Abstract
When deploying applications in Kubernetes, storage is traditionally attached and consumed from a remote source such as a SAN or NAS, and that remote source is to be accessible from the entire cluster. While this is a good model for rapid deployment of resources, it is less desirable for high-performance applications that would work better with locally accessible low-latency storage. The open source Linux SDS “LINSTOR” provides a framework for integration with Kubernetes that allows containers to prefer or require automatically provisioned local storage while still maintaining the capability of accessing that same storage remotely.
This is a very powerful solution for enabling distributed storage with varied locality for latency-sensitive applications which require persistence, such as machine learning workloads or databases.
In this talk, I will be demonstrating the configuration and environmental requirements for accomplishing such a solution, the impact that this methodology has on workload performance, and analyzing traditional as well as more exotic applications of this new capability in realistic conditions.