Recommended number of OSDs in CEPH cluster

[nitaish]

I want to setup a Proxmox cluster with Ceph storage for HCI based infrastructure. To start with I will be creating a 3 nodes cluster. I have decided to go with the following hardware configuration

2 x Intel Xeon Decacore processors
128 GB DDR4 RAM
120 GB Enterprise SSD for Proxmox
2 x 1.92 TB Enterprise SSD for the OSDs
1 x 8 TB Enterprise SATA for backups

I want to know if 2 OSDs per server is enough or I should go for more OSDs like say 4 or 6. I will be using this cluster for Production and will be adding more nodes as and when the requirement increases. Kindly suggest.

 

[Alwin]

I have decided to go with the following hardware configuration

On what bases did you decide?

I want to know if 2 OSDs per server is enough or I should go for more OSDs like say 4 or 6. I will be using this cluster for Production and will be adding more nodes as and when the requirement increases.

Smaller disks and more disks, means less time during recovery of a disk and more disks to distribute work. But depending on workload, two SSDs could be enough (especially with NVMe).

See our docs for the precondition section.
https://pve.proxmox.com/pve-docs/chapter-pveceph.html#_precondition

 

[nitaish]

On what bases did you decide?

Based on the current and the estimated workload in the next quarter.

Smaller disks and more disks, means less time during recovery of a disk and more disks to distribute work. But depending on workload, two SSDs could be enough (especially with NVMe).

Do you think going with four 1 TB Enterprise SSDs will be better than going with two 2 TB Enterprise SSDs?

See our docs for the precondition section.
https://pve.proxmox.com/pve-docs/chapter-pveceph.html#_precondition

 

[Alwin]

Based on the current and the estimated workload in the next quarter.

Yes, I meant more, what are the numbers.

Do you think going with four 1 TB Enterprise SSDs will be better than going with two 2 TB Enterprise SSDs?

Depends on the model of SSDs, even if they are called Enterprise, they might just not do the job.