guest device to which information on a task to be processed and information on a target virtual GPU are added by using the communication unit 1220. Receives the resource allocation request from the target virtual GPU and allocates the target physical GPU to the target virtual GPU based on the information of the target virtual GPU.

An embodiment of the present invention receives a resource allocation request from a guest device to which information on a task to be processed and information on a target virtual GPU are added, and targets the target virtual GPU based on the information on the target virtual GPU. Allocate a physical GPU. In the embodiment of the present invention, the fixed mapping relationship between the virtual GPU and the physical GPU is not built in advance, and the corresponding physical GPU is dynamically allocated to the virtual GPU after receiving the resource allocation request from the guest device. The GPU resources can be flexibly allocated based on the characteristics of the current task, the resource utilization of the physical GPU has been improved, and the overall performance of the GPU system has been improved.

In one optional embodiment, when the target physical GPU is assigned to the target virtual GPU based on the information of the target virtual GPU, the processing unit 1210 is specifically based on the video memory capacity required for the target virtual GPU. And assign a physical GPU to the target virtual GPU.

In one optional embodiment, when allocating a physical GPU to the target virtual GPU based on the video memory capacity required for the target virtual GPU, the processing unit 1210 specifically has the video memory required for the target virtual GPU. A physical GPU is allocated to the target virtual GPU based on the capacity, the resource allocation policy, and the second resource state table, and the second resource state table is the current state information of each physical GPU in the plurality of physical GPUs. including.

In one optional embodiment, when allocating a physical GPU to the target virtual GPU based on the video memory capacity required for the target virtual GPU, the resource allocation policy, and the second resource status table, the processing unit 1210 Specifically, based on the second resource state table, according to the arrangement order of the plurality of physical GPUs, until the remaining video memory capacity determined meets the video memory capacity required for the target virtual GPU. The remaining video memory capacity of the physical GPU is sequentially determined, and the physical GPU whose remaining video memory capacity satisfies the video memory capacity required for the target virtual GPU is determined as the physical GPU to be assigned to the target virtual GPU.

In one optional embodiment, the processing unit 1210 further does not meet the video memory capacity required for the target virtual GPU in the remaining video memory capacity of the last physical GPU arranged among the plurality of physical GPUs. If it is determined that the virtual GPU has no available physical resources.

In one optional embodiment, when allocating a physical GPU to the target virtual GPU based on the video memory capacity required for the target virtual GPU, the resource allocation policy, and the second resource status table, the processing unit 1210 Specifically, based on the second resource state table, the number of tasks already allocated to each physical GPU among the currently multiple physical GPUs of the target user, and the remaining video memory capacity of each physical GPU. The physical GPU having the smallest number of tasks and the remaining video memory capacity satisfying the video memory capacity required for the target virtual GPU is determined as the physical GPU to be assigned to the target virtual GPU, and the target user determines. , The user corresponding to the target virtual GPU.

In one optional embodiment, the processing unit 1210 further has a plurality of physical GPUs with the least number of tasks and the remaining video memory capacity satisfying the video memory capacity required for the target virtual GPU. When the number of tasks is the smallest, and the remaining video memory capacity satisfies the video memory capacity required for the target virtual GPU, the front-arranged physical GPU among the physical GPUs is allocated to the target virtual GPU. Determined as GPU.

In one optional embodiment, the processing unit 1210 further updates the second resource state table based on the physical GPU assigned to the target virtual GPU, or if the task processing state changes, the task processing. The second resource status table is updated based on the change in status.

In one optional embodiment, the processing unit 1210 further performs task scheduling based on the task queue of each physical GPU in the plurality of physical GPUs.

The embodiments of the present invention further provide a computer recording medium, the computer recording medium storing a computer program, wherein the computer is the computer of any one of the methods described in the embodiments of the above-mentioned methods. Try to perform some or all steps. The computers described above include electronic devices.

The embodiments of the present invention further provide a computer program product, the computer program product described above comprising a non-temporary computer-readable recording medium in which the computer program is stored. The computer program described above, by being manipulated, causes the computer to perform some or all steps of any one of the methods described in the embodiments of the method described above. The computer program product may be one software installation package, the computer described above comprising an electronic device.

Examples of each of the above-mentioned methods are represented as a combination of a series of actions for the sake of brevity, but those skilled in the art are not limited to the order of the described actions and are described in the present invention. It should be explained that some steps should be understood that they can be performed in other order or at the same time. Next, one of ordinary skill in the art should understand that all of the embodiments described herein belong to the preferred embodiments and that the associated operation and module are not necessarily required by the present invention.

In the embodiments described above, the description of each embodiment is focused on itself. For parts not described in the citadel in one embodiment, the relevant description in another embodiment can be referred to.

It should be understood that in some of the embodiments provided by the present invention, the disclosed devices may be implemented in other ways. For example, the embodiments of the device described above are only exemplary. For example, the unit division described above is only a logical function division, and in a real implementation there may be other divisions, for example, multiple units or components can be combined or integrated into another system. Alternatively, some features can be ignored or not implemented. Further, the interconnect or direct coupling or communication connection displayed or described can be an indirect coupling or communication connection of devices or units via some interface, and can be electrical or other form.

The unit described as the above-mentioned separated part may or may not be physically separated, and the part displayed as a unit may or may not be a physical unit. That is, they can be located in the same location or distributed to multiple network units. Depending on the actual need, some or all of the units may be selected to achieve the objectives of the solution of this embodiment.

Also, each functional unit in each embodiment of the invention may be integrated into one processing unit, each unit may be physically independent, or two or more units may be integrated into one unit. Can be done. The integrated units described above can be implemented in the form of hardware or software functional units.

When the integrated unit described above is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in one computer-readable memory. Based on this understanding, the technical solution of the present invention, a part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in one memory, and by including several instructions, a computer device (which may be a personal computer, a server, a network device, or the like) is described in each embodiment of the present invention. Try to perform all or part of the steps in. The above-mentioned memory includes various media such as U disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), mobile hard disk, magnetic disk, or optical disk, which can store program code. Is included.

Those skilled in the art may complete all or part of the steps in the various methods of the embodiments described above by a program that directs the relevant hardware, which program may be stored in computer-readable memory and memory. Can be understood to include flash disks, read-only memory, random access memory, magnetic disks, or optical disks.

Examples of the present invention have been described in detail as described above, and in the present specification, the principles and embodiments of the present invention have been described by applying specific examples. The above description of the examples is used only to aid in understanding the method of the invention and its gist. At the same time, one of ordinary skill in the art may change the specific embodiments and scope of application in accordance with the ideas of the present invention, and as described above, the contents of the present specification should not be understood as a limitation to the present invention.

Claims (51)

It ’s a resource management method.
Reading the configuration file containing the configuration parameters of the virtual image processing unit (GPU) of the guest device,
Generating at least one available virtual GPU for the guest device based on the configuration file.
Including assigning a physical GPU to a target virtual GPU corresponding to the task to be processed in response to receiving a task to be processed by the target user.
The resource management method, wherein the at least one usable virtual GPU includes the target virtual GPU. Assigning a physical GPU to the target virtual GPU corresponding to the task to be processed is
Includes allocating a physical GPU to the target virtual GPU based on the video memory capacity and resource state table required for the target virtual GPU.
The resource management method according to claim 1, wherein the resource status table includes current status information of each physical GPU among a plurality of physical GPUs. The current state information of the physical GPU is
The video memory capacity of the physical GPU, the user information of each virtual GPU corresponding to the already occupied video memory of the physical GPU, and the video memory capacity of each virtual GPU corresponding to the already occupied video memory of the physical GPU. The resource management method according to claim 2, further comprising. Allocating a physical GPU to the target virtual GPU is based on the video memory capacity required for the target virtual GPU and the resource status table.
By allocating a physical GPU to the target virtual GPU based on the video memory capacity required for the target virtual GPU and the resource state table, the video memory of the plurality of physical GPUs is minimized. The resource management method according to claim 2 or 3, wherein the resource management method includes the above. Allocating a physical GPU to the target virtual GPU is based on the video memory capacity required for the target virtual GPU and the resource status table.
Based on the resource state table, according to the arrangement order of the plurality of physical GPUs, the remaining video memory capacity of each physical GPU is set until the remaining video memory capacity determined meets the video memory capacity required for the target virtual GPU. To decide sequentially and
Any of claims 2 to 4, wherein a physical GPU whose remaining video memory capacity satisfies the video memory capacity required for the target virtual GPU is determined as a physical GPU to be assigned to the target virtual GPU, and the like. The resource management method described in Section 1. If it is determined that the remaining video memory capacity of the last arranged physical GPU among the plurality of physical GPUs does not meet the video memory capacity required for the target virtual GPU, the physical resources available for the virtual GPU. The resource management method according to claim 5, further comprising determining that there is no such thing. Allocating a physical GPU to the target virtual GPU is based on the video memory capacity required for the target virtual GPU and the resource status table.
By allocating a physical GPU to the target virtual GPU based on the video memory capacity and resource state table required for the target virtual GPU, at least one task of the target user allocates to the plurality of physical GPUs as evenly as possible. The resource management method according to claim 2 or 3, wherein the at least one task including the above-mentioned task includes the task to be processed. Allocating a physical GPU to the target virtual GPU is based on the video memory capacity required for the target virtual GPU and the resource status table.
Based on the resource state table, the number of tasks currently assigned to each physical GPU among the plurality of physical GPUs of the target user, and the remaining video memory capacity of each physical GPU are determined.
It is characterized by including determining a physical GPU having the smallest number of tasks and having a remaining video memory capacity satisfying the video memory capacity required for the target virtual GPU as a physical GPU to be assigned to the target virtual GPU. The resource management method according to claim 2, 3, or 7. When the number of tasks is the smallest and the remaining video memory capacity is a plurality of physical GPUs that satisfy the video memory capacity required for the target virtual GPU, the number of tasks is the smallest and the remaining video memory. It further comprises determining the front-arranged physical GPU among a plurality of physical GPUs whose capacity meets the video memory capacity required for the target virtual GPU as the physical GPU to be assigned to the target virtual GPU. The resource management method according to claim 8. Updating the resource status table based on the physical GPU assigned to the target virtual GPU, or
The resource management method according to any one of claims 2 to 9, further comprising updating the resource status table based on the change in the task processing status when the task processing status changes. A resource management method that applies to guest devices
Reading the host's configuration file, which contains the virtual GPU configuration parameters for the guest device,
Generating at least one available virtual GPU for the guest device based on the configuration file.
When a task to be processed by the target user is received, it includes sending a resource allocation request to the host based on the target virtual GPU corresponding to the task to be processed.
The resource allocation request is used to request the host to allocate a physical GPU to the target virtual GPU, and the at least one available virtual GPU includes the target virtual GPU. Management method. Before sending a resource allocation request to the host based on the target virtual GPU corresponding to the task to be processed.
The resource management method is
Determining which virtual GPU is idle in at least one available virtual GPU of the guest device based on the first resource state table.
Further including determining the target virtual GPU from the virtual GPU in the idle state.
The resource management method according to claim 11, wherein the first resource status table includes current status information of each physical GPU among a plurality of physical GPUs. Receiving the first update instruction to the first resource status table from the host and
Further including updating the first resource status table based on the first update instruction.
The resource management method according to claim 12, wherein the first update instruction is transmitted when the host detects that the task processing state of at least one physical GPU has changed. The at least one physical GPU includes a first physical GPU having a newly added task, and the first update instruction includes user information corresponding to the newly added task of the first physical GPU, said first. At least one of the physical GPU information and the video memory capacity of the virtual GPU corresponding to the newly added task has been added and / or
The at least one physical GPU includes a second physical GPU having a newly completed task, and the first update instruction includes user information corresponding to the newly completed task and the second physical GPU. The resource management method according to claim 13, wherein at least one of the information is added. Scheduling the virtual GPU to execute the task to be processed,
The resource management method according to any one of claims 11 to 14, further comprising receiving a processing result of the task to be processed from the host. A resource management method that applies to the host
Receiving a resource allocation request from a guest device with information on the task to be processed and information on the target virtual GPU.
A resource management method comprising allocating a target physical GPU to the target virtual GPU based on the information of the target virtual GPU. Assigning a target physical GPU to the target virtual GPU based on the information of the target virtual GPU is possible.
The resource management method according to claim 16, further comprising allocating a physical GPU to the target virtual GPU based on the video memory capacity required for the target virtual GPU. Allocating a physical GPU to the target virtual GPU based on the video memory capacity required for the target virtual GPU
Includes allocating a physical GPU to the target virtual GPU based on the video memory capacity required for the target virtual GPU, a resource allocation policy, and a second resource status table.
The resource management method according to claim 17, wherein the second resource status table includes current status information of each physical GPU among a plurality of physical GPUs. Allocating a physical GPU to the target virtual GPU is based on the video memory capacity required for the target virtual GPU, the resource allocation policy, and the second resource status table.
Based on the second resource state table, according to the arrangement order of the plurality of physical GPUs, the remaining video of each physical GPU until the remaining video memory capacity determined meets the video memory capacity required for the target virtual GPU. Determining the memory capacity sequentially and
18. The resource according to claim 18, wherein the remaining video memory capacity determines a physical GPU to be assigned to the target virtual GPU as a physical GPU that satisfies the video memory capacity required for the target virtual GPU. Management method. If it is determined that the remaining video memory capacity of the last arranged physical GPU among the plurality of physical GPUs does not meet the video memory capacity required for the target virtual GPU, the physical resources available for the virtual GPU. 19. The resource management method according to claim 19, further comprising determining that there is no such thing. Allocating a physical GPU to the target virtual GPU is based on the video memory capacity required for the target virtual GPU, the resource allocation policy, and the second resource status table.
Based on the second resource status table, the target user's number of tasks already allocated to each physical GPU among the currently multiple physical GPUs, and the remaining video memory capacity of each physical GPU are determined.
Including determining a physical GPU with the least number of tasks and having the remaining video memory capacity satisfying the video memory capacity required for the target virtual GPU as the physical GPU to be assigned to the target virtual GPU.
The resource management method according to claim 18, wherein the target user is a user corresponding to the target virtual GPU. When the number of tasks is the smallest and the remaining video memory capacity is a plurality of physical GPUs that satisfy the video memory capacity required for the target virtual GPU, the number of tasks is the smallest and the remaining video memory. A claim further comprising determining the front-arranged physical GPU among the physical GPUs whose capacity meets the video memory capacity required for the target virtual GPU as the physical GPU to be assigned to the target virtual GPU. 21. The resource management method. Updating the second resource status table based on the physical GPU assigned to the target virtual GPU, or
The resource management according to any one of claims 16 to 22, further comprising updating the second resource status table based on the change in the task processing status when the task processing status changes. Method. The resource management method according to any one of claims 16 to 23, further comprising performing task scheduling based on the task queue of each physical GPU among a plurality of physical GPUs. It is a resource management device
Equipped with a communication unit and a processing unit
The processing unit is
Obtain the configuration file using the communication unit and
Read the configuration file containing the configuration parameters of the virtual image processing unit (GPU) of the guest device.
Generate at least one available virtual GPU for the guest device based on the configuration file.
In response to receiving the task to be processed by the target user, the physical GPU is assigned to the target virtual GPU corresponding to the task to be processed.
The resource management device, wherein the at least one usable virtual GPU includes the target virtual GPU. When allocating a physical GPU to the target virtual GPU corresponding to the task to be processed, the processing unit specifically
A physical GPU is assigned to the target virtual GPU based on the video memory capacity required for the target virtual GPU and the resource status table, and the resource status table displays the current status information of each physical GPU in the plurality of physical GPUs. 25. The resource management apparatus according to claim 25. The current state information of the physical GPU is
The video memory capacity of the physical GPU, the user information of each virtual GPU corresponding to the already occupied video memory of the physical GPU, and the video memory capacity of each virtual GPU corresponding to the already occupied video memory of the physical GPU. 26. The resource management device according to claim 26. When allocating a physical GPU to the target virtual GPU based on the video memory capacity required for the target virtual GPU and the resource status table, the processing unit specifically
By allocating a physical GPU to the target virtual GPU based on the video memory capacity required for the target virtual GPU and the resource state table, the video memory of the plurality of physical GPUs is minimized. The resource management device according to claim 26 or 27. When allocating a physical GPU to the target virtual GPU based on the video memory capacity required for the target virtual GPU and the resource status table, the processing unit specifically
Based on the resource state table, according to the arrangement order of the plurality of physical GPUs, the remaining video memory capacity of each physical GPU is set until the remaining video memory capacity determined meets the video memory capacity required for the target virtual GPU. Determined sequentially,
The invention according to any one of claims 26 to 28, wherein the physical GPU whose remaining video memory capacity satisfies the video memory capacity required for the target virtual GPU is determined as the physical GPU to be assigned to the target virtual GPU. Resource management unit. The processing unit further
If it is determined that the remaining video memory capacity of the last arranged physical GPU among the plurality of physical GPUs does not meet the video memory capacity required for the target virtual GPU, the physical resources available to the virtual GPU. 29. The resource management device according to claim 29, wherein it is determined that there is no such thing. When allocating a physical GPU to the target virtual GPU based on the video memory capacity required for the target virtual GPU and the resource status table, the processing unit specifically
By allocating a physical GPU to the target virtual GPU based on the video memory capacity and resource state table required for the target virtual GPU, at least one task of the target user allocates to the plurality of physical GPUs as evenly as possible. The resource management device according to claim 26 or 27, wherein the at least one task includes the task to be processed. When allocating a physical GPU to the target virtual GPU based on the video memory capacity required for the target virtual GPU and the resource status table, the processing unit specifically
Based on the resource status table, the number of tasks currently assigned to each physical GPU among the plurality of physical GPUs of the target user and the remaining video memory capacity of each physical GPU are determined.
26. , 27, or 31. The processing unit further
When the number of tasks is the smallest and the remaining video memory capacity is a plurality of physical GPUs that satisfy the video memory capacity required for the target virtual GPU, the number of tasks is the smallest and the remaining video memory. 32. Described resource management unit. The processing unit further
Update or update the resource status table based on the physical GPU assigned to the target virtual GPU.
The resource management device according to any one of claims 26 to 33, wherein when the task processing state changes, the resource status table is updated based on the change in the task processing state. It is a resource management device
Equipped with a communication unit and a processing unit
The processing unit is
Obtain the host configuration file using the communication unit and
Read the host's configuration file, including the guest device's virtual GPU configuration parameters.
Generate at least one available virtual GPU for the guest device based on the configuration file.
When a task to be processed by the target user is received, the communication unit is used to send a resource allocation request to the host based on the target virtual GPU corresponding to the task to be processed.
The resource allocation request is used to request the host to allocate a physical GPU to the target virtual GPU, and the at least one available virtual GPU includes the target virtual GPU. Management device. Before sending a resource allocation request to the host based on the target virtual GPU corresponding to the task to be processed, the processing unit further
Based on the first resource state table, the virtual GPU that is idle in at least one available virtual GPU of the guest device is determined.
The target virtual GPU is determined from the virtual GPUs in the idle state, and the target virtual GPU is determined.
The resource management device according to claim 35, wherein the first resource status table includes current status information of each physical GPU among a plurality of physical GPUs. The processing unit further
Upon receiving the first update instruction for the first resource status table from the host,
The first resource status table is updated based on the first update instruction, and the first resource status table is updated.
The resource management device according to claim 36, wherein the first update instruction is transmitted when the host detects that the task processing state of at least one physical GPU has changed. The at least one physical GPU includes a first physical GPU having a newly added task, and the first update instruction includes user information corresponding to the newly added task of the first physical GPU, said first. At least one of the physical GPU information and the video memory capacity of the virtual GPU corresponding to the newly added task has been added and / or
The at least one physical GPU includes a second physical GPU having a newly completed task, and the first update instruction includes user information corresponding to the newly completed task and the second physical GPU. The resource management device according to claim 37, wherein at least one of the information is added. The processing unit further
Schedule the virtual GPU to execute the task to be processed.
The resource management device according to any one of claims 35 to 38, wherein the processing result of the task to be processed is received from the host. It is a resource management device
Equipped with a communication unit and a processing unit
The processing unit is
The communication unit is used to receive a resource allocation request from a guest device to which information on a task to be processed and information on a target virtual GPU are added.
A resource management device characterized by allocating a target physical GPU to the target virtual GPU based on the information of the target virtual GPU. When assigning a target physical GPU to the target virtual GPU based on the information of the target virtual GPU, the processing unit specifically
The resource management device according to claim 40, wherein a physical GPU is assigned to the target virtual GPU based on the video memory capacity required for the target virtual GPU. When allocating a physical GPU to the target virtual GPU based on the video memory capacity required for the target virtual GPU, the processing unit specifically
A physical GPU is allocated to the target virtual GPU based on the video memory capacity required for the target virtual GPU, the resource allocation polysheet, and the second resource status table.
The resource management device according to claim 41, wherein the second resource status table includes current status information of each physical GPU among a plurality of physical GPUs. When allocating a physical GPU to the target virtual GPU based on the video memory capacity required for the target virtual GPU, the resource allocation policy, and the second resource status table, the processing unit specifically
Based on the second resource state table, according to the arrangement order of the plurality of physical GPUs, the remaining video of each physical GPU until the remaining video memory capacity determined meets the video memory capacity required for the target virtual GPU. Determine the memory capacity sequentially,
The resource management device according to claim 42, wherein the physical GPU whose remaining video memory capacity satisfies the video memory capacity required for the target virtual GPU is determined as the physical GPU to be assigned to the target virtual GPU. The processing unit further
If it is determined that the remaining video memory capacity of the last arranged physical GPU among the plurality of physical GPUs does not meet the video memory capacity required for the target virtual GPU, the physical resources available for the virtual GPU. The resource management device according to claim 43, wherein it is determined that there is no such thing. When allocating a physical GPU to the target virtual GPU based on the video memory capacity required for the target virtual GPU, the resource allocation policy, and the second resource status table, the processing unit specifically
Based on the second resource status table, the number of tasks already allocated to each physical GPU among the current plurality of physical GPUs of the target user and the remaining video memory capacity of each physical GPU are determined.
A physical GPU having the smallest number of tasks and having the remaining video memory capacity satisfying the video memory capacity required for the target virtual GPU is determined as the physical GPU to be assigned to the target virtual GPU.
The resource management device according to claim 42, wherein the target user is a user corresponding to the target virtual GPU. The processing unit further
When the number of tasks is the smallest and the remaining video memory capacity is a plurality of physical GPUs that satisfy the video memory capacity required for the target virtual GPU, the number of tasks is the smallest and the remaining video memory. 45. The 45. Resource management unit. The processing unit further
Update or update the second resource status table based on the physical GPU assigned to the target virtual GPU.
The resource management device according to any one of claims 40 to 46, wherein when the task processing state changes, the second resource status table is updated based on the change in the task processing state. The processing unit further
The resource management device according to any one of claims 40 to 47, wherein task scheduling is executed based on the task queue of each physical GPU among a plurality of physical GPUs. It ’s a resource management system.
It has a guest device and a host with multiple physical GPUs.
The guest device is the device according to any one of claims 35 to 39.
The resource management system, wherein the host is the apparatus according to any one of claims 40 to 48. It ’s an electronic device,
It has a processor and a memory to store computer programs.
The method according to any one of claims 1 to 10 or 11 to 15 or any one of 16 to 24 when the computer program is executed by the processor. An electronic device characterized by running. A computer-readable recording medium
The computer-readable recording medium stores a computer program and stores the computer program.
The computer program is characterized in that the computer performs the method according to any one of claims 1 to 10 or any one of claims 11 to 15 or any one of 16 to 24. Computer-readable recording medium.

Images