ITEC -  Operating Systems Group

Interconnect Adapter State Migration for Virtual HPC Clusters

  • Type:Diploma Thesis
  • Date:18.01.2013
  • Supervisor:

    Prof. Dr. Frank Bellosa, Marius Hillenbrand, Dr. Jan Stoess

  • Graduand:Simon Sturm
  • Links:PDF
  • Abstract:

    Live migration of virtual machines is considered as one of the most powerful tools available in the context of machine virtualization. Specifically in data centers, it can enable higher fault tolerance through fail-over semantics and better load-balancing within a cluster.

    High performance computing infrastructure as a service (HPC IaaS) has gained substantial momentum in the recent past, as it promises to bring elasticity and flexibility to the area of high performance computing (HPC). In contrast to traditional Ethernet-based cloud data centers, however, HPC clusters typically use high-speed interconnects with intelligent network adapters, such as InfiniBand. They provide cutting-edge communication performance and use aggressive protocol offloading, OS-bypass techniques, and advanced features such as remote DMA. While the latest generation of such cards already provides hardware capabilities to ease their virtualization (with SR-IOV, for instance), support for transparent live migration of these self-virtualizing adapters is still an open question in virtualization research.

    In this thesis, we propose a novel design enabling guest-transparent live migration of a virtual machine configured with a directly assigned virtual device instance of self-virtualized hardware. We take advantage of the self-virtualization features to perform the actual state migration of the assigned instance in software. Despite we focus on HPC IaaS, our approach can be also employed in other scenarios. A prototypical evaluation, based on Mellanox’s ConnectX-3 VPI InfiniBand HCA and the Linux KVM virtual machine monitor (VMM), proofs that this design integrates with recent VMMs.

    BibTex:

    @mastersthesis{sturm13interconnectadapter,
     author = {Simon Sturm},
     title = {Interconnect Adapter State Migration for Virtual HPC Clusters},
     type = {Diploma Thesis},
     school = {System Architecture Group, Karlsruhe Institute of Technology (KIT), Germany},
     month = jan # "18",
     year = 2013,
     note = {\url{http://os.ibds.kit.edu/}}
     }