The hardware used to produce graphics in computers may be the key to managing the massive amounts of processing power needed by future supercomputing efforts, such as the Square Kilometre Array (SKA).
A project like the SKA needs massive amounts of processing power, with terabytes of data flowing into the central computing core every second.
The difficulty with processing radio astronomy data from multiple telescopes is that a massive number of calculations are involved, taking massive amounts of time using conventional methods.
By splitting the data into chunks to be processed simultaneously, the processing time can be drastically reduced. This method is called parallel processing, as opposed to a single processor doing all the work in conventional serial processing.
Chris Harris, PhD student at the Western Australian Supercomputer Program (WASP), at UWA, is completing his doctoral thesis on this topic.
“My thesis investigated the use of GPUs in radio astronomy data processing,” he says
“If you’re trying to get image of the sky using an interferometer, you’re using multiple telescopes to get a better image.
“What that means is that you need to take those separate signals and combine them to form an image, and there’s a great deal of computation involved in that.
“My research shows that when I use a GPU with a parallel correlation algorithm, it’s ten to a hundred times faster.”
A Graphics Processing Unit (GPU), the graphics card in a computer, is tailor-made for this kind of work, since the number crunching involved in displaying graphics has the same requirements.
These small, powerful cards are already catching up to traditional, room-sized supercomputers.
Professor Karen Haines, Director of the WASP at UWA, says the ability to quickly process data makes GPUs invaluable for projects like the SKA.
“The problem is that scientists go to the big computers to do their number-crunching but then they use a different computer to look at their results.
“In cases like the Square Kilometre Array, you’re talking stream computing. They don’t have time to store the data on disk, they need it to stream straight through.
“And not only that, but the scientists will want to see the intermediate steps during processing.
“GPUs are letting us figure out not only how to process the data in real-time,but how to let us look at the data while it is processing rather than at just the end. And you can do it all in real-time.
“The computing revolution’s on.”
For ‘stream computing’ architecture, massive networking bandwidth and powerful parallel processing is required.
Two years ago, the WASP purchased a Cray XT supercomputer, taking up three filing-cabinet sized containers, with a capacity of 0.85 TeraFLOPs (0.85 trillion calculations per second), and costing 2 million dollars.
Now, a 2 thousand dollar graphics card has two thirds the processing power of that supercomputer.
To test the capability of these cards, the WASP team has a dedicated GPU lab, with multiple processors connected by a high-speed network. They have also installed a GPU into a Cray XT system, dubbed the FrankenXT, and will be refining this design based on their prototype’s success.
A story provided by ScienceNetwork WA - Activate your connections to science. This article is under copyright; permission must be sought from ScienceNetwork WA to reproduce it. To comment on this article go to the original story here.