Ready to switch to production - need advice

[russellkt]

I manage IT at a very small company and have primarily been using a napp-it AIO setup for backing up a managed Dell R620 VMWare 6.5 setup for the last 3 years. Everything has worked great and I'd like to now switch the current setup to a production napp-it AIO on vmware 6.5 u2

Current Hardware:
Dell R620
2 x Intel e5-2620
64GB Ram
H710
4 x 600gb Hardware Raid


Dell R720
2 x Intel e5-2650 processors
128 gb RAM
H710 Flashed to LSI IT
16 x 1.2 TB SAS 2.5 10k
Intel DC P3700 800 GB
Intel 700P 800 GB
Mellanox 10gbe SFP
Intel x540 10gbe

Supermicro x9dri
2 x Intel e6-2660
128 gb RAM
LSI SAS Controller Flashed to IT
8 x 3tb 3.5 SAS 7200rpm
Mellanox 10gbe SFP

Current VMs
Windows Server 2012 r2 as Domain Controller
Windows Server 2012 as Office File Server
- 3tb vmdk
- 81,370 pdfs avg file size 37mb min 500K max 250mb
2 x Win 10 Pro for Remote Desktop
Napp-it
Linux webserver for legacy rails app
Win 2003 for a java webapp
All of the vms are relatively static and a nightly replica will be sufficient

Waiting for delivery - Optane 900p 280gb

My initial thoughts for the next version are:

Retire R620 and use memory in R720
Make R720 Production Server
- napp-it Pro license
- 2 x 8 Raid Z2
- Intel DC P3700 as zil
- Use Intel 700P for VMs
- current vms would remain the same except replace the Win 2012 File Server with a SMB share running on the zfs pool

Use Supermicro as Backup Server
- 1 x 8 Raid Z2 napp-it Replication or Pro License

I've got everything setup as mentioned but have been unable to find the best way to use the Intel 700P to host the vms and still get the benefits of napp-it: replication, snapshots, and nfs sharing. Options I am considering:
- run napp-it vm on 30gb vmdk, create a 650gb data vmdk
- get another drive to host napp-it vm and passthrough 700P to napp-it
- don't fool with 700p and just host vms on the zfs spinning pool

Any and all thoughts are greatly appreciated. Thank you to all on this forum for their work and information as I would still be stuck in a suboptimal vendor lock in without it.

 

[gea]

I've got everything setup as mentioned but have been unable to find the best way to use the Intel 700P to host the vms and still get the benefits of napp-it: replication, snapshots, and nfs sharing. Options I am considering:
- run napp-it vm on 30gb vmdk, create a 650gb data vmdk
- get another drive to host napp-it vm and passthrough 700P to napp-it
- don't fool with 700p and just host vms on the zfs spinning pool

The 700P, is this onboard soldered?
I have not looked at the specs but it should have powerloss protection if you intend to place critical VMs on it. It is ok hold ESXi and some uncritical VMs like the napp-it storage VM, a second storage VM for tests or backup with the option to switch the pool between and some other uncritical local VMs (without ZFS, ZFS snaps and replication).

I would just start with the disk pool and the 3700 as slog. If it is not fast enough for VMs, use the disk pool for filer and backup usage without Slog and with sync disabled. Then build a high performance pool from the 3700 (has powerloss protection) with sync enabled. Optionally look for a second P3700 to bild a mirror.

 

[russellkt]

Thanks for the reply @gea. I setup as you described and it appears that the disk pool will be fine for the vm usage. When the optane arrives, I plan to replace the 3700 with it as the slog. In this case is it best to pass through the optane as described in the manual or better to just create a 20G vmdk and use it as the SLOG.

If not passing through is the vmware SCSI Controller ok to use or should I use the vmware NVME Controller?

 

[gea]

With pass-trough, OmniOS can access the Optane directly while with a vmdk the traffic goes over the ESXi driver. The Optane is so fast that the difference should be quite small. If you have the choice and it works without problems, pass-through is recommended while vmdk give more options (esxi bootdisk, slog, l2arc etc). A Slog size of 20GB is big enough.

The ESXi NVMe vdisk driver is newer and may give a lower latency. I have not done performance benchmark between (use the default scsi usually)