Virtual CPU and memory management methods for data access performance of virtual machines on a single NUMA system비균등 메모리 접근 구조에서의 가상머신 메모리 데이터 접근 성능 보장을 위한 가상 CPU 및 메모리 관리 기법에 관한 연구
Recently, a manycore system with NUMA (Non-Uniform Memory Access) architecture is being used widely. Especially, due to cost efficiency and resource management capability, service providers consider virtualization on the NUMA system as their underlying target system.However, a data access on the NUMA system may experience different latency depending on node locations of the core issued it and the target data. The virtualization needs some optimizations considering this characteristic to accomplish its goals such as performance isolation, resource utilization, and so on on the NUMA system. Therefore, through this dissertation, we will present a research topic to enhance the virtualization on the NUMA system: "Credit-based Runtime VM Placement on a NUMA system for QoS of data access performance".
$\bullet$ Credit-based Runtime VM Placement on a NUMA system for QoS of data access performance: Each virtual machine (VM) running on the NUMA system may have different data access performance depending on the shared resource contentions and remote access conditions. The data access performance also varies with time due to changeable data access pattern of other VMs and core scheduling and memory placement decisions of a hypervisor. On existing hypervisors, such as Xen, VMware, and KVM, users of VMs cannot recognize or predict their data access performance that they have received or will receive.
In the first part of this dissertation, we attempt to resolve these issues pertaining to
irregular data access performance of VMs running on the NUMA system by relocating virtual CPU and memory of VMs dynamically. A hypervisor with our scheme can provide the illusion of a private memory subsystem to each VM, which guarantees the latency defined at initial time of the VM for incoming data access requests on average. For this purpose, using hardware PMUs, the hypervisor evaluates the average data access latency of each VM. We define a per-VM metric of data access performance, ca...