A Research On Resourse Virtualization Technique In Heterogenerous Computing

With the rapid growth of the data volume,the demand for computing power in the data center is also increasing.The computing power demand of various applications such as encryption and decryption computing,deep learning online reasoning,cloud VR,cloud game business,etc is far beyond the capabilities of general-purpose processors.At this time,the use of heterogeneous computing technique to increase the computing power of the system has become a trend in the industry.However,there are differences in various types of heterogeneous computing units in heterogeneous computing systems.Users directly using multiple computing resources need to be exposed to the details of the mutual computing system,which ultimately leads to problems such as a high degree of customization of the computing system and difficulty in development and deployment.This thesis studies the virtualization technology of various heterogeneous computing nodes.The main work of the thesis includes the following aspects:First,the construction of the heterogeneous computing platform is completed.The specific work is to complete the construction of the basic operating environment of the heterogeneous computing nodes in the heterogeneous computing platform,including the system migration of the X86 board,the underlying firmware development of the FPGA board and the transplantation of embedded Linux system.Second,this thesis analyzes the basic concepts of virtualization technology.A detailed analysis of the principles of various virtualization schemes on general-purpose processor-type computing nodes is done.By comparison,this article uses Docker as the virtualization technique for processor nodes in this platform,and designs the implementation of Docker technique in this platform.Third,an interconnection scheme of virtual computing resources based on RDMA technique is proposed.The framework creates a virtual channel based on QP SEDN/RECV operation between task modules,realizing a low-latency,highthroughput data exchange scheme.At the same time,we designed the proxy module of RDMA for Docker container and tested its related performance.Fourth,the FPGA partial dynamic reconfigurable technique is analyzed,and the related process of FPGA dynamic and reconfigurable is completed.At the same time,based on this technique,the FPGA resource virtualization solution design has been completed,including the FPGA virtualization logic framework design,the FPGA virtualization framework driver layer design,and the controller programming design.Fifth,completed the design and implementation of the system resource management framework,and designed a diagram-based task description scheme,which provides users with a unified task deployment and monitoring interface while shielding the underlying details of the system hardware,thereby simplifying user development and the deployment process.This paper completes the abstract virtualization of various heterogeneous computing resources,and implements a heterogeneous computing resource management framework,which simplifies the application development and application deployment process of each heterogeneous computing node,and at the same time improves the resource utilization of the system,which has a certain engineering significance.