Improving of shared storage disk speed e.g. with LVM caching

[tobic]

Hello,

we are using the following setup as a Proxmox cluster for running ~6 Windows 10 VMs used heavily as VDI for software development and additionally some server VMs and containers for our infrastructure.

Servers:
2x Dell PowerEdge R730 (each: 2x Xeon-E5-2670v3 (24 Cores), 256GB RAM)

Storage:
SAN Fujitsu DX90 S2 (4GB Cache, Dual Controller, RAID10 with 14x 600GB 10K 2,5" HDDs)

The storage is connected via multipath FC 8Gb, providing one big LUN as a shared LVM to the two Proxmox hosts. Sequential read operations will be easily able to reach theoretical wire speed 15-1600MB/s via the two multipathed FC channels. But the problem is, with e.g. a lot of small file disk operations in one VM the storage becomes sometimes slow and we are getting a poor reaction time in other VMs as well.

Is there any possibility with a shared LVM storage to add SSD cache drives (e.g. with dm-cache or lvmcache)? Either by adding them to the SAN as an additional LUN accessible to both hosts in parallel or adding NVMe SSDs in both hosts.

Alternatively, are there any suggestions how the storage could be rearranged for high performance but still HA without e.g. upgrading to a costly new hybrid SAN?


Thanks and regards,
tobi

 

[czechsys]

Second ssd Lun and migrate VMs there. 14x600GB = 4,2TB in R10, thats 4x2TB ssd. Performance + power consumption benefit.

 

[tobic]

Second ssd Lun and migrate VMs there. 14x600GB = 4,2TB in R10, thats 4x2TB ssd. Performance + power consumption benefit.

Thanks for your reply. Adding SSDs to the SAN is actually limited due to needed SAS dual channel drives and a vendor lock in the system. I already tried attaching non-OEM hard drives without success.
The only way I found is, to buy 400GB SSDs which are compatible with the SAN: CA07339-E703 - These are available (used!) for about 379€. Which brought me to the idea of using them in some way as a shared cache LUN (like a ZFS SLOG device - but with dm-cache, lvmcache, ...).
Another idea would be migrating to some kind of a hyperconverged storage, but e.g. Ceph should at least consist of a three nodes cluster and provides only 1/3 of the physical storage...