Just over a year ago, I built a new home file-server since I was quickly outgrowing my existing storage capacity. Rather than just dropping more HDD's into the existing hardware I had running, I succumbed to geek techno-lust and opted to explore new technologies: ZFS and VMWare ESXi.
Virtualization is huge in the enterprise-space these days. After talking with some friends from work about what kind of virtualization strategies we're using for our in-house datacenter, I became enticed by the idea of running a "bare metal" hypervisor: install virtualization software as the primary OS on the physical hardware and spinning up VM's for different logical needs.
Over the years, I've accumulated a fair amount of old computers. Back in the day, those extra computers gave me an opportunity to try-out new OSes/software: playing with different Ubuntu configurations, tinkering with different *BSD flavors, etc. But once my spare-time dried up, all that old computer hardware just became clutter taking-up space and it seemed like such a pain to fire up an old computer just to tinker with some new configuration.
So, going virtualized held a lot of (obvious) appeal to me: minimize my physical hardware (less power consumed, less physical space, etc.) while letting me easily tinker with new configurations.
Based on previous research, I knew that I really wanted to move to some kind of ZFS-backed storage solution, to protect/maintain the integrity of my ever-increasing digital collection.
After following the "zfs-discuss" mailing list for a few months, given my simple home-needs a simple mirrored configuration (along with proper backups) seemed like the best solution. If I need more space, I can expand horizontally: add another pair of drives (of whatever size is appropriate) to expand my pool. It was this easy expansion which pushed me towards a mirrored configuration rather than a RAID-Z style configuration. The trouble with RAID-Z is that you can't add new devices to a vdev after you've initially created it; the only way to "grow" a RAID-Z vdev is by replacing all of the individual drives (one at a time, waiting for each to resilver) and setting the "autoexpand" property on the pool so that ZFS will auto-expand the pool based on the new common maximum size of each of the individual drives. But, that's a whole lot of moving parts (pun intended) and replacing all the drives in the pool isn't quite my idea of easy expansion. Mirroring just seems easier, given that I don't need lots of individual drives for raw performance reasons.
Somewhere along the line in my research, I stumbled upon napp-it, which introduced me to this idea of an "all-in-one" fileserver: using VMWare ESXi as a bare-metal hypervisor, having a Solaris-based VM running on the ESXi datastore which will control the mass-storage ZFS pool, and then exporting an NFS share back out to ESXi so that you can store the bulk of the VMs on ZFS-backed storage. It's a bit complicated at first glance, but it performs great (thanks to hardware-passthrough) and lets me easily manage the bulk of my VMs on ZFS-backed storage.
To make use of ESXi's hardware-passthrough support, you need to run server-grade hardware. For me, that meant going with an Xeon-based CPU with appropriate motherboard chipset. But that also gave other server-grade wins like using ECC memory and IPMI (easy remote-control of the console, which is awesome).
Here is the hardware I ended-up going with:
- Intel Xeon X3440 Lynfield 2.53GHz
- Supermicro X8SIL-F-O
- 2x Kingston 4GB DDR3 ECC Unbuffered
- Intel SASUC8I (LSI 1068E) SATA/SAS card
- SeaSonic X650 Gold PSU
- Fractal Design Define R3 Mid-Tower Case
- 2x 2TB Hitachi Deskstar 5K3000 (primary pool)
- 2x 2TB Hitachi Deskstar 5K3000 (backup pool)
- 2x 2TB Hitachi Deskstar 5K3000 (off-site backup pool)
The Xeon chipset was needed for some of the hardware passthrough features in ESXi, e.g. passing through the HBA card directly to the Solaris-based VM so that ZFS can have direct access to the physical drives.
The 650W power-supply ended-up being way more than I needed. The server uses less than 100W when idle, though that PSU is still quite efficient even with such low power draw.
The setup has been awesome so far. Even just getting to play around with server-grade hardware has been an eye-opening experience. Being able to remote-control the server, e.g. mounting an ISO remotely over the Java-based client and installing the OS on the computer all without needing to hook-up a keyboard or monitor, has been a revelation. (Goodbye old CRT monitor that I used to keep around for my server closet!)
I love the flexibility of ESXi. I have an Ubuntu VM for development/testing, a FreeBSD VM for running Mediatomb (PS3 media-server), etc. It even let me play around with different Solaris flavors while trying to figure out what OS I wanted to use for the ZFS back-end. It's just so easy to spin-up new VMs to test an idea or just to play around.
I started out using Solaris 11 Edition (free), which ran fine for the better part of a year. I tried upgrading to Solaris 11/11 earlier this year only to find that Solaris 11 apparently doesn't play nicely with ESXi. From there, I jumped over to using OpenIndiana, the open-source spin-off from the now-defunct OpenSolaris lineage.
ZFS has been a huge win too. Taking nightly snapshots makes it dead-easy to look back in time to see how day was several months ago. The snapshots also make it easy to send the incremental differences to a backup pool. The built-in CIFS server makes it dead-easy to mount the shares on Windows, and the filesystem snapshots are easily accessible via the "Previous Versions" tab in Windows Explorer. I also really love the idea of the "pool". I can create different filesystems to group/organize my data (e.g. photos vs. music vs. backups), enable compression on a per-filesystem basis, set quotas per filesystem, etc.
I really love the "all-in-one" idea: ZFS reliability combined with the flexibility of VMWare ESXi.