Research On Resource Management Optimization Strategy For NUMA Architecture In Virtualized Environment

The continuous development of cloud computing and virtualization technology has made it a key technology to support the efficient operation of various applications.Cloud computing can not only provide flexible resource supply and secure spatial isolation for applications,but also improve data center resource utilization and reduce resource maintenance costs.However,with the ever-increasing scale and complexity of resource access for cloud applications,current virtualization technologies are still not able to meet the performance requirements of applications,especially the lack of an efficient virtual machine(VM)resource management mechanism.At the same time,due to the transformation of the data center server architecture to non-uniform memory access architecture(NUMA),the performance overhead of resource access has become more and more complicated,especially the remote access latency and shared resource contention have a great impact on the performance of VM.Moreover,the transparent resource management mechanism of virtualization technology brings new challenges to the optimization of resource access to VM.This study analyzes the performance overhead of various current cloud applications on NUMA platforms by combining the development trend of server architectures with the shortcomings of existing VM resource management methods.Meanwhile,with quantitative experiment and qualitative analysis,three optimization strategies for VM resource management based on NUMA architecture are proposed in the application layer,operating system layer and virtualization layer.Firstly,due to the complex and variable resource requirements and resource access behaviors in the application layer,current VM resource management methods cannot accurately optimize the performance bottlenecks of different types of applications on the NUMA platform.Therefore,this study proposes a dynamic adaptive virtual resource management policy(v DARM).Based on the resource requirements of different applications,a bandwidth performance model of the VM is established,and an adaptive performance bottleneck positioning method is used to determine the performance bottleneck of the application running inside VM.For the current NUMA performance bottlenecks,we also designed a dynamic resource management strategy to eliminate the corresponding performance bottlenecks,so as to achieve performance optimization for different applications in the VM.Secondly,for the inaccurate underlying node interconnection topology description mechanism in the operating system,the current VM resource scheduling mechanism cannot achieve efficient resource scheduling and performance optimization on heterogeneous NUMA architectures.This study proposes a topology-aware virtual resource management mechanism(v TRMS)for heterogeneous NUMA multi-core servers.First,the performance impacts of symmetric and asymmetric NUMA interconnection topology is studied,we conclude that the system information locality table(SLIT)provided by the operating system is inaccurate and the scheduling mechanism is inefficient.To this end,this study uses the resource access latency between each node as the distance metric,and designs a latency detection method.Based on the detected distance metric and the runtime VM resource access behavior,the VM’s virtual CPU scheduler accurately schedule VM’s v CPU to the optimal NUMA node to avoid the performance degradation on NUMA servers.Finally,in view of the fact that the current virtualization layer resource management methods cannot efficiently optimize the performance of complex multi-resource access applications,this study proposes a load-aware global resource affinity management system(LG-RAM).We first analyze various resource access paths of VMs on the NUMA host by running typical applications.Experimental results show the performance impacts of resource affinity and shared resource load balancing on VM performance in memory and I/O subsystems.Based on the experimental conclusions,LG-RAM uses a load-aware global resource affinity optimization method to optimize the affinity of memory and I/O resource access through v CPU and memory page rescheduling.At the same time,the load rebalancing mechanism is used to balance the load during resource access,which improves both VM Performance and host resource utilization.