Building inexpensive server clusters

When you have a serious computing problem to solve but limited funds, here are some options, courtesy of heavy hitters like MIT and Sandia National Laboratories.

When you need serious in-house computational power, there are two main options.

One is to build a supercomputer, a vast conglomerate of high-speed servers, extremely fast storage arrays and lightning-quick data connections -- all of which are very expensive. Another, lower-cost approach: Build a cluster out of low-power and inexpensive computers.

That's how four heavy hitters addressed their need for top-of-the-line computation. Their clusters provide near-real-time processing power to scan for signs of the early universe, develop next-generation radars or simply run network tests as fast as economically possible.

1. GPU cluster for astronomy research

Here's a unique challenge: Say you want to set up a high-performance computing cluster in the Australian outback because it's easier to scan for signs of the early universe without any radio interference in the night sky there.

As you might expect, there are only a few options. For the Murchison Widefield Array, researchers knew that power was going to be a major issue. The array is about 50 kilometers (around 32 miles) from the nearest settlement and about 300km from the nearest town. A cluster that consisted of standard high-performance computing nodes would use too much power, given the lack of infrastructure on-site.

Instead, the Murchison Widefield Array consists of about 80 individual GPUs (graphics processing units) -- the Nvidia Tesla S1070 -- in two clusters. The whole thing runs on diesel generators that provide about 40 kilowatts of power.

There's a whopping 2.5 teraflops of astronomical data that runs through the array at a top speed of 3GB/sec. The image processing, which has to be done on-site because there are no fiber channels in the Australian outback, is fairly intense. Antennas capture the radio data (hence the need for a radio-quiet site) and feed the data into a device called a correlator, which then provides the input to the GPUs.

The image processing involves converting the mathematical equations from Fourier data into real space, correcting ionospheric distortion and calibrating the instruments.

"Contemporary CPUs such as Xeon and Opteron are not going to provide the computing muscle required within that power budget," says Richard Edgar, a researcher who helped set up the cluster. "We considered using low-power MIPS processors instead, but they could not supply the required performance either. The GPU route was the only one which offered a reasonable chance of success, given the extremely high compute requirements and low electrical power."

The cluster is not completed, although a test prototype is up and running. Edgar says the cluster is the fastest in the world in terms of teraflops per watt.

1 2 3 Page 1
Page 1 of 3
  
Shop Tech Products at Amazon