News

Fifteen years ago, the PC you're probably, like most of us, complaining is too slow would probably have qualified as a supercomputer.

Don't believe me? Twice a year since June of 1993, a group of researchers at the U.S. National Energy Research Scientific Computing Center (NERSC), the University of Tennessee and the University of Mannheim have compiled a list of the 500 most powerful supercomputers in the world. On the first list, the number 500 system clocked in at less than the speed of a good Pentium 4 desktop today, as measured by the Linpack benchmark.

How things have changed. Today's Top 500 machines are literally thousands of times faster — the latest number one machine is over 4700 times the speed of number one in 1993.

Despite this, they do much the same things. The systems on the current list forecast the weather, produce climate simulations, help design cars, do complex financial modeling, biological research and drug development, process images, or work in a myriad of other industry segments. Although many are in the United States, eight live in Canada (the highest-ranked, at number 96, is at the University of Sherbrooke — to look at a Google map of locations of the top 100, check out this link.

The top ten systems include four from IBM, two from Cray, and one each from SGI, NEC/Sun, Bull, and Dell.

Yes, I did say Dell. Surprisingly, it produced 17 of the top 500 supercomputers on the most recent list, making it the fourth most popular vendor (IBM and HP are the top dogs, with 236 and 158 machines, respectively; SGI is third, with 20).

Unexpected vendors such as Dell can only appear in the list because the whole architecture of a supercomputer has changed since the elegant Cray 1 supercomputer, a work of art whose look was a consequence of the need to dissipate a lot of heat and move signals along as short a path as possible to tweak up the speed. The industry has now moved away from these single huge computers to collections of many less powerful machines connected to produce one virtual computer of immense processing might. Almost 400 of the top 500 are now these high performance computing clusters (HPCCs).

And that is how Dell got onto the list. According to Dr. Reza Rooholamini, director of enterprise solutions engineering, Dell's HPCC group got its start with a little creative scavenging during an office move. Groups put equipment they no longer needed into the hallway for collection and disposal, and from those bits and pieces, his team was able to cobble together a proof-of-concept system that convinced the Powers That Be that Dell could, indeed, play in the emerging HPCC marketplace.

Today, Dell's HPCC lab is a lot better equipped, but it still relies on the assembly of relatively modest commodity hardware into the powerful clusters that earned it its places in the Top 500.

How modest? Well, the company wanted to prove that Real People (or, at least, ordinary companies, with ordinary techies) could have their own HPCCs, even if they didn't quite make the Top 500, so after its annual new product show and tell, it invited me to come into the lab and build a supercomputer myself.

How could any geek resist an invitation like that?

So it was that I found myself in the HPCC lab with a diminutive engineer named Garima Kochhar, and she and I created a very nice little HPCC in about half a day. Ladies, take note - supercomputing is not a guys-only field.

Admittedly, our cluster had nowhere near the oomph of IBM's Blue Gene/L supercomputer, number one on the current Top 500 list, but it only consisted of 14 machines, each with 4 GB of RAM, and a total of 96 processors. It was constructed from a bunch of interconnected PowerEdge servers just like those many companies use as basic file and print servers (mind you, the high-speed InfiniBand interconnection was definitely not standard, but if the systems can't talk to each other fast enough, the cluster's performance suffers). IBM's system has 131,072 processors and 32768 GB of memory, so it had a bit of an advantage. Still, our cluster was a bit faster than the number two machine in the initial 1993 Top 500.

Blue Gene fills a room, our baby cluster filled a four foot high enclosure. But the greatest benefit of clusters is that they can be expanded. Garima and I could have easily added more nodes to our cluster by simply popping in more servers and connecting them, then installing the clustering software. And, as with any supercomputer, when you submit a job you can choose how many of the installed processors it should use; part of the art of running these systems is in resource scheduling.

Of course, clusters themselves have been around for a long time. It's specialized connections and software that have turned them into "supercomputers". And they only function as supercomputers with the right software, proper tuning and the right kind of problem. But, given those conditions, even our desktop machines could be part of a supercomputer.