Cephalocon 2022

We have a customer that required to know what are Ceph's limitations top to bottom. In the benchmark process, we've tested Ceph at all levels, starting from the underlying hardware, to the S3 object storage service exposed by Ceph RGW. The main cause of this benchmark was to create the connection between all layers: Hardware (RAM, Disk, CPU, Network) --> OS (sysctl tunning) --> Ceph RADOS --> Ceph RGW --> User experience. The artifact from this benchmark should be the SLA and SLO our customer can expose to his own customers within the organization. This customer has pushed Ceph's architecture to the edge, while allowing his customers to consume Amazon-like storage classes (on for fast processing and low latency response time, and the other for archival, quite similar to Amazon S3 storage classes) while dividing the cluster to performance domains to prevent noisy neighbors (using Realms, Zones, and different root buckets). Having this architecture pushed the benchmark complexity even further, and we'd like to share this with the Ceph community to ease their understanding on how to manage Ceph at scale.