Faster, Leaner, Smarter, Better Data Centers

Friday, March 07, 2014 · Posted by at 1:30 AM

Four years ago, Facebook broke ground on its first greenfield data center project in Prineville, Oregon. In the years since, we’ve deployed six iterations of that design, culminating in the first building currently under construction at our new campus in Altoona, Iowa. With facilities around the world, we constantly challenge ourselves to improve our data center designs to maximize efficiency, reduce material use, and speed up build times.

At this year’s Open Compute Summit, we previewed what we believe will be a step change in those ongoing efficiency efforts: a new “rapid deployment data center” (RDDC) concept that takes modular and lean construction principles and applies them at the scale of a Facebook data center.

We expect this new approach to data center design will enable us to construct and deploy new capacity twice as fast as our previous approach. We also believe it will prove to be much more site-agnostic and will greatly reduce the amount of material used in the construction. And with today’s exciting news from my colleague Joel Kjellgren, we will get to test these theses: Our newly announced second building at our Luleå, Sweden, campus will be the first Facebook data center to be built to our RDDC design.

In true Facebook style, the RDDC concept began with a hack. In October 2012, our data center strategic engineering and development team and several experts in lean construction came together to hack on a design for a data center that would look less like a construction project and more like a manufactured product. From this hack, a couple of core ideas for streamlining and simplifying the build process emerged.

The “chassis” approach

The first idea developed during the hack was employing pre-assembled steel frames 12 feet in width and 40 feet in length. This is similar in concept to basing the assembly of a car on a chassis: build the frame, and then attach components to it on an assembly line. In this model, cable trays, power busways, containment panels, and even lighting are pre-installed in a factory.

photo1

The “chassis” would be assembled in a factory and transported to the site

These chassis support all the services that are found overhead above the racks. Unlike containerized solutions, which are a full volumetric approach that includes a floor, this idea focuses solely on the framework that exists above the racks, to avoid shipping the empty space that will eventually be occupied by the racks. When the chassis arrives on site, it’s set atop posts mounted to the slab. Two of these chassis attached end to end create the typical 60-foot-long cold aisle, with 10 feet of aisle space at each end. For context, a typical data hall is composed of 52 total chassis, attached in a 4 x 13 grid configuration, with 13 cold aisles.

Chassis being placed on post

Chassis being placed on post

Sectional view of two 40’ chassis

Sectional view of two 40’ chassis

Plan view of 40’ modules being placed in a grid pattern

Plan view of 40’ modules being placed in a grid pattern

We concentrated our modular production efforts on the framework chassis over the racks – which define the cold aisle – to avoid having to ship modules that also include the hot-aisle space. Instead, the width of the hot aisle is established by the pitch at which chassis are set. Using a pitch of 15 feet, we’re able to decrease the hot aisle width to 3 feet – translating to a reduction of approximately 28 feet in overall data hall length over our previous stick-build designs. In conjunction, to further reduce the amount of structural steel needed in the stick-build design, this floor-supported approach removes the supply air penthouse. In its place, manufactured air handling units are installed at grade flanking the data halls.

The “flat pack” approach

The second concept developed during the hack was inspired by the flat-pack assemblies made famous by Ikea. Our previous data center designs have called for a high capacity roof structure that carries the weight of all our distribution and our cooling penthouse; this type of construction requires a lot of work on lifts and assembly on site. Instead, as Ikea has done by packing all the components of a bookcase efficiently into one flat box, we sought to develop a concept where the walls of a data center would be panelized and could fit into standard modules that would be easily transportable to a site.

In this scheme, we employ standard building products like metal studs and preassembled unitized containment panels that are then erected onsite and are fully self-supporting. These panels are limited to a width of 8 feet to maximize the amount of material that can be shipped on a flatbed trailer without requiring special permits for wide loads.

A mock-up of a flat-pack test assembly

A mock-up of a flat-pack test assembly

The wall panels – which are 14 feet tall – have been simplified using off-the-shelf components and easily mate with each other. Careful attention was paid to minimizing the number of unique components. For example, 364 identical wall panels are used in each data hall. The ceiling panels use Epicore metal deck product, which spans the 12 feet width of the cold aisle and racks. This serves the additional duty of carrying the loads of the trays, power bus, and light fixtures below it using a proprietary hanger clip for the threaded rods.

This flat pack concept is still early in its development, but the current evaluation has already identified great time and material savings. Expected results Both of these key RDDC concepts (the chassis and the flat pack) should allow us to make a number of measurable gains, including:

- Site-agnostic design: By deploying pre-manufactured assemblies, a majority of the components can be used interchangeably. The goal is to be deployable wherever we seek to build next. It’s our hope that by standardizing the designs of our component assemblies much like we do with OCP servers we can deploy a unitized data center into almost any region in the world faster, leaner, and more cost-effectively. Performing more of the assembly in a controlled environment and at ground level also reduces assembly time.

- Reduced on-site impact: The RDDC concept will deploy pre-engineered unitized modules that minimize the amount of time required for heavy equipment on site and overall time to complete a data hall. The modules reduce the generation of on-site waste and the impacts associated with the delivery and staging of individual construction materials common to traditional construction techniques.

- Improved execution and workmanship: Having a predictable and repeatable product delivered to the site allows local teams to easily replicate the quality and fit from one region to another. Our RDDC design will produce this result by using explicit assembly instructions with established tolerances.

We expect to begin construction on our second data center in Luleå soon, using these RDDC designs. We will continue to share our learnings about RDDC design and construction so the OCP community can contribute their ideas and help advance data center design and construction that much more quickly.

A rendering of Facebook’s Luleå 2 Rapid Deployment Data Center (RDDC)

A rendering of Facebook’s Luleå 2 Rapid Deployment Data Center (RDDC)

Marco Magarelli is a design engineer on Facebook’s data center design team. For more background, watch his presentation from the OCP Summit in January (begins at 21:50 mark).