Ceph 18.2.2 - Performance Windows 2022

[kellogs]

Hello,

Today i tested a windows 2022 vm performance on Proxmox 8.2.2 Ceph 18.2.2 using CrystalDiskInfo

https://crystalmark.info/en/software/crystaldiskinfo/

Total 7 Ceph nodes.
Spec per node.
1 x 1.92TB PM863a SATA 6.0 Gbps 1.3 DWPD
9 OSD x 8TB Seagate ST8000VN0022 IronWolf NAS SATA 6Gb/s which each uses 96 GiB of the 1.92TB PM863a as WAL/DB device

Each compute and Ceph nodes have 4x10G with 2x10G for public and 2x10g bonding layer3+4 - LACP to a pair of Huawei stacked cloudengine switches for cluster network

MTU size 1500

This is the result of the crystalmark on the Windows 2022 while 2 OSD were active+remapped+backfilling

 

[kellogs]

Would love to anyone ceph setup benchmark speed

 

[Falk R.]

Hello,

Today i tested a windows 2022 vm performance on Proxmox 8.2.2 Ceph 18.2.2 using CrystalDiskInfo

https://crystalmark.info/en/software/crystaldiskinfo/

Total 7 Ceph nodes.
Spec per node.
1 x 1.92TB PM863a SATA 6.0 Gbps 1.3 DWPD
9 OSD x 8TB Seagate ST8000VN0022 IronWolf NAS SATA 6Gb/s which each uses 96 GiB of the 1.92TB PM863a as WAL/DB device

Each compute and Ceph nodes have 4x10G with 2x10G for public and 2x10g bonding layer3+4 - LACP to a pair of Huawei stacked cloudengine switches for cluster network

MTU size 1500

This is the result of the crystalmark on the Windows 2022 while 2 OSD were active+remapped+backfilling

1. ChrystalDiskMark is not the best Tool to Benchmark Server Systems.
2. A Simple SATA SSD as Bluestore is definitely your Bottleneck. I have Ceph Setups (Backup Server) with SAS SSDs and 1 SSD is Bluestore for 6 HDDs. The SSD has often the same Load as the HDDs and with Full Load, the SSD is the Bottleneck. Please use Fast SSDs and Best Practice is 4 HHD per SSD and Bluestore = 4% of Disk Capacity.
3. Ceph with HDDs is good for Backups or Fileservices, but no good idea for VMs.
4. The Next Bottleneck are your Network, but with this slow Disk is it OK.

The Chrystal Benchmark Results are not Bad for a Single VM. Ceph is a Scaleout System and Perform Better with multiple Workers (VMs)