OS-independent live migration scheme for bare-metal clouds

• Published in: Proceedings of the 8th International Conference on Utility and Cloud Computing pp. 80 - 89
• Main Authors: Fukai, Takaaki, Omote, Yushi, Shinagawa, Takahiro, Kato, Kazuhiko
• Format: Conference Proceeding
• Language: English
• Published: Piscataway, NJ, USA IEEE Press 07.12.2015
• Series: ACM Conferences
• Subjects: virtualization; live migration; bare-metal clouds
• ISBN: 0769556973; 9780769556970
• DOI: 10.5555/3233397.3233411

Bare-metal clouds are an emerging and attractive platform for cloud users who demand extreme computer performance. Bare-metal clouds lease physical machines rather than virtual machines, eliminating a virtualization overhead and providing maximum computer hardware performance. Therefore, bare-metal clouds are suitable for applications that require intensive, consistent, and predictable performance, such as big-data and high-performance computing applications. Unfortunately, existing bare-metal clouds do not support live migration because they lack virtualization layers. Live migration is an essential feature for bare-metal cloud vendors to perform proactive maintenance and fault tolerance that can avoid long user application downtime when underlying physical hardware is about to fail. Existing live migration approaches require either a virtualization overhead or OS-dependence and are therefore unsuitable for bare-metal clouds. This paper introduces an OS-independent live migration scheme for bare-metal clouds. We utilize a very thin hypervisor layer that does not virtualize hardware and directly exposes physical hardware to a guest OS. During live migration, the hypervisor carefully monitors and controls access to physical devices to capture, transfer, and restore the device states while the guest OS is still controlling the devices. After live migration, the hypervisor does almost nothing to eliminate the virtualization overhead and provide bare-metal performance for the guest OS. Experimental results confirmed that network performance of our system was comparable with that of bare-metal machines.