Slashdot Mirror
Open Compute Hardware Adapted For Colo Centers
1sockchuck writes "Facebook has now adapted its Open Compute servers to work in leased data center space a step that could make the highly-efficient 'open hardware' designs accessible to a broader range of users. The Open Compute Project was launched last year to bring standards and repeatable designs to IT infrastructure, and has been gaining traction as more hardware vendors join the effort. Facebook's move to open its designs has been a welcome departure from the historic secrecy surrounding data center design and operations. But energy-saving customizations that work in Facebook's data centers present challenges in multi-tenant facilities. To make it work, Facebook hacked a rack and gave up some energy savings by using standard 208V power."
The World is Distributed. People are Distributed. The web is Distributed. Centralized Computing / Centralized Storage is irrelevant. Resistance is futile, you will be distributated.
Clouds, virtual systems, clusters, stand-alone servers all benefit from being in an environmentally friendly facility where there's lots of networking capacity and sufficient power and cooling. While home users have dedicated desktop or laptop computers, it's far more power efficient to use technologies like blade systems to package computing power. Regardless, everything's still in a data center where the equipment can be protected.
I used to work at a very large ISP where there were a half dozen data centers, each containing racks and racks of servers, storage and backup. The data centers I visit now still resemble the old ones, but they're more power efficient and the equipment has much higher densities and the networks much higher capacity.
Unless someone can make a computer or cloud that doesn't require much in the way of power, cooling or physical security, then the data centers will probably continue their current trend for the foreseeable future.
Don't anthropomorphize computers, they don't like it.
The equipment was originally designed to run at 277V (1 leg of a 3-phase 480V system), but is instead running at 208V (3-phase system where each leg is 120V). So while 208V may be higher than most US equipment, it's still lower than what they typically use.
dom
Data centers likely won't be going anywhere anytime soon. Businesses [1] tend to like keeping their critical stuff in a secured spot.
What I see happening in a data center are a few changes:
1: Data center rack widths will increase. This allows more stuff to be packed in per rack unit.
2: There will be a standard for liquid cooling where CPUs, RAM, GPUs, and other components that normally use heat sinks will use water jackets. Instead of a HVAC system, just the chilled water supply and a heat exchanger would do the trick. Of course, the issue is someone making valves and fittings that are leak resistant, are quick connects (disconnect hose, it shuts off the water flow), and can handle a number of connection and disconnection cycles before giving up and leaking. There would be leak sensors to shut off automatically any damaged cooling part and the machines attached to it similar to how a CPU shuts down if its heat sink gets bumped off.
3: A move to DC power because it means that every rack unit just needs to step up and down the incoming voltage. No power supply needed. Of course, there are dangers with DC power (muscle lock), but telcos already use 48VDC. Of course, switching DC power is a PITA due to no zero crossings, thus having to deal with arcs and pitted contacts. However, there are always rack-level PDUs which can take the 208 VAC power and turn it into 12-48 VDC, with low amounts of voltage loss due to the relative short distances.
4: A move to a passive backplane type of architecture. This way, specialized CPU boards can be added as needed, as well as "external RAM" [2]. It will allow the latest/greatest network and disk protocols to be changed out as need be.
5: More high end SAN features like real time block level deduplication making it into the onboard motherboard RAID chips.
6: Hypervisors built into all motherboards where a utility like Xen or vSphere will be more of an admin shell.
7: More security appliances, which are specialized in tasks. For example, an appliance that just stores username and password hashes so when a Web server authenticates a user, it uses that. Too many wrong guesses of a user's password would result in blocks/delays on the appliance level (something even a compromised Web server could not get around.) This would be used to ensure that an intruder couldn't make off with the /etc/shadow equivilent.
8: A resurgance of tape. Disk media was cheap and improved exponentially for a while. Now, tape is starting to catch up, and offers a lot more surface area, so ariel densities are not as critical compared to reliability and storage. No matter how one slices it, tape is not going anywhere soon because nothing beats it for reliability and price. D2D2T will still remain the norm provided there is no new media revolution (like a new optical format.)
9: More technologies for deduplication. IBM has tape deduplication as well as an appliance which sits between machines and the SAN fabric and deduplicates data on the fly.
10: A push for more technologies that can be remotely run via a Web page or a SSH connection. This allows for unmanned data centers to not just be possible, but easily done.
[1]: Those businesses that didn't trade their heroes for ghosts and move to the cloud, that is. However, cloud providers use data centers.
[2]: This may be DRAM, or it may be some other mass media technology. We have been hearing about holographic storage for decades now. It would be a tier of storage with a speed level between disk and normal RAM that would be used as swap or cache.
Low voltage DC is piss-poor for distribution because power losses in wires increases at the SQUARE of the current. 120V@1A will have far lower losses than 12V@10A - 100 times less.
The big AC to DC places use high-voltage DC for that reason - lower current cables are far easier to handle than high current cables (the thickness of a conductor depends on its current - ampacity. The insulator does have to get thicker for higher voltages, but it's a lot more flexible than a thick 00-gauge wire.
DC-DC converters are fairly efficient and converting down to where you need has less losses than trying to shove 100A of 12VDC to a rack (assuming said rack only consumes 1200W. I think a modern rack can easily draw 3600/4800W fully loaded with servers which would mean up to 400A at 12V to the rack - calling for seriously thick cabling).
Oh, and what happens when you have high currents flowing at low voltages? You get welding. Because IIR heating is far more effective when you're passing huge currents through.
Colo means co-location, in which customers rent rack space, and they move their own hardware into the data center.