What does a supercomputer do?

June 27, 2007 2:21 AM

What do supercomputers actually do? Specifically...

In this BBC story about IBM's new Blue Gene/P supercomputer, it states that its predecessor, the Blue Gene/L, is housed at the Lawrence Livermore National Laboratory in California, where it is 'used to ensure that the US nuclear weapons stockpile remains safe and reliable'. My question is, how?

posted by jonathanbell to technology (11 comments total)

Supercomputers are commonly used to perform numerical simulations of physical processes, in lieu of being able to test those physical processes in the real world.

In the case you mention, they seem to be running physics simulations to see how various materials react to extreme environments (1, 2), which presumably helps them design and improve nuclear weapons systems.

Also, the Stockpile Stewardship program is responsible for the reliability and safety of all the old nuclear weapons that are lying around. They study the ageing of nuclear materials (3) -- again, something that simulation could help with.

If you want to get broad-brush about it, they're choosing small-scale simulation on a computer as an alternative to actually doing nuclear tests.
posted by chrismear at 2:55 AM on June 27, 2007 [1 favorite]

Wikipedia has a nice overview of stockpile stewardship.
posted by chrismear at 2:56 AM on June 27, 2007

Supercomputers are generally used to do simulations. The problem with computers is that they are great at discrete simulations (what happens at point x at time y given inputs z?), but they tend to suck at analogue simulations (what will happen in this area over a time period?).

You can approximate analogue simulations by doing discrete/digital simulations over thousands or millions of points of space and time. The more points you can calculate in a simulated hurricane or nuclear explosion, the closer you get to the real thing. Hence bigger and badder supercomputers can do simulations that are closer and closer to true analogue representations.

As chrismear says, the stewardship program are probably running nuclear explosion simulations with varying inputs based around the age of the stockpile and the half-lives of the component compounds in the warheads.

Rest assured that the 'safe and reliable' thing is mostly based around "will they still do as much damage as reliably as they did when we tested the real thing?". 'Safety' is not really an issue when the warheads need external detonation to get them anywhere near critical mass.
posted by pivotal at 3:20 AM on June 27, 2007 [1 favorite]

Thank you both.
posted by jonathanbell at 5:52 AM on June 27, 2007

Blue Gene/P is being used for weather/atmospheric simulation. Basically, anything that can't be reduced to a sufficiently simple model for desktop/cluster computation has to be thrown at a supercomputer to do it in better-than-realtime. I remember working on one watershed groundwater/irrigation model that we scrapped because it actually ran slower than realtime - by the time it would have predicted anything, it would have already happened.
posted by devilsbrigade at 5:58 AM on June 27, 2007

Historically, supercomputers were distinguished by their vector processing capability - doing the same operation very fast over a stream of data as opposed to doing it one element at a time.

Modern supersomputers are sometimes vector but are mostly massively parallel - Wikipedia indicates that BlueGene/L is a cluster of 65,536 computer nodes, each with two processors. The advantage of the supersomputer is that it simplifies running a single program across all these processors and coordinates the data streams efficiently.

This is useful for simulations, as other posters have noted. It seems that the US government has gone soft and would rather simulate nuclear explosions rather than simply blowing up Nevada repeatedly as they did in the 40's. Wimps.
posted by GuyZero at 6:15 AM on June 27, 2007

To add to pivotal's answer: the European Centre for Medium-Range Weather Forecasts had (may still have) a policy of generating its seven-day forecasts in a fixed period of time, I believe it was 3 hours.

As they acquired faster and faster supercomputers (up to the mid-90s they'd always had the latest model Cray), they were able to increase the number of discrete points they used in their model, while retaining the same runtime. I'm sure that they have a lot more data points in their model today than they did in 1995.
posted by lowlife at 7:21 AM on June 27, 2007

Also mentioned in the article you link to is Roadrunner, a "next generation" of supercomputer. Built for Los Alamos National Labs, it will be the fastest computer in the world when it hits top speeds (planned to be early 2008). One of the things Roadrunner will be doing, in addition to simulating how nuclear weapons will age as chrismear describes, is building probability models of vaccines--vaccines that would be the most likely to be effective on the greatest number of people. In some cases it would take about a decade to compute the same information on a desktop PC, based on the complexity of the disease and the global aspect of a vaccine.
posted by cocoagirl at 7:32 AM on June 27, 2007

I'm a pretty big fan of Japan's "Earth Simulator" that simulates global climate models to evaluate how climate change would change the earth.
posted by MonkNoiz at 7:35 AM on June 27, 2007

There was an article in WIRED about this: This is Not a Test
posted by daser at 8:31 AM on June 27, 2007 [1 favorite]

Supercomputers, or more commonly large clusters, are also used quite frequently in computational biology applications. Modeling protein folding, or molecular interactions is one example. Doing large-scale genomics is another (3 billion base pairs per human makes for huge data sets!)
posted by chrisamiller at 9:51 AM on June 27, 2007