Shark Bait

Knowledge Centers

Operating Systems

Networking & Internet

Mobile & Wireless

Security

Storage

Business Intelligence

Servers & Data Center

Computerworld Reports

Industry

See your link here

Steven J. Vaughan-Nichols

Cyber Cynic

Linux powers the fastest computers on the planet

228 comments

TAGS:IBM, Linux, Novel, Red Hat, supercomputer, SUSE
IT TOPICS:Hardware, Linux, Mainframes & Supercomputers, Open Source, Operating Systems, Software

Once upon a time, supercomputers used special vector model processors to achieve their remarkable speeds. Then, at the dawn of the 21st century, people began working out how to achieve record-breaking computer speed by linking hundreds or thousands of commercial microprocessors running Linux and connected with high-speed networking in MPP (massively parallel processor) arrays. The supercomputing world has never been the same. Today, Linux rules supercomputing.

The latest "Top 500 supercomputer" list of the fastest computers on the planet makes that abundantly clear. Broken down by operating system, this latest ranking has 469 of the top 500 running one kind of Linux or another.

To be exact, 391 are running their own house brand of Linux. Sixty-two are running some version of Novell's SUSE Linux, including such variants as UNICOS/lc and CNL (Compute Node Linux). Red Hat and its relatives, including CentOS, come in second with 16 supercomputers.

As for the non-Linux members of the fastest computer club, IBM's AIX Unix, with 22 computers, is the only serious competitor. Microsoft and Sun, with Windows HPC 2008 and OpenSolaris, are barely in the running, with fives supercomputers for Windows and a mere pair for OpenSolaris.

Linux isn't just setting the standards; it's breaking the record books. The fastest of the fast is now the Cray XT5 supercomputer, known as Jaguar. Jaguar, which runs CNL, didn't just take first place; it blew away the competition with a top speed of 1.75 petaflops per second, leaving the previous record of 1.04 petaflops per second in the dust. (A petaflop is 1,000 trillion, a quadrillion, floating point calculations per second.)

What's even more amazing is that the IBM Roadrunner, another Linux system, which has held the top record, had only broken the petaflop barrier in the summer of 2008. Or perhaps it isn't so amazing when you consider that, with Linux leading the way, the slowest member of this new list can do 20 teraflop (trillion floating point calculations per second). In other words, this list's slowest system would have ranked No. 336 in the last Top 500 list from six months ago.

Linux and improvements in Linux-based MPP programming techniques can't take all the credit. Jaguar, which is located at the Department of Energy's Oak Ridge Leadership Computing Facility, was upgraded from quad-core to six-core AMD Opteron processors. After this upgrade, Jaguar is running almost a quarter of a million CPU cores. In theory, Jaguar can reach a peak speed of 2.3 petaflop per second.

Between the continued improvements in multi-core processor speeds, network fabric throughput, and Linux performance, we can expect to see supercomputers gaining speed at this remarkable rate for quite some time to come. At this rate, we may see a Linux-powered exaflop (one quintillion calculations per second) computer by the early 2010s. That's not just a theory: IBM is already working on the design for such a supercomputer monster for the Square Kilometre Array telescope project. The proposed operating system? Linux, of course.

Email this

Digg this

What People Are Saying

Add new comment

Rocking

Submitted by Anonymous on November 27, 2009 - 11:05 P.M.

Linux absolutely rocks!

Reply | Report this comment

RE:Rocking

Submitted by Anonymous on November 28, 2009 - 2:44 P.M.

Yeah,

But Linux is just for us really bright people.
Mass market junk is good enough for average folks.

Reply | Report this comment

Microsoft pays me to say bad

Submitted by Anonymous on November 19, 2009 - 2:12 P.M.

Microsoft pays me to say bad stuff about Linux on random blogs. It's totally true.

Reply | Report this comment

What's with $JVN's obsession

Submitted by Bob Saget on November 19, 2009 - 1:47 P.M.

What's with $JVN's obsession with speed? FLOPs isn't so much a measure of speed, as it is a measure of computational power. And what's with the weasel wording suggesting that Linux is the reason for it? Jaguar has more than double the amount of cores Roadrunner does, but only provides a 70% performance increase, while both operate at only 75% efficiency (maximum sustained FLOPs vs. Peak operating FLOPs, vs. Theoretical maximum FLOPs). Also worth pointing out is that Jaguar eats up 3x the power that Roadrunner does.

Contrast Roadrunner, to say, JAXA (powered by OpenSolaris) ~12000 cores (vs. Roadrunner's ~122k cores) which operates at 90% efficiency, pumping out 40 GFLOPs/CPU vs. Roadrunner's 12. less than 10% the amount of cores, for more than 10% the output, and more than 3x the output per CPU.

Contrast that to the Super-UX powered Earth Simulator 2, which operates at an astounding 93% efficiency, pumping out 102 GFLOPs/CPU, with a paltry 1280 cores, 1% the amount of cores (compared to RoadRunner), dishing out 10% of the computational power (FLOPs). Even a JAXA-sized cluster would put Roadrunner to shame, and rival Jaguar.

It's fun to focus on only one set of numbers, while ignoring the big picture, and to pretend that it means something else. The more CPUs you toss at a cluster, the more FLOPs it's going to generate, regardless of the OS. You'd expect a cluster utilizing 100x the cores, would put out 100x the FLOPs, but obviously, this isn't the case, 122,000 cores pumping out a petaFLOP is impressive on the surface, but not nearly as much so as 1280 cores punching way above it's weight and putting out 130 TFLOPs.

Similarly a cluster of 224,000 cores cores putting oput 1.7 PFLOPs may be impressive on the surface, until you notice the gross inefficiency.

Beyond that, there's the denial over the distinction between CNL 1 (which IIRC powers Jaguar) and CNL2, where SLES is used on the slave nodes, while all of the heavy lifting is done by the UNICOS kernel), and there's no mention of Jaguar's proprietary interconnect, which I'd hate to break it to you, but the interconnect is much more important regarding efficiency than the operating system itself.

Then there's the dually misleading headline "fastest" while talking about computational power, rather than speed, and "computers" as though these are massive NUMA systems rather than clusters. Sorry, Steven, the "fastest" computer in the world is actually the Sun/Fujitsu-Siemens 64-way SPARCVII-based behemoth that is the M9000, clocking in over a TeraFLOP on a single board (targeted at the enterprise, not HPC mind you), which currently holds the record for the most powerful single computer.

The "fastest" supercluster, however is Jaguar, which is powered by a proprietary interconnect, and has UNICOS at its core. And Linux is nowhere near setting efficiency records. That's Solely between NEC and Sun/Fujitsu, otherwise Roadrunner and Jaguar would have breached the 10 PetaFLOP barrier by now. (Something Fujitsu-Siemens plans to do within the next two years, might I add).

Thanks for playing.

Reply | Report this comment