Testing you'd like to see?

[NewDude]

My Situation: I've got 3 Xenserver hosts in a datacenter that's overkill for my needs, and I'm tired of paying for a full rack. So I'm going to host out of my house - gigabit fiber and generator backup are already here, so it just makes sense. It'll be a step down as far as uptime, but for the money I'm saving I'm fine with it. (I think.)

Anyway, as part of the move I'm doing away with Xenserver because, well, it's Citrix, and I don't particularly like the company and I really don't like the direction they're taking Xenserver. Proxmox is the platform I'll migrate to if I can make it work. So with that in mind I've got the following coming:

• At least* one Dell T30 server. Nothing fancy - E3-1225, 64G max ram.
• A Logitech SATA card that I can flash into IT mode.
• Some Toshiba 4TB drives
• Samsung SM863a drive for SLOG and maybe L2ARC use
• 64G RAM from Crucial.

My plan is to

• Install as RAIDZ-1 using the on-board SATA controller, install a VM, and measure drive performance.
• Move the drives from the motherboard SATA to the SATA card, and redo the test. I'd like to know if the third-party card outperforms the on-motherboard controller in any measurable way as some on the Internet claim.
• Add the SSD as a SLOG, and retest the same way.
• Tweak kernel.sched_migration_cost and kernel.sched_autogroup_enabled because why not? I probably won't see much change on a single VM, but it caught my eye today and if the results are noticeable I'll post something about it.
• I'll probably install Ceph, migrate the testing VM's storage to that, and test again.
• Once I migrate my servers and vacate my rack, I'll install the QNAP rack-mount server in my house, add NFS and iSCSI stores from it, and test those as well. How does Ceph perform versus file- and block-based shares from a separate device? Can you even notice the difference over a 1 gbit connection?

At the end of the day I'll hopefully have some comparative performance data on what should be an inexpensive server that's suitable for smallish loads.

So, my question: are there any particular numbers that would be most useful? Any other tests I should run while I'm at it?

* I say "at least" because the server I ordered last week never got confirmed, and Dell's site didn't have it listed. So I called Dell and got a nice Indian woman who told me there was no order by that number and that my card wasn't charged. So I ordered another one, and paid with Paypal. Paypal did get charged, but no order listing. With lots of cursing I ordered the same thing from an Amazon reseller. Now I've got something arriving via UPS tomorrow (the original arrival date) from a company that offers "the largest excess storage supply chain network across the globe." So, yeah. I hope I like the box because I probably just committed to a HA cluster of the things...

 

[totalimpact]

Sounds like fun, if you run windows VM, I might suggest Anvils storage benchmark, or on linux, fio.

You will want to run an initial quick test, and a 2nd test that exceeds RAM cache size (it may take a while to finish).

Those T30 towers are monsters with no guts, have you looked at 2nd hand R720, great bang for the buck, much slimmer, mainly the T30 is limited to 4 drives, ZFS does so-so on 4 drives, and actually mediocre on raidz 4 drives, I would suggest a zfs raid10 for some better io, the SLOG will work well, but in production you may notice some data will go straight to disk and not the slog.

 

[LnxBil]

Sounds like fun, if you run windows VM, I might suggest Anvils storage benchmark, or on linux, fio.

fio does also run on Windows :-D

 

[NewDude]

Thanks, ti. Once I have shut down the datacenter I've got a bunch of boxes I can add to the cluster instead - supermicros with dual E5 CPUs and 128-144GB RAM, though they all run hardware RAID. I can migrate that to the home office as well and maybe save some money, but if the T30's are enough for my limited needs I'm inclined to save on power and my ears as my servers aren't necessarily quiet...