This session explains lessons learned integrating HPC workloads with unikernels on top of OpenStack.
HPC workloads are traditionally executed in general purpose operating systems running directly on the host. Virtualized and containerised environments using these base systems are being adopted to improve flexibility and manageability of the infrastructure. To this end, the MIKELANGELO project is using a lightweight unikernel (OSv) providing bare minimum, but fully virtualized environment to these workloads. We are going to present a flexible application management that has been implemented specifically to support integration with cloud middlewares, such as OpenStack. This allows dynamic composition and execution of workloads based on user's needs. OpenFOAM CFD is being used to showcase our approach, including an integration into OpenStack dashboard. Finally, we will show how the use of a flexible telemetry system delivers in-depth insights into the infrastructure and application.
This session is going to explain best practices, pros and cons of using unikernel from the perspective of the HPC. It will explain the unikernel concept and approach, explain the architecture chosen in this case study and present its major limitations. It will use a real world application based on OpenFOAM CFD simulation to introduce the steps necessary to ensure full compatibility with Linux-based guests.
We are also going to delve into the highly flexible management layer of the OSv itself, as well as the newly developed application packaging mechanisms enhancing flexibility of running arbitrary workloads independent of the actual environment: local, private/public cloud or HPC. Pilot integration with OpenStack dashboard will be used as a real-world showcase.
The session will conclude by outlining several other benefits that the MIKELANGELO project focuses on and are tightly coupled with our choice of the unikernel-based HPC: improved paravirtual I/O and virtual RDMA.