Consideration: large scalable fs

[dennisTGC]

Hi,

Wonderfull project, this kind of projects with curousity the primairy focus are in my opinion the most worthfull research projects out there.

Having said that, i have some questions/considerations:

• you didn’t disclose much about your testsetup. I think the problem you are trying to solve is related to large multiuser filesystems where you actually have the problem of the infra layering you described in your presentation. But in your presentation it seems (i’m assuming) you tested on a single system without the raid controllers, lots of systems/disks. So I would suggest testing against large systems. On small single systems the fs is already optimised i recon and obviously you would not run gelera on a single system anyway. I’m guessing the performance hit will be less at larger workloads, multiple user/systems where scale can benefit from the less layered infrastructure.
• Although i admire the project, I think your goals have been superceded by Ceph, which has CephFS, which is a scalable multiuser system which can (or choose not to) know about the hardware underneath. The current implementation of Ceph does away with having a fs on the disks and has dedicated metadata servers. Ceph has at least a horizontal scalable posix fs and can be optimised for the hardware it uses. It’s (i think) a better way to test against as your idea, as it tries to solve the same issue at the same scale.

If you get great results, im sure the Ceph team is interested and they have much more manpower to make something of the project.

Some disclaimer here: Im no way an expert, so i can easilly be proven wrong, so comments welcome!