_ __ _ __ __ _ _ __ ___ _ __ ___ ___ _ __ ___ | '_ \| '__/ _` | '_ ` _ \| '__/ __/ _ \| '_ ` _ \ | |_) | | | (_| | | | | | | | | (_| (_) | | | | | | | .__/|_| \__, |_| |_| |_|_|(_)___\___/|_| |_| |_| |_| |___/ We don't assume you are stupid.
Update: Well, I'm back. for real this time. The deal is now slightly different, it's $206.25 for a quarter-rack with a 20 amp 120v circuit, so you have 80%, or 4 amps to play with. The pdu is monitored, but not remotely-accessable. You get a shelf. Photos to follow.
In most co-location setups here is what happens: the owner has a full rack with a metered PDU. Usually the PDU doesn't meter by outlet, just by circuit, so when someone wants to put in a new server, they call up their provider. "Hey provider" they ask, "can I put up a server of spec X?" the provider makes some guesses about how much power it will draw, and says "Sure, give me Y dollars a month for it" -
then the guy shows up with a more powerful server, you detect it with the metering PDU, back and fourth. It's a huge pain in the ass, and half the time the customer ends up feeling like they got pushed into buying more than they wanted. I can't count the number of people who tried to get dual xeons in my $50/month 80w 1u co-location deal, because, after all, the server /was/ 1u. I'm not trying to say the customer is being dishonest or trying to pull a fast one; the amount of rack space used is certainly the most visually identifiable standard for deciding how "big" a server is, right? I mean, I did say that 1u was $50 a month, right? to the non-technical, this certainly is more concrete than how much power it may draw.
Even hooking up a kill-a-watt to the server and seeing how much it draws is more complicated than it looks; servers use a whole lot less power at idle than under load, and even under load, they use more power when they are hot than when they are cold. I remember once, a long time ago, I filled a circuit with dual netburst xeons. I started memtest86 on the bunch, and later, after everything looked fine, I left for the night. I came back the next morning? the breaker had tripped. Nothing else on the circuit; the amount of power the servers running memtest were using apparently rose while I slept, either due to the servers becoming warm, or due to some variation in how much CPU memtest86 was using.
This can easily happen to the unwary shared-circuit co-location provider, because your customers usually want to pay for the amount of power their system uses when the turn it on, plug it in to the meter, and run some cpu burn program, which might be less than what the system uses once it has warmed up, or when the CPU is exercised in a manner that differs radically from the CPU burn program. If you call them back and tell them they are using more, they usually want to re-negotiate. This, I think, is why per-customer metering, warning, and if worst comes to worst, shut-off is essential to ensure that if a customer goes over their power budget, they don't take out the whole rack.
So, here's what I'm thinking I can do to cut down on the negotiation (and thus labor cost and discomfort for both customer and provider) involved:
what if I simply split each rack into quarters, and built some technical means such that if your equipment drew more than 1/4 the circuit's capacity, it shut you down automatically? (Obviously, I'd want to build in something that warned you when you reached 75% or something)
The most obvious means would be a metering, switched power supply with a front end like powerman, so that users can see their power usage and hard reboot servers when required. (also, so my script can shut users down when they are over their power quota and in danger of popping the circuit.) There are two routes I can go here: first, I can buy a very expensive power supply that is metered per outlet. Outlets 1-3 are for customer 1, 4-6 customer 2, 7-9 are customer 3, etc... the obvious problem with this is that it requires a more careful customer to make sure they only use plugs 4-6 than to say "this 8 port PDU is for you. Don't touch the others" and it appears that used APC 79xx 1u 8 port PDUs are less than 1/4th the cost of per-outlet metering PDUs, so at least for this test run, and for 1/4th racks, I'll be giving everyone their own 8 port APC. It's kinda silly, just 'cause in this power budget, you will be talking one powerful server, or maybe up to four really small servers. Without resorting to SheevaPlugs or other ARM chips, there is no way you are going to get 8 servers in the thing.
It also serves my higher-end customers. I always say that if you need more than 8-16GiB ram, you really are better off buying your own servers. I think $150/month is a reasonable price for that much power in this area, and it's still within range of the small customers that might be interested in doing business with me. (obviously, if you have real money, you aren't going to go with me, so renting full racks and other more expensive products is right out.)
So, after what, a year and a half? I've gotten this going. Two customers are live now.
Note, there are some downsides: 3 other people will have physical access to your servers, though all users of a rack will share contact info with all other users of a rack, and I /will/ throw someone out if they touch someone elses server. This is true of most, but not all co-location where you have physical access and rent less than a full rack. I'm only giving you one Ethernet port, and The port speed I'm selling is 10Mbps and the next easily available port speed up is 100Mbps, both full pipe commits. I'm doing this because I want to add as few points of complexity and failure as possible. At some future point I may sell a 95th percentile billing bandwidth product, but I'm not going to do that just yet.
Anyhow, signups are now open for 1/4 rack co-location at SVTIX in san jose - $150 setup, $206.25/month, includes 10Mbps and a few serial ports. I've got two quarter racks ready to go, and an empty rack right next to mine waiting for me to wire up the patch panels, so this is something you can sign up for now and reasonably expect to have soon.