Practical study of bare metal virtualization platforms

• Author(s): Pousa, Duarte ; Rufino, José
• Date: 2017
• Persistent ID: http://hdl.handle.net/10198/22523
• Origin: Biblioteca Digital do IPB
• Subject(s): Bare metal virtualization; Benchmarking; Performance assessment
• Document Type: Conference object
• Language: English
• Contributor(s): Biblioteca Digital do IPB

Description

With the hardware breakthroughs accomplished through the years, the idea of “software defined hardware” has become a reality. Hypervisors such as KVM, Xen, Hyper-V and ESXi enable the cloud of today, with hardware consolidation bringing a reduction in operating costs. In this scope, it is imperative to address the performance of all the different virtualization implementations, in order to discover any potential bottlenecks or bugs. In this work, the performance of all the prominent Type-1 virtualization platforms was analyzed, using guests representative of the Windows NT and Linux kernels, in the form of Windows 10 LTSB and Ubuntu Server 16.04 LTS. The effectiveness of the CPU scheduler of each hypervisor is put to the test, as well as the storage backend performance under multiple scenarios (iSCSI, NFS and local). To achieve this end goal, the tests were all performed under the same conditions. As such, an effort was made in terms of automation. Additionally, the benchmarks were performed using platforms representative of each hypervisor: KVM is represented by oVirt and Proxmox, Xen by XenServer and ESXi by VMware Vsphere. These platforms cover most of the bare metal virtualization market. The outcome of this study revealed that all the current implementations provide near native performance, at least when there is no resource overcommitment. That being said, there are some outliers, namely Hyper-V, which seems to have significant CPU/memory access overhead, while displaying a clear lead in terms of disk I/O performance.