Cephalocon 2020

People would usually think of Ceph being disaggregated - after all it's not a single storage box where you need to rely on proprietary magic to keep the thing going. But what if we could disaggregate it even further? What if when a node dies we could just quickly restart the OSD on an other host?

Large scale cloud providers usually have dedicated storage boxes due to various space, performance and density constraints on their compute farm - which is usually much denser and far larger in numbers than their storage cluster.
Some organically grown storage clusters might have started their life as 4U, HDD-dense boxes but now people are looking at how to add NVMe.

The common in these two? There is a lot of extra compute and network capacity on those compute nodes. What if...

What if we could use some of the latest storage and network technology to move these NVMes in the storage boxes magically around?

Would it improve performance?
Can we do it natively BlueStore and SPDK or must we rely on the kernel?

And the most important question: does it all worth it?

Come and find out!