Experiences of supporting HPC/HTC workloads on private cloud resources, with ideas for how to do this better and description of trends for non-traditional HPC resource provision.
I will discuss my work as part of the Operations Team for the eMedLab private cloud, https://www.emedlab.ac.uk, which is a large-scale (6000-core, 5PB) biomedical research cloud using HPC hardware, aiming to support HPC workloads. This includes an account of the technical challenges of marrying traditional HPC components such as parallel filesystems with the rather different expectations of the OpenStack cloud management framework. Other challenges faced include educating users about the basics of virtual environments, providing access to restricted-license large-scale (800TB) datasets in a multi-tenant environment, and achieving good performance inside virtual HPC clusters.
Looking to the future, container orchestration engines such as Kubernetes provide a very promising approach, as the containerisation of scientific applications gains momentum, and I will describe existing and future work in this area.
Speakers: Adam Huffman