Anchor: Diskless Cluster Provisioning Using Container Tools
SessionSoftware Provisioning
Event Type
State of the Practice Talk
Best Practices
Containers
System Management
TP
TimeTuesday, 17 November 20202pm - 2:30pm EST
LocationTrack 6
DescriptionLarge scale compute clusters are often managed without local disks to ease configuration management across several hundred nodes. This diskless management frequently relies on a collection of in-house scripts designed to build client compute images. By leveraging container building tools matured by the tech community we can reduce internal technical debt while allowing cluster installations to be more flexible and resilient. Deploying container images to compute clusters, however, remains an unsolved problem. To this end we present Anchor, an extensible initrd module designed to boot clusters from an immutable squashFS image with a read-write overlay. The code referenced in this paper is available at https://github.com/olcf/anchor.
