Friday, July 17, 2009

Here I am, brain the size of a planet...

... and they ask me to run a shock tube test.

Contrary to its name, no actual astronomical observing is done at the National Observatory in Tokyo. As the city spread, the domes scattering the campus were converted into museums or declared part of Japan's history and left to be (somewhat ironically) observed. Real observing, such as at the Japanese Subaru telescope, is conducted in places like Hawaii while the city observatory buildings operate as offices for astrophysicists and house the tools for a superior different kind of astrophysics; the theoretical group's Cray supercomputer. (The security surrounding access to this machine made me wonder whether information about its existence was controlled. If so, me blogging about it... not so smart. But Google revealed that Cray advertised its installment at the Observatory in their newsletter, so they're totally going down before me.)

Half the battle in running large numerical simulations is getting your code installed on a new computer system. It does not matter how many times I do this, it never seems to be simple. Of course, this time had the added complication of all the system details being in Japanese.

Once I'd bashed everything into submission (including the Kanji dictionary), I set about trying a test simulation to check the code did not fail when run. This produced the response:

Job exceeds queue resource limits

Huh? This was a small programme that followed the motion of a fluid along a pipe (known as a shock tube test) and I'd only requested a single processor for it. It should not have taken more than a minute to complete. In fact, I could have solved it myself... though in a bit more than a minute. Upon investigation, I discovered that the error message above is misleading. What the Cray was really saying was:

Your job is too small for me to consider it worth my time to compute.

In fact, it transpired that the Cray wouldn't have anything to do with me or my simulations until I requested a minimum of 20 cores each with 4 processors on them. Which I did .... for this one shock tube test. It was the numerical equivalent of squishing a crippled ant with a heavily armored tank.

Well, fine. Two can play at this game. You want a real problem, Cray? One universe demand coming up...

No comments:

Post a Comment