CN117707693A - Heterogeneous intelligent computing platform virtualization management system and method - Google Patents

Abstract

The invention provides a heterogeneous intelligent computing platform virtualization management system and method, comprising the following steps: the chip virtualization module is used for virtualizing the CPU, the GPU and the FPGA; the computing virtualization module is used for providing a dynamic computing resource pool based on the super-heterogeneous platform; and the network virtualization module is used for creating a virtual network above the physical network and providing communication between the virtual machines and an external network. And the dynamic scheduling management module is used for determining required calculation, memory, storage, network resources and the like according to the application scene and the requirements. The invention is oriented to the image intelligent processing and computing requirement, constructs a heterogeneous intelligent computing resource pool combining virtual and real, and constructs a virtualization system for generating, debugging and optimizing an intelligent computing system. Meanwhile, the resource allocation of the virtual machine is automatically expanded or contracted by periodically monitoring the change of the resource utilization rate in the virtualized environment. The method realizes the maximum utilization of computing resources, reduces the deployment cost and improves the service performance.

Description

A heterogeneous intelligent computing platform virtualization management system and method

Technical field

The present invention relates to the field of computer technology, and in particular to a heterogeneous intelligent computing platform virtualization management system and method.

Background technique

As the complexity and dynamic capabilities of computing platforms continue to increase, it is necessary to efficiently manage the computing resources and network storage of end-side platforms. Traditional virtualization technology may encounter the following problems in practice: First, heterogeneous platform virtualization needs to support a variety of different operating systems and hardware architectures, and each architecture has its unique technical requirements and challenges. This requires virtualization technology to be highly adaptable and flexible and able to cope with various workloads and environments; second, virtualization technology will introduce additional overhead, including the operation of the virtual machine manager, resource allocation, and task scheduling. . These overheads may affect the performance and efficiency of the system, especially when resources are limited; third, different virtual machines may require different configuration and management strategies, which requires administrators to have a high degree of technical knowledge and experience. In addition, the deployment, management and maintenance of virtual machines also require a lot of time and resources. At present, traditional virtualization technology runs inefficiently on domestic heterogeneous platforms and cannot meet the needs of real-time applications.

Contents of the invention

The purpose of the present invention is to provide a heterogeneous intelligent computing platform virtualization management system and method to solve the problem of heterogeneous platform virtualization system resource allocation in view of the shortcomings of the existing technology.

The object of the present invention is achieved through the following technical solutions: a heterogeneous intelligent computing platform virtualization management system, including:

Chip virtualization module, used for virtualization of CPU, GPU, and FPGA;

The computing virtualization module is used to monitor and manage the status information and availability information of computing resources, and provide the information to the dynamic scheduling management module;

The network virtualization module is used to use virtual software-defined network technology to enable different users to share the network resources of the same physical network, and segment the physical network resources according to needs to form a logically independent virtual flow distribution network vSDN for users to use;

The dynamic scheduling management module is used to determine the host load status, and then select virtual machines that meet the migration conditions based on the minimum migration number policy to achieve static placement where the mapping relationship between the virtual machine and the host remains unchanged or dynamic placement where the mapping relationship between the virtual machine and the host is variable. ; After completing the virtual machine selection, the host with the least increase in energy consumption will be used as the target host for virtual machine migration.

Furthermore, CPU virtualization allows a single physical CPU to be virtualized into multiple vCPUs, that is, virtual CPUs. The user operating system of each virtual machine uses one or more parallel vCPUs. Each vCPU runs independently of each other and supports the CPU. Time-sharing multiplexing uses real-time scheduling strategies to achieve CPU sharing of tasks.

Furthermore, GPU virtualization uses hardware-assisted virtualization technology based on SR-IOV to virtualize PCIe devices. By enabling the SR-IOV function on the GPU, multiple virtual GPUs are divided, and each virtual GPU has its own The identification and resources of the virtual machine or container directly access the virtual GPU to realize the isolation and sharing of GPU resources.

Furthermore, FPGA virtualization divides a single FPGA into multiple fine-grained partially reconfigurable vFPGAs, that is, virtual FPGAs. Each vFPGA is managed by a separate controller and is connected to the outside world through the AXI bus. On the one hand, it is used to access memory, and on the other hand, it is used for data interaction with the dynamic scheduling management module.

Further, the computing virtualization module regularly updates the load status, performance indicators and availability information of the resources, and provides this information to the dynamic scheduling management module. According to the needs of the task or application, the task scheduler and decision engine perform task scheduling. Scheduling and allocating resources.

Furthermore, the virtual network controller in the network virtualization module simultaneously accepts user requests and divides physical network resources according to needs to form logically independent vSDN for users to use. The establishment of vSDN mainly includes two aspects: network management program NVH and virtual network mapping VNE; NVH is located between the user SDN controller and the physical network, and is implemented through Flowvisor, the network virtualization platform of the network communication protocol Openflow; VNE allocates network resources to each virtual network, which is divided into two parts: node mapping and link mapping; for Node mapping ensures that physical resource nodes do not exceed capacity limits, while link mapping maps a virtual link to a physical path; every time the user status changes, virtual nodes and links are mapped to new physical paths based on vSDN reconfiguration. nodes and physical links.

Further, the host status monitoring is used to determine whether the host status is in an overload or underload state, using the adaptive dynamic threshold method to make the determination; using machine learning methods to learn dynamic and adaptive resource utilization thresholds, and at the same time during the learning process The learning results are strengthened through interaction and trial and error with the dynamic environment to adapt to the changing environment, and a dual threshold method of overload and underload is used to trigger virtual machine scheduling when overloaded and underloaded.

Furthermore, the specific process of selecting virtual machines that meet the migration conditions based on the minimum migration quantity policy is: first, sort the virtual machines in descending order according to the resource requirements of the virtual machines, and then select the virtual machines that meet the conditions. After the virtual machine migration is finally completed, the remaining virtual machines on the host If the resource requirements of the machine are less than the maximum capacity of the host, select the virtual machine with the smallest resource requirements.

On the other hand, the present invention provides a heterogeneous intelligent computing platform virtualization management method, including:

Based on the needs of human-computer interaction end users, obtain corresponding CPU, GPU and FPGA computing resources;

Monitor and manage the status information and availability information of computing resources, and generate virtual machine resource allocation instructions based on the required computing resources;

Send resource allocation instructions to each chip module to virtualize computing resources, and then mount the virtualized resources to the target virtual machine;

The use of virtual software-defined network technology allows different users to share the network resources of the same physical network, and splits the physical network resources according to needs to form a logically independent virtual flow distribution network vSDN for users to use;

Based on the minimum migration quantity strategy, select virtual machines that meet the migration conditions, and implement static placement where the mapping relationship between the virtual machine and the host remains unchanged or dynamic placement where the mapping relationship between the virtual machine and the host is variable; after completing the virtual machine selection, minimize the increase in energy consumption The host is used as the target host for virtual machine migration.

Further, the load status, performance indicators and availability information of computing resources are collected and analyzed; this information is used to evaluate resource utilization, bottlenecks and performance bottlenecks, and is used in the decision-making engine to make scheduling decisions and optimization; through continuous monitoring and memory, network And storage optimization, improve resource utilization, reduce task execution time and cost.

Beneficial effects of the present invention:

(1) Break through the bottleneck of heterogeneous hardware cluster computing power management and scheduling, achieve integrated scheduling of heterogeneous computing power, and improve hardware resource utilization efficiency

(2) For deep learning model training, heterogeneous computing power virtualization can help achieve distributed training and greatly improve training efficiency.

(3) For multi-task concurrency scenarios, heterogeneous computing power virtualization can realize multi-task parallelism without interfering with each other, ensuring safe and reliable operation of tasks.

Description of the drawings

In order to explain the embodiments of the present invention or the technical solutions in the prior art more clearly, the drawings needed to be used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings in the following description are only These are some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without exerting creative efforts.

Figure 1 is a schematic structural diagram of a heterogeneous intelligent computing platform virtualization management system provided by the present invention;

Figure 2 is a schematic diagram of CPU virtualization provided by the present invention;

Figure 3 is a schematic diagram of GPU virtualization provided by the present invention;

Figure 4 is a flow chart of the status monitoring system provided by the present invention;

Figure 5 is a flow chart of the heterogeneous intelligent computing platform virtualization management method provided by the present invention;

Figure 6 is a possible hardware schematic diagram provided by the present invention;

Detailed ways

The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only some of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts fall within the scope of protection of the present invention.

It should be noted that, as long as there is no conflict, the features in the following embodiments and implementation modes can be combined with each other.

Figure 1 is a schematic structural diagram of a heterogeneous intelligent computing platform virtualization management system provided by the present invention. As shown in Figure 1, the system includes a chip virtualization module, a computing virtualization module, a storage virtualization module, a network virtualization module, a dynamic Scheduling management module and underlying KVM module.

The chip virtualization module mainly includes a CPU virtualization sub-module, a GPU virtualization sub-module and an FPGA virtualization sub-module, which are used to virtualize CPU, NPU, GPU and FPGA; the computing virtualization module is used to virtualize the CPU, NPU, GPU and FPGA according to the user's needs. Based on the domestic heterogeneous platform, it provides a dynamic computing resource pool; the storage virtualization module is used to build a virtual storage management layer and realize unified management and scheduling of virtual storage resources. And build a scalable storage framework that can be seamlessly expanded with business development; the network virtualization module is used to create a virtual network on top of the physical network to provide communication between virtual machines and between virtual machines and external networks. The dynamic scheduling management module is used to determine the required computing, memory, storage and network resources based on application scenarios and needs.

It should be noted that before using the system of the present invention, a KVM underlying system can be deployed and built on a physical server, thereby providing virtualization services through the above KVM underlying system.

It can be understood that in the above-mentioned CPU virtualization, the CPU may be an ARM architecture or an X86 architecture, and this embodiment does not limit this.

It should be understood that the above-mentioned CPU virtualization sub-module allows a single physical CPU to be virtualized into multiple vCPUs (virtual CPUs), see Figure 2. Each virtual machine's user operating system can use one or more vCPUs. Each vCPU runs independently of each other without interfering with each other, achieving multi-core parallelism. At the same time, it supports time-sharing multiplexing of CPU, that is, through real-time scheduling strategy, CPU sharing of tasks is realized.

The CPU virtualization includes a real-time scheduling unit. Mainly using the prioritization method, first classify the priorities according to the task type and urgency. During task execution, tasks with high priority are executed first. If a higher-priority task is inserted during the running process, an interrupt can be used to ensure that the high-priority task runs.

It should be understood that in the above-mentioned GPU virtualization, the corresponding virtual GPU is obtained according to the resource allocation instruction, and the virtual GPU is mounted to the default virtual machine, see Figure 3. After use is completed, uninstall the virtual GPU mounted in the preset virtual machine and release the virtual GPU resources. For NPU virtualization, you can refer to GPU virtualization, and the technical solutions are basically the same.

In specific implementation, the GPU virtualization sub-module uses hardware-assisted virtualization technology based on SR-IOV (Single Root I/OVirtualization) to virtualize PCIe devices. By enabling the SR-IOV function on the GPU, multiple virtual GPUs can be divided, each with its own identity and resources. First, based on the login command sent by the user, obtain the user's corresponding GPU resource information, generate resource allocation instructions based on the GPU resource information, obtain the corresponding virtual GPU resources based on the resource allocation instructions, and mount the virtual GPU resources to the preset virtual machine. Provide users with GPU computing power based on preset virtual machines. Virtual machines or containers can directly access virtual GPUs to isolate and share GPU resources.

FPGA virtualization divides a single FPGA into multiple fine-grained partially reconfigurable virtual FPGAs (vFPGA), which can be individually configured and used by users. Each vFPGA has a separate controller for management, which is connected to the outside world through the AXI bus. On the one hand, it is used to access memory, and on the other hand, it is used for data interaction with other modules. In specific implementation, the FPGA virtualization sub-module mainly divides FPGA logical resources into static areas and dynamic areas. The static area is a non-configurable area, and the dynamic area is a plurality of dynamically reconfigurable vFPGA areas; the host CPU side Software design: Application deployment, FPGA resource management, and communication with the vFPGA area through user application APIs; application scheduling in space and time through the runtime manager; multiple dynamically reconfigurable devices are instantiated through the driver vFPGA area, set up the required data structures for the vFPGA area respectively, and create a virtual memory map for the application to communicate with the FPGA through the PCIe bus. A single vFPGA area contains two parts: user logic and dynamic wrapper. User logic is a bit stream that has been synthesized and system verified. Users can develop applications using languages including HLS, Verilog, VHDL, and OpenCL; the dynamic wrapper provides a standard interface for user logic so that applications can run across vFPGAs.

As an embodiment, the computing virtualization module is responsible for monitoring and managing the status and availability of computing resources. It regularly updates the load status, performance indicators and availability information of resources, and provides this information to the task scheduler and decision engine of the dynamic scheduling management module. It requires the discovery and description of heterogeneous resources existing in the system. This includes identifying various resource types in the system (such as CPU, GPU, FPGA, etc.) and obtaining their characteristics and performance information for subsequent management and scheduling. And according to the needs of the task or application, tasks are reasonably scheduled and allocated to appropriate resources. The computing virtualization module includes a resource discovery and description submodule, a resource monitoring and adjustment submodule, and a fault handling and fault tolerance submodule.

Among them, the resource discovery and description submodule mainly discovers and describes the heterogeneous resources existing in the system. This includes identifying various resource types in the system (such as CPU, GPU, FPGA, etc.) and obtaining their characteristics and performance information for subsequent management and scheduling of the dynamic scheduling management module.

It can be understood that the resource monitoring and adjustment sub-module is mainly used to monitor the usage, load status and performance indicators of heterogeneous resources, and adjust the allocation and configuration of resources in a timely manner. Through real-time monitoring and feedback mechanisms, resource allocation strategies can be dynamically adjusted to adapt to system changes and optimize resource utilization efficiency.

It should be understood that the fault handling and fault tolerance sub-module is to consider the processing strategy of resource faults and abnormal situations. When a resource fails, corresponding fault tolerance mechanisms and alternatives need to be adopted to ensure the reliability and continuity of the system.

As an embodiment, the network virtualization module allows different users to share network resources of the same physical network, using virtual software-defined network technology (vSDN). A virtual network is a virtual topology composed of a set of virtual nodes and virtual links. It is a subset of the physical topology and an entity based on the network virtualization environment. The mapping relationship is that virtual nodes are mapped to physical nodes, and multiple virtual nodes can coexist. Virtual links are mapped onto physical links. Virtual network mapping is to establish a virtual network (including topology, resource requirements, location restrictions and other elements) based on user requests. The virtual nodes are logically deployed on the corresponding physical nodes of the base network whose location and resources meet the requirements. The links are connected by paths that meet resource constraints and other conditions between the physical nodes where the corresponding virtual nodes are deployed. At the same time, the nodes and Link allocation corresponds to virtual network requests, and resource virtual network mapping implements the instantiation or initialization of a virtual network request.

Since the network environment used by users is in dynamic change, as user requests continue to arrive and leave, the resource requirements of network slicing will also change accordingly. The traditional optimal solution of virtual network mapping (VNE) cannot solve the problem of unreasonable resource allocation. asked. The use of virtual software-defined networks can control the network in real time, enhance the programmability of the network, and improve user flexibility. In addition to resource allocation, virtual software-defined networks also have good applicability in fault recovery and mobility management.

The virtual network controller in the present invention accepts user requests at the same time, and can segment physical network resources according to needs to form logically independent vSDN for users to use. Establishing vSDN mainly includes two aspects: network hypervisor (NVH) and virtual network mapping (VNE). NVH is located between the user SDN controller and the physical network and can be implemented through Openflow's Flowvisor; VNE allocates network resources to each virtual network and is divided into two parts: node mapping and link mapping. Node mapping ensures that physical resource nodes do not exceed capacity limits, while link mapping maps a virtual link to a physical path. In order to solve the dynamic problem of users constantly arriving and exiting, vSDN reconfiguration is proposed to map virtual nodes and links to new physical nodes and physical links every time the user status changes.

As an embodiment, the dynamic scheduling management module mainly includes a host status monitoring sub-module, a virtual machine selection sub-module and a virtual machine placement sub-module.

It is understandable that in order to achieve efficient utilization of heterogeneous resources and optimal allocation of tasks, the system needs to monitor the status and performance indicators of resources in real time. The host status monitoring sub-module is responsible for collecting and analyzing resource load, performance indicators and availability information. This information can be used to evaluate resource utilization, bottlenecks, and performance bottlenecks, and is used by the dynamic scheduling management module to make scheduling decisions and optimizations. Through continuous monitoring and optimization, the system can improve resource utilization and reduce task execution time and cost. The optimization mainly includes memory, network and storage optimization. Network optimization refers to pooling physical network cards to reduce the number of data copies; storage optimization refers to preventing virtual machines from overusing shared resources and preventing the performance of other virtual machines from being affected. Specific strategies include: establishing performance logs, minimizing the impact of performance monitoring itself on the server, analyzing monitoring results, establishing performance baselines and creating alarms.

The host status monitoring sub-module is mainly used to determine whether the host status is overloaded or underloaded, and the adaptive dynamic threshold method is proposed to be used for determination. Use machine learning methods to learn dynamic and adaptive resource utilization thresholds. At the same time, during the learning process, the learning results are strengthened through interaction with the dynamic environment and trial and error to adapt to the changing environment. A dual-threshold method is used to trigger virtual machine scheduling when overloaded and underloaded.

In the specific implementation, as shown in Figure 4, the host status monitoring sub-module is used to monitor performance indicators in the virtualized environment in real time, such as CPU, memory and network latency, and set up an alarm mechanism to handle abnormal situations in a timely manner. The present invention uses Prometheus for system monitoring and alarming. Prometheus is an open source monitoring and alarming system based on a time series database and is very suitable for monitoring Kubernetes clusters. The basic principle of Prometheus is to periodically capture the status of monitored components through the HTTP protocol. Any component can be accessed for monitoring as long as it provides the corresponding HTTP interface. No SDK or other integration process is required. This is very suitable for virtualization environment monitoring systems, such as VM, Docker, Kubernetes, etc. By installing the relevant Agent or Exporter on the virtual machine, the performance data of the virtual machine is pushed to Prometheus and visually displayed and analyzed through Grafana.

It should be understood that if the host status monitoring sub-module can monitor and predict that the host will be overloaded, it can stop allocating virtual machines to the host, or migrate some virtual machines on the host in advance to avoid possible SLA violations; if it can accurately predict If it is predicted that the host will be lightly loaded at the next moment, a series of operations can be performed to save energy. You can stop allocating virtual machines to the host, then migrate all virtual machines on the host, and finally shut down the host or convert it to Sleep mode to save energy.

As an embodiment, the virtual machine selection sub-module adopts the minimum migration number strategy to reduce the number of online migration virtual machines. First, sort the virtual machines in descending order according to their resource requirements, and then select the virtual machines that meet the conditions. The main method is: only when the computing requirements of all virtual machines running on the host exceed the maximum capacity of the host, the virtual machine will be migrated; secondly , after the virtual machine migration, the computing requirements (CPU, GPU, NPU, FPGA, etc.) of all remaining virtual machines on the host should be less than the maximum capacity of the host; finally, among all virtual machines that meet the above two conditions, select computing resources The smallest virtual machine is migrated to another host, and if there is no virtual machine that meets the conditions, the virtual machine with the largest computing capacity is migrated.

In the specific implementation, the virtual machine selection sub-module first obtains a list of all virtual machines on the host. If the virtual machine list is empty, it exits because the host has no virtual machines to migrate at this time. If the virtual machine list is not empty, continue with virtual machine selection. The amount of computing resources that the computing host needs to release is the difference between the total computing resources requested by the virtual machine and the total computing resources of the host multiplied by the reservation coefficient. If the current amount of computing resources is less than 0, exit because the host has sufficient resources and there is no need to migrate the virtual machine. If the current amount of computing resources is greater than 0, that is, the host resources cannot meet the needs of the virtual machines, and virtual machine migration needs to be performed at this time, so the virtual machine selection process continues. Then start traversing the virtual machine list on the host to select the appropriate virtual machine.

As an embodiment, the virtual machine placement sub-module can be divided into static placement and dynamic placement. Static placement means that the mapping relationship between the virtual machine and the host remains unchanged during the entire life cycle; and dynamic placement means that the mapping relationship between the two is variable. It is planned to adopt the energy consumption-aware best adaptation strategy. After completing the virtual machine selection, the host with the least increase in energy consumption will be used as the target host.

The specific implementation of the virtual machine placement sub-module is as follows: first, arrange the virtual machines to be placed in descending order according to the computing resource utilization, and then select the target placement host for the virtual machine according to the virtual machine computing resource utilization from high to low. For a virtual machine, the host list is traversed looking for a suitable target host.

In the specific implementation, for the target host, first use the host status detection module to detect. If the detection result is that the host will be overloaded or lightly loaded at the next moment, then the host will not be used as the target host for virtual machine placement because it is overloaded or lightly loaded. The host is not an ideal target host for the virtual machine. If the host status is normal, it will be further determined whether the host has enough resources to accommodate the virtual machine. If not, the host will not be placed as the target of the virtual machine; if the host has sufficient resources to accommodate the virtual machine, then the virtual machine will be calculated. The energy consumption increase value after the virtual machine is placed on the host is calculated, and the host with the smallest energy consumption increase value is selected from all target hosts that meet the conditions as the target host for placing the virtual machine.

Please refer to Figure 5. Figure 5 is a flow chart of a heterogeneous intelligent computing platform virtualization management method provided by an embodiment of the present invention. As shown in Figure 5, the present invention provides a heterogeneous intelligent computing platform virtualization management method. include:

Based on the user needs of the human-computer interaction end, obtain the corresponding computing (CPU, GPU, NPU, FPGA, etc.), memory, storage and network resources, etc.;

It should be noted that the execution subject of the method of this embodiment can be a computer terminal device with data processing, network communication and program running functions, or a server device with the same similar functions, or a server with similar functions. The embodiment does not limit this. To facilitate understanding, this embodiment and the following embodiments will be described using a server device as an example.

Based on the required computing resources, define resource allocation strategies based on requirements, including resource quotas and priorities, and generate virtual machine resource allocation instructions;

Send resource allocation instructions to each chip module for virtualizing computing resources. Based on requirements, use Hypervisor to create a KVM virtual machine instance and allocate appropriate computing, memory, and storage resources to it. After creation, the virtual machine needs to be initialized. The purpose of initialization is to provide the necessary software and hardware environment for the creation and operation of the virtual machine. Configure and manage virtual machines through the virtual machine manager, including network settings, operating system installation, etc. Then mount the virtual computing resources to the computer.

Regularly monitor resource utilization in virtualized environments, such as computing resources, memory, and storage usage.

According to changes in resource utilization, automatically expand or shrink the resource allocation of virtual machines to meet actual needs. Among them, the system can automatically adjust the memory size according to operating conditions to optimize resource utilization.

It should be noted that the memory adjustment strategy is set based on actual needs and performance optimization goals. It mainly includes the following aspects: memory threshold, which sets a threshold for memory utilization and triggers memory adjustment when the memory utilization of the virtual machine exceeds or falls below the threshold; adjustment amplitude, which determines the increase or decrease of memory adjustment, for example The memory block size increased/decreased each time; adjustment frequency, set the detection frequency of memory adjustment, that is, how often to monitor memory usage and make adjustments.

According to the set memory adjustment policy, take corresponding operations to adjust the memory size of the virtual machine. When the memory utilization of the virtual machine exceeds the threshold, the memory size of the virtual machine is dynamically increased. This can be done through the API provided by the virtual machine manager (libvirt) to allocate additional memory to the virtual machine. When a virtual machine's memory utilization is low, you can reduce the virtual machine's memory size. This can be done through the API provided by the virtual machine manager or by using tools in the virtual machine to release some memory resources.

The use of virtual software-defined network technology enables different users to share the network resources of the same physical network, and splits the physical network resources according to needs to form a logically independent virtual flow distribution network vSDN for users to use;

Based on the minimum migration quantity strategy, select virtual machines that meet the migration conditions, and implement static placement where the mapping relationship between the virtual machine and the host remains unchanged or dynamic placement where the mapping relationship between the virtual machine and the host is variable; after completing the virtual machine selection, minimize the increase in energy consumption The host is used as the target host for virtual machine migration.

After the step of providing required computing power to the user based on the preset virtual machine, the method further includes:

Based on the user's shutdown instruction, obtain the virtual machine corresponding to the shutdown instruction;

Uninstall the virtual computing resources in the virtual machine (CPU, GPU, NPU, FPGA, etc.), and close the virtual machine after the uninstallation is completed;

After the step of shutting down the virtual machine after the uninstallation is completed, the method also includes:

Detect the mounting status of the virtual computing resource of the virtual machine that is in the shutdown state in the current virtual environment. If it is being mounted, uninstall the virtual GPU of the virtual machine.

Please refer to Figure 6, which is a possible hardware schematic diagram provided by the present invention. As shown in Figure 6, an embodiment of the present invention provides an electronic device, including a memory, a CPU processor, a GPU processor, an FPGA, an NPU, and a computer program stored in the memory and capable of running on the CPU processor. The processor When the computer program is executed, a heterogeneous intelligent computing platform virtualization management method provided by the present invention is implemented.

The above embodiments are used to illustrate the present invention, rather than to limit the present invention. Within the spirit of the present invention and the protection scope of the claims, any modifications and changes made to the present invention fall within the protection scope of the present invention.

Claims (10)

1. A heterogeneous intelligent computing platform virtualization management system, which is characterized by including: Chip virtualization module, used for virtualization of CPU, GPU, and FPGA; The computing virtualization module is used to monitor and manage the status information and availability information of computing resources, and provide the information to the dynamic scheduling management module; The network virtualization module is used to use virtual software-defined network technology to enable different users to share the network resources of the same physical network, and segment the physical network resources according to needs to form a logically independent virtual flow distribution network vSDN for users to use; The dynamic scheduling management module is used to determine the host load status, and then select virtual machines that meet the migration conditions based on the minimum migration number policy to achieve static placement where the mapping relationship between the virtual machine and the host remains unchanged or dynamic placement where the mapping relationship between the virtual machine and the host is variable. ; After completing the virtual machine selection, the host with the least increase in energy consumption will be used as the target host for virtual machine migration. 2. A heterogeneous intelligent computing platform virtualization management system according to claim 1, characterized in that CPU virtualization allows a single physical CPU to be virtualized into multiple vCPUs, that is, virtual CPUs, and user operations of each virtual machine The system uses one or more parallel vCPUs, each vCPU runs independently of each other, and supports time-sharing multiplexing of the CPU, that is, through real-time scheduling strategy, the CPU sharing of tasks is realized. 3. A heterogeneous intelligent computing platform virtualization management system according to claim 1, characterized in that GPU virtualization adopts hardware-assisted virtualization technology based on SR-IOV to realize virtualization of PCIe devices. Enable the SR-IOV function on the GPU and divide it into multiple virtual GPUs. Each virtual GPU has its own identity and resources. The virtual machine or container directly accesses the virtual GPU to isolate and share GPU resources. 4. A heterogeneous intelligent computing platform virtualization management system according to claim 1, characterized in that FPGA virtualization divides a single FPGA into multiple fine-grained partially reconfigurable vFPGAs, that is, virtual FPGAs. , Each vFPGA has a separate controller for management, which is connected to the outside world through the AXI bus. On the one hand, it is used to access memory, and on the other hand, it is used for data interaction with the dynamic scheduling management module. 5. A heterogeneous intelligent computing platform virtualization management system according to claim 1, characterized in that the computing virtualization module regularly updates the load status, performance indicators and availability information of resources, and provides these information to The dynamic scheduling management module schedules tasks and allocates resources through the task scheduler and decision engine according to the needs of tasks or applications. 6. A heterogeneous intelligent computing platform virtualization management system according to claim 1, characterized in that the virtual network controller in the network virtualization module simultaneously accepts user requests and divides physical network resources into logical An independent vSDN is provided to users. The establishment of vSDN mainly includes two aspects: network management program NVH and virtual network mapping VNE; NVH is located between the user SDN controller and the physical network, and is implemented through the network communication protocol Openflow's network virtualization platform Flowvisor; VNE will Network resources are allocated to each virtual network and are divided into two parts: node mapping and link mapping; for node mapping, it is ensured that the physical resource nodes do not exceed the capacity limit, while link mapping ensures that one virtual link mapping corresponds to one physical path; in each When the secondary user status changes, virtual nodes and links are mapped to new physical nodes and physical links based on vSDN reconfiguration. 7. A heterogeneous intelligent computing platform virtualization management system according to claim 1, characterized in that the host status monitoring is used to determine whether the host status is in an overload or underload state, and an adaptive dynamic threshold method is used to determine. ; Use machine learning methods to learn dynamic and adaptive resource utilization thresholds, and at the same time strengthen the learning results through interaction and trial and error with the dynamic environment during the learning process to adapt to the changing environment, using dual thresholds of overload and underload method to trigger virtual machine scheduling when overloaded and underloaded. 8. A heterogeneous intelligent computing platform virtualization management system according to claim 1, characterized in that the specific process of selecting virtual machines that meet migration conditions based on the minimum migration number strategy is: first, perform descending order according to the resource requirements of the virtual machines. Arrange, and then select the virtual machines that meet the conditions. After the virtual machine migration is completed, the resource requirements of the remaining virtual machines on the host are less than the maximum capacity of the host. Select the virtual machine with the smallest resource requirements. 9. A virtualization management method for heterogeneous intelligent computing platforms, which is characterized by including: Based on the needs of human-computer interaction end users, obtain corresponding CPU, GPU and FPGA computing resources; Monitor and manage the status information and availability information of computing resources, and generate virtual machine resource allocation instructions based on the required computing resources; Send resource allocation instructions to each chip module to virtualize computing resources, and then mount the virtualized resources to the target virtual machine; The use of virtual software-defined network technology allows different users to share the network resources of the same physical network, and splits the physical network resources according to needs to form a logically independent virtual flow distribution network vSDN for users to use; Based on the minimum migration quantity strategy, select virtual machines that meet the migration conditions, and implement static placement where the mapping relationship between the virtual machine and the host remains unchanged or dynamic placement where the mapping relationship between the virtual machine and the host is variable; after completing the virtual machine selection, minimize the increase in energy consumption The host is used as the target host for virtual machine migration. 10. A heterogeneous intelligent computing platform virtualization management method according to claim 9, characterized by collecting and analyzing load conditions, performance indicators and availability information of computing resources; these information are used to evaluate resource utilization and Performance bottlenecks are used in the decision-making engine for scheduling decisions and optimization; through continuous monitoring and memory, network and storage optimization, resource utilization is improved and task execution time and costs are reduced.