Practical Technology

for practical people.

The fastest computers are Linux computers


What do the BlueGene/L System, the BlueGene/P system and the New Mexico Computing Applications Center SGI system, based on the Altix ICE Integrated Compute Environment 8200 model, have in common?

First, they’re the top three fastest computers in the world, according to the latest Top 500 supercomputer list. Second, they all run Linux.

It’s not just the crème de la crème of supercomputers that run Linux, however. 85.2 percent of the Top 500 supercomputers run Linux. In other words, 426 of the world’s fastest 500 supercomputers depend on Linux.

Linux-powered supercomputers don’t require any particularly special version of Linux. New Mexico’s SGI Altix ICE, for example, run pretty much standard Novell SLES (SUSE Linux Enterprise Server) and Red Hat’s RHEL (Red Hat Enterprise Linux).

One of the reasons why Linux owns HPC (high-performance computing) is that today’s supercomputers tend to be made up of clusters of multi-cored processors. Typically, these are dual-core CPUs. The one notable exception is that Intel’s quad cored Clovertown chips now power 102 of the Top 500. The vast majority of the fastest of the fast supercomputers, 354 systems, or 70.8 percent, now use Intel processors. Intel is followed by the AMD Opteron family, 78 system,s or 15.6 percent. After AMD, IBM takes third place with 61 systems, or 12.2 percent, that use IBM Power processors.

What does all this have to do with Linux? For starters, Linux runs natively on all these processors. The days when supercomputers from Cray and Control Data Corp. depended upon customized chips for their speed are long past. Today, speed comes from clustering standard chips into better and faster processor arrays and more effective parallel processing programming. For example, a single SGI Altix ICE 8200 rack can hold up to 512 common-as-dirt Intel Xeon processor cores. Today’s supercomputer processors are then tied together into a three-dimensional matrix by high-speed intraconnect fibre.

Linux has excelled at getting the maximum of performance from the minimum of commodity equipment for more than a decade now. In 1994, Thomas Sterling and Don Becker, at Goddard Space Flight Center’s CESDIS (Center of Excellence in Space Data and Information Sciences) created the first Beowulf Linux-powered clustered supercomputer. Even with that first system, which was made up of 16 486-DX4 processors connected by channel bonded Ethernet, it was clear you could deliver supercomputing performance with COTS (Commodity off the Shelf) based systems.

Since then, the story of supercomputing is really the story of Starling and Becker’s initial ideas, implemented with Linux and ever-faster COTS processors and networking, taking computers to ever-faster speeds. Thirteen years later, the BlueGene/L has achieved a Linpack benchmark performance of 478.2 Tflop/s (teraflops, or trillions of calculations per second).

That record won’t stay up for long, though. In just the last six months, to be a top 100 supercomputer, almost all of them Linux-powered, the entry point for the top 100 in March 2007 was 9.29 Tflop/s. Now, the slowest of the top 100 runs at 12.97 Tflop/s. With Linux, the sky, and perhaps Moore’s law, are the only limits for how fast computers can go.

A version of this story first appeared in Linux-Watch.