Sun-seeking NASA boffins steer FLEETS of satellites... BY HAND

With a little help from some serious iron, naturally

HPC blog

Looking for something to test your Star Wars video game skills? How about this: launch four satellites and fly them in eccentric orbits around the Earth.

Each satellite is 1km apart from its neighbour, which sounds like a lot of space until you consider that they’re travelling at several kilometres per second during the fastest part of their orbits.

The satellites aren’t just in static orbits; you have to vary their course and the formation at least once every two weeks – more often if you see a collision coming up.

And the cherry on the cake? There isn’t any fancy collision prediction or avoidance gear on board the satellites; they’re essentially just flying lightning rods, out there to capture a magnetic reconnection event.

If this sounds like fun to you, you need to watch the GTC14 session given by the folks from NASA’s Goddard Space Flight Center. In it, they talk about the first rule of this mission (“don’t crash the satellites”) and the difficulties they face in making sure they don’t wreck a \$1bn mission.

There are quite a few things that can go wrong when trying to keep the sats in a tight formation. For example, the satellites' thrusters might not burn at exactly the right time, or provide the rated thrust, or may cut off early – or late. Any of these conditions could spell doom for the formation and/or the satellites.

NASA deals with this uncertainty by running a massive Monte Carlo-like simulation of every maneuver daily. This yields a probabilistic location for each satellite at the current moment in time.

How do they figure all this out?

In the third group from the left, you can see where the groupings of blue and brown dots might overlap. This shows that there are potential collisions in this time slice - a very bad thing.

Calculating these points turns out to be a pretty darned large job, as the Goddard Space Flight Center points out:

So after 34 billion calculations, we have a pretty good idea where the satellites are now. But we also need to know where they’ll be later on in the week, just to make sure that the formation is tight and they won’t be crashing against each other at any point.

To run the full 10-day location forecast, you essentially calculate the 34 billion possible collision points 172,800 times. Which is, not surprisingly, an even larger computational load:

As noted on the slide above, using traditional CPU-based systems, it takes at least two weeks to run a 10-day collision forecast – which turns their “forecast” into essentially a history report. Not useful.

Fortunately for them, their code is embarrassingly parallel, meaning it can be chopped up and run simultaneously on multiple systems without losing fidelity or performance. Given that they needed maximum speed, they decided to optimise their code for use on GPUs.

The results were pretty good. The same calculations that took two weeks on their traditional systems now take 20 minutes on dual NVIDIA K20s. That’s about 1,000 times faster, which isn’t bad at all.

Goddard also noted that the GPU solution gave them the ability to map more points and run more calculations, meaning that their results are many times more precise than before.

Check out the brief 20 minute presentation to learn more about the magnetic reconnection mission. Towards the end of the presentation, they also discuss their upcoming missions. One is to replace the Hubble telescope with the new and improved James Webb scope.

Another intriguing future NASA mission is to find “zero energy pathways”, which are points between universes that are frictionless and supposedly require no energy to transit. If true, these passages would give us the ability to live in a different universe and commute back to this one for work and shopping. ®