Dude where's my vm?
summary
G‑Research describes using OpenStack (Train -> Xena), Kolla Ansible, Ironic and Kyobi to pass whole GPUs and raw NVMe into VMs via PCI passthrough, plus nvme, GPU and vGPU workloads scheduled with Placement using placement traits and PCI aliases. The team builds flavors with passthrough properties, automates trait assignment with Ansible/Kyobi, stacks instances to reduce host count for a monthly rebuild cadence, and reports working GPU/NVMe passthrough but ongoing issues around device cleaning and stable device mapping.
AI-written summary · about
chapters (9)
- 0:00Introduction to G Research
- 1:10OpenStack infrastructure
- 3:09Cloud operational challenges
- 5:43GPU and NVMe hardware support
- 8:38PCI passthrough configuration
- 11:41Scheduling and weighing
- 13:28Using traits for automation
- 17:35Review of architectural goals
- 18:48Future work and Q&A
related talks
01h 50mKeynote
17mBuilding the Rackspace Open Cloud with XenServer and OpenStack
29mCeph, Xen, and CloudStack: Semper Melior
30mOpenStack Works, so now what?
33mDelivering continuous deployment of xen-api at cloud scale
18mAchieve a single Openstack deployment in multiple datacenter
25mTraining Through Failures: How Meta Keeps 100k-GPU Jobs Alive With Open-Sou...
44mLINE's Journey; Road to 4 Million Cores in t...