Sydney's 20-tonne cloud keeps Harry Potter flying

What weighs 20 tonnes and gives Harry Potter the ability to fly? A very heavy animation and special-effects rendering cloud.

System integrator Frontline has built a high-performance computing infrastructure-as-a-service (IaaS) called Steam Engine to provide compute on demand for the processor-intensive film and visual effects industry in Australia and overseas.

The platform was recently used as a satellite facility to render Harry Potter and the Deathly Hallows.

Steam Engine used HP's C7000 C-Class servers which, fully populated, weigh almost 250 kilos per chassis.

"You put four of those in a rack and you're talking about one metric tonne per square metre," said Stefan Gillard, equity partner and commercial director for Steam Engine.

"Even with a raised floor there aren't very many floorplates within facilities that can accommodate dumping 20 of those on the deck in one place," Gillard said.

Gillard said he considered moving Steam Engine's servers into an APC hot-aisle containment system (known by its acronym, HACS) which prevents hot air recirculating into the server racks and can improve density.

The densities from HP's C Class were "far and away beyond what we could get from any other vendor at the moment", Gillard added.

Each rack of dual quad-core or hex-core servers requires 35kW - far above the 8KW conventional supply in most data centres. Steam Engine has 1000 nodes deployed in Sydney's Global Switch data centre with several terabytes of high-speed storage connected by 10Gbit switching.

The company planned to add 1500 nodes by December and another 1500 in early 2011.

Steam Engine has eight customers and expanded beyond entertainment to financial services and science. The greatest traction was from SMEs for whom high-performance computing was beyond the realms of their financial capacity, but "now they can play with the big boys", Gillard said.

"Customers are using anything from 100 nodes (at 12 cores per node) right down to six nodes for simulation analysis work on the Large Hadron Collider data," Gillard said.

Gillard claimed Steam Engine can deliver similar amounts of computing capacity at a lower cost than Amazon. However, Steam Engine was giving customers physical nodes and not just virtualised instances of nodes.

The heavy input-output (IO) requirements in high-performance computing require physical infrastructure which can provide better performance than virtual infrastructure, he said.

Outsourcing animation

Gillard has extensive experience in building render farms for the movie industry; his last big project was building a new facility to process animation and special effects for Happy Feet and Mad Max. Gillard kept thinking of the financial burden that building a high-performance computing environment placed on a film studio and what an alternative might look like.

"I was looking at moving to a model where our computing IT requirements become more like a utility company where you could bolt on additional capacity during peak demand rather than have it as a sunk cost.

"The rule of thumb for motion picture production on an animated feature, your negative cost is about $100 million. $30 million goes to cap. ex., $30 million to op. ex., $30 million in people and 10 percent tends to be margin," Gillard said.

Movie-making facilities were discussing collaborative approaches to pool resources and spread usage between films "to save them from having to capitalise and build out infrastructure in their own data centre, with their own network, with their own dime", Gillard said.

Gillard teamed up with Frontline to build Steam Engine and used his connections to find customers in the entertainment industry.

Steam Engine charged a single by-the-node price that included power, cooling, data transmission and the physical kit itself. Discounts kicked in based on the number of nodes and the length of the rental period.

"We can provide a real savings to an operation over even a short-term, mid or long-term of about 30 percent of what the market can provide that for. When you don't need it you just turn it off. There's no penalty for saying I'll have 100 nodes this month and none next month," Gillard said.

The initial plan has changed from providing bare-metal computing requirements to a managed service. Customers want queue management, asset management and a managed platform as a service.

Steam Engine has been approached by studios in Vancouver and India to build an HPC cluster which they could rent.