I’ve been using virtualization technology for several years now – first with VMWare (Workstation, Server, and Player), and then with VirtualBox. I tried Xen very briefly while looking at OpenSUSE, but never did use it on a day-to-day basis. Until now, I had only ever used desktop OSes (operating systems) (mostly Linux, but some Windows 7 Pro and Mac OS X also) as the virtual machine host – the OS that was running on the actual physical hardware.
Then, video podcast Hak5 released Episode 519 – Building the Ultimate White Box [EXSi] Server for Under $2000. ESXi is an installable VMWare hypervisor that can be downloaded and used for free. The catch with ESXi is that there is a somewhat limited list of supported hardware, and some of it – specifically the RAID controller – is pretty pricey. There have been many successful efforts to get ESXi installed and running on hardware that is not officially supported, but it can be a bit tricky to get the right combination. Another hardware component that seems to be problematic is a supported network adapter.
The idea behind building this type of server is to take advantage of the fact that most server hardware is underutilized – particularly the CPU and network bandwidth. With adequate memory and fast, reliable disk access, it is possible to run a few to several virtual servers on a single physical server. After watching the episode, I started thinking about building an ESXi server, but my budget was only about $1000. I priced all the components I’d need, and the cost was going to be close to the $2000 that the guys at Hak5 had spent. Because I really could not afford to spend that much, I started wondering if I could do something similar using beefy workstation class components running Debian as the host OS. As you might guess from the fact that this is the first in a series of blog posts about it, I was successful! (Well, okay, I didn’t quite make it under $1000; also, in the interest of full disclosure, I used a full tower case that I already had, so that wasn’t included in the cost, but overall it came pretty close.)
Server-class -vs- Workstation-class Trade-offs
By beefy workstation class components, I mean that instead of a server-class motherboard that uses ECC RAM and has dual power connections I opted for a well-made ATX motherboard. Instead of using Intel’s server-class Xeon CPU, I opted for a Core i7-930 CPU. Instead of redundant power supplies, the case has only a single 750 watt unit. Instead of hardware RAID-5 using hot-swappable SCSI hard drives, I’m using software RAID-1 on hot-swappable SATA II drives. Finally, instead of an ultra-fast server-class Gigabit network interface card (NIC), I opted to use the Gigabit NIC on the motherboard with the option of adding a PCIe Gigabit NIC if needed.
So what did I give up by making these trade-offs (other than saving about $800 or so)? I gave up redundant power – but none of my clients that have “real” servers with redundant power supplies have a UPS for each; they all have both power supplies plugged into a single UPS, so the only thing I’ve given up is some possible uptime if the power supply needs to be replaced. I gave up Error Correcting RAM, but that is mitigated somewhat by the fact that today’s RAM is very reliable, and for the most part follows the “if it works for two weeks it’ll probably work for 10 years” rule-of-thumb for electronics. With the CPU, RAID, and NIC, I’m giving up some performance, but I doubt that I’ll notice. And in addition to the performance hit, with the software RAID-1 I’m taking a chance that both of the disks might fail at the same time, but I think that’s unlikely.
I’ll look for ways to mitigate the reliability issues and future articles in this series will detail any I find. The next article in this series will detail the components I chose and provide an overview of the hardware build with photos. Check back soon!
