Automated Instantiation of Heterogeneous Fast Flow CPU/GPU Parallel Pattern Applications in Clouds

摘要

Parallel scientific workloads typically entail highly-customised software environments, involving complex data structures, specialised systems software, and even distinct hardware, where virtualisation is not necessarily supported by third-party providers. Considering the expansion of cloud computing in different domains and the development of different proprietary (e.g. Amazon Web Services, Azure) and open source cloud platforms (Eucalyptus, OpenStack, OpenNebula), users should arguably be able to automatically and seamlessly migrate their parallel workloads across cloud platforms using standardised virtual machines. However, even if it is easier to migrate the workload between nodes when the nodes have a similar configuration on the same platform, the transition between different platforms typically raises different issues such as vendor lock-in, portability, and interoperability. In this paper, we describe our work to automatically deploy a complex parallel software stack on heterogeneous hybrid cloud platforms. We have elastically deployed FastFlow -- a C/C++ pattern-based programming framework for multi-/many-core and distributed platforms -- using virtual machines on both CPU and GPU-based architectures between heterogeneous virtualised platforms. Our approach relies on the standard Open Virtualization Format (OVF) in order to achieve a universal description of virtual appliances. Such a description is not only useful for elastically migrating and deploying, but also to determine the hardware/system software configuration needed switching to any new (cloud) image format. We have successfully evaluated our work using virtual machines based on VirtualBox and Amazon Web Services on local cluster and public cloud providers.