How much aircraft fatigue is too much?
HPC eases the strain
Here’s another “How HPC saves your worthless hide” type of story - our pals at InsideHPC publicized a collaboration between the Federal Aviation Administration and the Texas Advanced Computing Center (TACC) that should increase air safety for people riding on planes and for the people standing around underneath them.
Briefly, when the FAA or a supplier discovers a problem with an aircraft part, the questions are, “So how big a problem is this? Will it fail? If so, when?” The big commercial manufacturers have the horsepower to figure this stuff out and get it fixed.
But for smaller planes, these questions have been hard – even impossible – to answer, since the simulations needed to figure them out are very lengthy to run, particularly when adding in lots of variables for load, velocity, different materials, and atmospheric conditions, to name a few.
These decisions are important and time-sensitive. Not every problem merits grounding an entire fleet of planes while figuring out whether or not metal fatigue in a particular part is an airworthiness issue. The cost of unnecessary groundings could cause significant economic damage; but not grounding them, if there is a critical problem, could cause things like death.
The TACC people lent a hand by using MPI to parallelize the code to the point where it can run on 256 processors, yielding a speed-up of 188x.
It’s a good story – something to tell your family when they get glassy-eyed during your blow-by-blow account of the latest interdepartmental systems architecture standards war. ®
"Norwegian SAS Pilot landing in a heavy crosswind on a very short runway at night?"
What about the Irish pilot landing on a very short runway, that is also incredibly wide?...
doesn't make sense
From the article: "The invididual simulations aren’t long, but the FCC wanted to create an ensemble analysis that looks at a range of probable values for the critical variables in each situation. This can mean tens of thousands of runs that need to be completed rapidly in response to a detected failure in a flight system."
If it's many runs then they can be farmed & run in parallel, no explicit paralising needed. Trivial, I've done similar myself.
Too easy. I guess something's gone missing in the original article.
Not the issue
It doesn't really matter that the code is probably less efficient. It is realtime sensitive, so what matters is getting it done quickly, not efficiently.