Across industries, clusters or networks of PCs or servers are being used to perform intensive tasks that were earlier in the domain of supercomputers. This opens up a niche opportunity for large as well as small players from the computer industry and for entrepreneurs who want to offer high performance computer services
Think supercomputer and you probably imagine a high-tech insulated setup outside India, inhabited by the nerdy professor or the military man. This is not true anymore, with supercomputers today being used by other industries than universities and the military. And more and more are coming up outside the US. Top500.org has been maintaining a list of the world’s 500 fastest supercomputers and India ranks ninth in the number of installations in this list, with the world’s fourth fastest being a system installed at the Computational Research Labs of the Tata group at Pune.
In fact, there is no strict definition of what a supercomputer is, and increasingly, we use the term high-performance computing for this class of computers. Another interesting point is that they no longer require specialized hardware or software. These number crunchers can be put together as a high speed network of hundreds of common PCs or servers. The trick is in making the application software to split the job at hand across each of these individual machines and then collect back and collate the results. Creating such software is also not a very difficult art. The process is fairly standardized.That brings us to the question: Who uses these machines besides the military and academia? A whole host of industries are today installing and using high performance computing systems to crunch through large volumes of data. Construction companies, animators, medical researchers, the financial sector, Internet businesses, media companies, retail players, the telecom industry… the list virtually spans the whole spectrum.
And what do they use them for? Let’s see some examples.
Any movie today has tons of post-production effects as well as animation. Each minute effect has to be created and finished (rendered is the industry term) frame by frame. Needless to say, these have to be of very high quality. The way this is done is to submit the job, frame by frame, to a render farm or a render wall—a group of servers dedicated to the job. Each frame could take hours to render. For example, for the blockbuster animation movie, Cars, Pixar Studios used up 2300 CPU hours on their render farm. Over the Hedge used up over 15 million rendering hours spread across thousands of servers. But you need not start off all that big. A small-sized render farm for animation work can be created using fifteen to twenty or even five machines.
Another opportunity is in the render services. These are render farms that allow users to use specific services, loading their projects over the Internet. These render farms charge on the basis of software used, number of frames and time consumed. A number of international render farms are available and recently, DuxSoft has started a similar service in Mumbai (see box).
|DuxSoft SPARX rendering cluster|
|SPARX from DuxSoft (dusoft.net) is a high-performance computing rendering solution for the animation industry. It currently has two deployments; on a pay-as-you-use model at renderarm.com and as an in-house solution at Accel Animation Studios. |
SPARX is a cluster environment, currently with 25 server class machines doing the rendering work and four server class machines doing cluster and job management. There is a storage pool for the large storage requirements of rendering. Additional workstations are connected to this network for development tasks.
|How much did it take to develop and deploy SPARX? |
According to Indivar Nair, Architect and VP, Technical Services at DuxSoft, the software took around five months to create, using a team of about seven people. Five months and seven people? Yes. Nair points out that all of them came with many years of expertise in their chosen domains and that unlike boxed software for mass use, such high-end applications are typically created in far lesser time. What was the big challenge? Integrating all the different software together and making it installable on multiple platforms and hardware.
Renderarm currently charges Rs 18 per gigahertz of rendering power used by a customer. A typical movie frame takes anywhere between 30 minutes to four hours to render, depending on the complexity. An average frame takes about two hours. Studios like Anibrains, Esco Toonz, Suyambu and p2rnet are customers of renderarm.com.
How much does it cost to set up a SPARX cluster? Development costs, consultancy charges, manpower costs and hardware costs and software included, the figure is under a crore, with hardware accounting for the bulk of it at fifty lacs.
|For the technically inclined|
Networking and storage are defined based on user requirements. Gigabit networking is normal, but if very high bandwidths are required, then InfiniBand can be used. Similarly, you could use SAN (Storage Area Network) storage connected to each node if you have very high storage requirements.
Deploying the Render Farm Management module on 25–50 node farm takes around three hours.
written by willieOutsourcing, November 17, 2010
written by Nick Pragmaticus, September 28, 2010
written by Anurag Jain, March 26, 2009
|< Prev||Next >|