CN102650950B - Platform architecture supporting multi-GPU (Graphics Processing Unit) virtualization and work method of platform architecture - Google Patents

Abstract

The present invention provides a platform framework supporting multi-GPU virtualization and its working method. The framework deploys middleware on the GPU server side and the virtual machine side, and uses socket or infiniband as the transmission medium to fill the original virtual machine The platform cannot take advantage of the disadvantages of GPU acceleration. This architecture manages GPU resources through one or more centralized control management nodes, and divides GPU resources in a fine-grained manner, and can provide the function of multi-task parallel tasks. The virtual machine requests GPU resources from the management node through the middleware, and uses the GPU resources to accelerate. The GPU server registers GPU resources with the management node through the middleware, and uses GPU resources to provide services. The invention introduces the parallel processing ability of the GPU into the virtual machine and combines the management mechanism to maximize the utilization rate of the GPU. This architecture can effectively reduce energy consumption and improve computing efficiency.

Description

A platform architecture supporting multi-GPU virtualization and its working method

technical field

The invention relates to a platform architecture and a working method for a virtual machine in a virtualized environment to accelerate computing by using multiple GPUs in a redirection manner, and belongs to the technical field of virtualization.

Background technique

Virtualization is the core technical basis of cloud computing, and its advantages such as cost saving and security enhancement have gradually been recognized by people, and it is a research hotspot in the field of computer science. Through the virtualization of hardware resources, virtualization technology can simulate multiple identical computer hardware platforms on one computer, so that multiple operating systems can run simultaneously and be isolated from each other, which improves the utilization efficiency of servers. There are numerous applications in areas such as network security, computational data protection, high-performance computing, and trustworthiness.

The performance and capabilities of Graphics Processing Units (GPUs) have increased significantly in recent years. The function of GPU is no longer limited to image processing, and has developed into a highly parallel processor with high computing peak and memory bandwidth. With the introduction of some technologies that support GPGPU (general-purpose graphics processing unit) computing (such as CUDA), the application of GPGPU is becoming more and more extensive. Due to the powerful parallel computing capability of GPU, the heterogeneous mode of CPU+GPU has been introduced into more and more high-performance computing. However, on the one hand, the power consumption of the GPU is large. If each node is equipped with a GPU, it may greatly increase the power consumption of the cluster; , only accelerates the parallel part of the code, so that the utilization rate of the GPU is not high; on the other hand, due to the closed nature of the GPU, the virtual machine cannot directly use the GPU to accelerate the calculation. This greatly limits the application of GPU in virtual machines.

Contents of the invention

technical problem

In order to solve the problem that the virtualization environment cannot use GPU acceleration, the present invention proposes a multi-GPU virtualization platform architecture and its working method applicable to clusters. Through the cooperative work of management components, client components, and server components, the virtual machine can Obtaining the powerful parallel computing capability of the GPU can realize the acquisition of GPU resources and fine-grained resource allocation, and through load balancing of the GPU, the utilization rate of the GPU can be improved and the energy consumption can be reduced. The present invention realizes the improvement of the processing ability of the virtual machine by using the GPU. The components in the virtual machine intercept the application program calling the GPU and redirect it to the local privilege domain or the remote GPU server, so that the GPU call of the application program is in the privilege domain or the remote GPU server. Execute on the GPU server, and return the result to the virtual machine after the execution is completed.

Technical solutions

The present invention adopts following technical scheme in order to solve the problems of the technologies described above:

A platform architecture that supports multi-GPU virtualization, including a GPU resource management module, a virtual machine client module, and a GPU server module. The GPU resource management module is deployed on the GPU resource management node, and the virtual machine client module is deployed on the virtual machine client. On the end, the GPU server module is deployed on the GPU server. The GPU resource management module is responsible for the registration of the GPU server and the processing of GPU resource requests. The virtual machine client module and the GPU server module perform data transmission. The virtual machine client module Interact with the GPU server module. The virtual machine client module intercepts the call of the virtual machine to the GPU and redirects it to the GPU server module. The GPU server module accepts the GPU call information intercepted by the virtual machine client module and calls the GPU for execution. And return the execution result.

The working method of the above-mentioned platform architecture supporting multi-GPU virtualization includes the following steps:

Step 1, start the GPU resource management module, monitor the registration request of the GPU server module and the resource request of the virtual machine client module, and maintain a status table for the registered GPU server.

Step 2, start the GPU server module, and send a registration request to the GPU resource management module.

Step 3, the GPU resource management module receives the registration request, creates a table, maintains the status of the current GPU server, and returns success after completion.

Step 4, the GPU server module receives the message of successful registration, and immediately monitors the designated port.

Step 5, start the virtual machine client module, and monitor calls to the GPU. When a GPU call occurs in the virtual machine, a resource request is sent to the GPU resource management module.

Step 6, the GPU resource management module receives the resource request, obtains the working status of the registered GPU server, and assigns the most matching GPU server to the virtual machine client module according to a certain algorithm.

Step 7, the GPU resource management module receives the assigned GPU server, and establishes a data transmission connection with the GPU server module of the GPU server, and the virtual machine client module encapsulates the intercepted call and sends it to the GPU server through the data transmission connection module.

Step 8: The GPU server module receives the encapsulated data, integrates the load information of each GPU, selects the most matching GPU for it, executes the call, and returns the result until the execution ends.

In step 9, the virtual machine client module receives the result and returns it to the application program until the execution ends.

Beneficial effect

The present invention utilizes the existing virtualization technology, introduces the parallel processing capability of the GPU into the virtual machine, and combines the management mechanism, and uses the commonly used socket, infiniband or the communication mode dedicated to each virtualization platform as the carrier of data transmission, which does not affect the existing virtual machine. It only needs to add components on the basis of the original platform, so that the virtual machine can use GPU to calculate, which is applicable to all virtualization platforms. The present invention is easy to use, users only need simple settings and configurations, easy to transplant, and can work on various virtualization platforms without modification; the present invention is designed and positioned to use GPU to assist in improving the processing capacity of virtualization in virtualization clusters It is suitable for teachers and students who need to use GPU to demonstrate programming technology or other content in the teaching process. It is also suitable for virtualization clusters to improve the processing capacity of virtual machines and increase the utilization of GPUs. The present invention has a wide range of application scenarios, and has good practicability and feasibility.

Description of drawings

Fig. 1 is a schematic diagram of functional modules of the present invention;

Fig. 2 is the transmission program flowchart of the present invention;

Fig. 3 is a flow chart of the real-time control program of the present invention.

Detailed ways

The specific embodiment of the present invention is illustrated below in conjunction with accompanying drawing:

As shown in FIG. 1 , the platform architecture supporting multi-GPU virtualization provided by the present invention includes a GPU resource management module, a virtual machine client module, and a GPU server module. The GPU resource management module is deployed on the GPU resource management node, the virtual machine client module is deployed on the virtual machine client, and the GPU server module is deployed on the GPU server (that is, the GPU server). The virtual machine client communicates with the GPU server and the GPU resource management module through socket, infiniband or a dedicated method of the virtualization platform. The GPU resource management module is responsible for the registration of the GPU server and the processing of GPU resource requests, data transmission between the virtual machine client module and the GPU server module, interaction between the virtual machine client module and the GPU server module, and interception by the virtual machine client module The call of the virtual machine to the GPU is redirected to the GPU server module, and the GPU server module receives the GPU call information intercepted by the virtual machine client module, calls the GPU for execution, and returns the execution result. The GPU resource management module processes the requests of the GPU server module and the virtual machine client module

It responds to the resource registration of the GPU server module, monitors the load of each GPU server in real time, adjusts the tasks of each GPU server, manages the computing resources of each GPU server, and responds to the virtual machine client module according to the load. request, and assign the best matching GPU server to it. The virtual machine client module and the GPU server module interact through socket or infiniband or communication methods dedicated to each virtualization platform, and complete the GPU acceleration of the virtual machine through interception and redirection. In addition, the GPU server module manages the resources of multiple GPUs in the GPU server. Each GPU supports multiple tasks running in parallel. The GPU server module counts the current task load of each GPU and performs load balancing among GPUs according to the current load.

Figure 2 is a workflow diagram of the GPU resource management node. The virtual machine client module requests resources from the GPU resource management module, and the GPU server module registers resources with the GPU resource management module. The GPU resource management module parses the request and responds according to the current state.

FIG. 3 is a flowchart of the interaction between the virtual machine client module and the GPU server module. The computing work is mainly completed by the virtual machine client module and the GPU server module. After the virtual machine client module establishes a connection with the GPU server module, both parties send and receive data. At this time, the virtual machine client module intercepts the call of the application program running on the local machine to the GPU, and sends data to the GPU server module. The GPU server module receives the data from the virtual machine client, parses it, selects the GPU, uses the GPU to calculate the result, and returns the execution result. The virtual machine client module receives the data and returns the execution result to the application program. The GPU server module determines whether the data sent by the client is a stop transmission command. If not, repeat the above process; if yes, close the connection.

The working method of the platform architecture supporting multi-GPU virtualization specifically includes the following steps:

Step 1, start the GPU resource management module, monitor the registration request of the GPU server module and the resource request of the virtual machine client module, and maintain a status table for the registered GPU server.

Step 2, start the GPU server module, and send a registration request to the GPU resource management module.

Step 3, the GPU resource management module receives the registration request, creates a table, maintains the status of the current GPU server, and returns success after completion.

Step 4, the GPU server module receives the message of successful registration, and immediately monitors the designated port.

Step 5, start the virtual machine client module, and monitor calls to the GPU. When a GPU call occurs in the virtual machine, a resource request is sent to the GPU resource management module.

Step 6, the GPU resource management module receives the resource request, obtains the working status of the registered GPU server, and assigns the most matching GPU server to the virtual machine client module according to a certain algorithm.

Step 7, the GPU resource management module receives the assigned GPU server, and establishes a data transmission connection with the GPU server module of the GPU server, and the virtual machine client module encapsulates the intercepted call and sends it to the GPU server through the data transmission connection module.

Step 8: The GPU server module receives the encapsulated data, integrates the load information of each GPU, selects the most matching GPU for it, executes the call, and returns the result until the execution ends.

In step 9, the virtual machine client module receives the result and returns it to the application program until the execution ends.

Claims (3)

1. A working method of a platform architecture supporting multi-GPU virtualization, characterized in that said platform architecture includes a GPU resource management module, a virtual machine client module, a GPU server module, and the GPU resource management module is deployed on a GPU resource management node In the above, the virtual machine client module is deployed on the virtual machine client, the GPU server module is deployed on the GPU server, the GPU resource management module is responsible for the registration of the GPU server and the processing of GPU resource requests, the virtual machine client module and the GPU The server module performs data transmission, the virtual machine client module interacts with the GPU server module, the virtual machine client module intercepts the virtual machine’s call to the GPU, and redirects it to the GPU server module, and the GPU server module accepts the virtual machine client The terminal module intercepts the GPU call information, calls the GPU for execution, and returns the execution result; The GPU resource management module processes the requests of the GPU server module and the virtual machine client module, responds to the registration request of the GPU server module, and monitors the load of each GPU server in real time, and adjusts the tasks of each GPU server , manage the computing resources of each GPU server, and at the same time respond to the request of the virtual machine client module according to the load, and assign the most matching GPU server to it; Described working method comprises the following steps: Step 1, start the GPU resource management module, monitor the registration request of the GPU server module and the resource request of the virtual machine client module, and maintain a status table for the registered GPU server; Step 2, start the GPU server module, and send a registration request to the GPU resource management module; Step 3, the GPU resource management module receives the registration request, establishes a status table, maintains the current working status of the GPU server, and returns the registration success information after completion; Step 4, the GPU server module receives the message that the registration is successful, and immediately monitors the designated port; Step 5, start the virtual machine client module, monitor the call to the GPU, and when the virtual machine calls the GPU, send a resource request to the GPU resource management module; Step 6, the GPU resource management module receives the resource request, obtains the working status of the registered GPU server, and assigns the most matching GPU server to the virtual machine client module according to a certain algorithm; Step 7, the GPU resource management module receives the assigned GPU server, and establishes a data transmission connection with the GPU server module of the GPU server, and the virtual machine client module encapsulates the intercepted call and sends it to the GPU server through the data transmission connection module; Step 8, the GPU server module receives the encapsulated data, integrates the load information of each GPU, selects the most matching GPU for it, executes the call, and returns the result until the execution ends; In step 9, the virtual machine client module receives the result and returns it to the application program until the execution ends. 2. The working method according to claim 1, wherein the virtual machine client module interacts with the GPU server module through socket or infiniband or a communication mode dedicated to each virtualization platform, by intercepting and redirecting , complete the GPU acceleration of the virtual machine. 3. working method according to claim 1, it is characterized in that described GPU server module carries out resource management to a plurality of GPUs in GPU server, and each GPU supports a plurality of tasks to run in parallel, and GPU server module counts each GPU The current task load, and perform load balancing among GPUs according to the current task load.