26
$\begingroup$

All of the top five answers to The Martian: Does it really take a supercomputer to calculate spaceflight maneuvers? are essentially "no, orbital mechanics isn't rocket science". Okay I have used some artistic license there, but these days even laptops do gigaflops, and that's not even counting the GPU; optimizing trajectories like a flight from Earth to the Moon are not likely going to need supercomputers as near as I can tell.

So I was surprised to read the following:

Aitken will consist of 1,150 nodes, with each node using two 20-core second generation Intel Xeon Scalable processors and Mellonox InfiniBand interconnects. Total numbers for Aitken come to 46,080 cores and 221TB of memory across 1,150 nodes for 3.69 petaflops of theoretical peak performance.

Aitken will reside at NASA Ames' new modular supercomputing facility, which had its grand opening last Thursday. The new facility is based on a Modular Data Center (MDC) design, and can accommodate 16 modules, with Aitken claiming the first. Aitken will aid in landing astronauts on the South Pole region of the moon by 2024, as part of NASA's Artemis program.

The new supercomputer will be used by more than 1,500 scientists and engineers from across the country, including on projects like developing a more efficient quadcopter or simulating the inside of our sun. The job at the top of the priority list will be running modeling and simulations of the entry, decent and landing to the moon for the Artemis project.

Question: Does the "modeling and simulations of the entry, decent and landing to the moon" really need 46,000+ cores, 3.69 petaflops and 221 TB of memory?

What does "entry" even mean when landing on the Moon?

$\endgroup$
  • 31
    $\begingroup$ If you're modelling the behaviour of the spacecraft structure (vibration, for instance) you could easily consume all that CPU $\endgroup$ – Steve Linton Aug 26 at 13:07
  • 5
    $\begingroup$ They may simulate 46000 variations of the trajectory from the cape in Florida to the south Pole of the Moon at once in less than a minute. Finding the optimal launch window by calculating the trajectory every 15 minutes within a full year would need 35040 variants. $\endgroup$ – Uwe Aug 26 at 14:03
  • 9
    $\begingroup$ Leaked image of Aitken simulation: fictionphile.com/wp-content/uploads/2019/06/… $\endgroup$ – Organic Marble Aug 26 at 14:07
  • 13
    $\begingroup$ Ages ago, Jack Ganssle ranted that from the time his shiny new Vaio laptop powered up until Windows let him log in, its CPU could have run every sim he ran in his Apollo career... a million times over. $\endgroup$ – Camille Goudeseune Aug 26 at 17:24
  • 8
    $\begingroup$ In response to the question in the title, landing on the Moon doesn't require floating-point math. $\endgroup$ – Mark Aug 26 at 21:00
35
$\begingroup$

First of all, it can require a lot of computer power to compute trajectories if they involve multiple gravitational slinghots to reduce fuel usage. This isn't because computing each segment is hard but because the search space is at least potentially exponential in the number of gravitational slingshots. I am pretty sure that this is what the thing in The Martian referred to. (There may be clever tricks for reducing the search space which I don't know about.)

However an Earth-Moon trajectory probably doesn't include a lot of slingshots!

But computing trajectories is not all of spaceflight: it's not even most of it. For instance, you might very well want know how your design behaves in the atmosphere, for instance, or how your engines behave under load, or what vibrational modes are going to be a problem in the structure of your vehicle and how you can minimise mass while not having the thing suffer from pogo or some other unfortunate vibration still less fail altogether. Is the thing strong enough to withstand impacts of various kinds – how does your fancy material fail when its hit by bits of foam, say? And there are many, many other engineering questions you need answers to.

You can solve these problems in several ways:

  • over-engineer the thing so that you are really pretty confident that it will not suffer from any of the above;
  • build lots of experimental things, fly them and see what goes wrong, refining the design successively;
  • build models of the thing, in a computer, and simulate their behaviour.

The last of these is hugely cheaper than the first two: it lets you explore a huge number of options, and results in something which may be close to optimal, without having to physically build lots of experimental vehicles & bits of them.

So you buy (time on) an HPC system, and do that.

$\endgroup$
  • 10
    $\begingroup$ Bit off-topic, but here's a good document on how they limit the search space for gravitational assist trajectories and optimize the algorithms to find them. It's a really dense, but ultimately cool, read. In artificial intelligence it's often important to specify what you know is incorrect if the thing you're searching for is not discernibly finite to speed up the algorithms. $\endgroup$ – Magic Octopus Urn Aug 27 at 17:05
  • $\begingroup$ If I recall, the situation in The Martian was a slingshot around Earth, since the ship was already on its way back there and that was actually the fastest way to reverse trajectory and get back to Mars (following a hasty resupply in the middle of the maneuver). Been a while and I remember the book more than the movie, so there might be differences between them. $\endgroup$ – Darrel Hoffman Aug 27 at 18:50
  • $\begingroup$ @MagicOctopusUrn: thank's, I will read that: I have some history in old-fashioned searchy AI, so I'm interested n this. $\endgroup$ – tfb Aug 27 at 19:42
  • $\begingroup$ @DarrelHoffman: I think that was what happened in the film too, and it seems to me that something like that can't require vast resources to compute, but perhaps I'm wrong. $\endgroup$ – tfb Aug 27 at 19:43
  • 1
    $\begingroup$ As for The Martian there's another factor at work, also--their objective was to get to Mars ASAP given the available delta-v. AFIAK there is no simple solution to the problem, you need to search the whole problem space. Since a life was on the line you're not going to risk getting caught in a local optimum, you'll simulate everything rather than try to converge on the right answer. It's not that any projection is that hard, it's that you run an awful lot of them. $\endgroup$ – Loren Pechtel Aug 30 at 15:26
18
$\begingroup$

All that computing power is not dedicated to the Artemis project. As you quote in the body of your question,

The new supercomputer will be used by more than 1,500 scientists and engineers from across the country, including on projects like developing a more efficient quadcopter or simulating the inside of our sun.

Not all of this computing power is used for the Artemis project, so it's clear you don't need all of it to simulate a moon landing. Without knowing more about the actual simulations they're running, it's near-impossible to say how much computing power is actually needed. Heck, I can fire up Kerbal Space Program on my 5-year old laptop and simulate a moon landing, but I have to assume NASA's simulations are a bit more detailed than that.

$\endgroup$
  • $\begingroup$ The problem with "Not all of this computing power is used..." is that it includes "zero of this computing power is used". I think a good answer should explain if Artemis needs some reasonable fraction, like 5% or 20%, or not. $\endgroup$ – uhoh Aug 26 at 13:31
  • 6
    $\begingroup$ HPE builds supercomputer for NASA, aimed at future moon missions is a comma splice error. NASA is aimed at the moon, not the computer, which unlike this clickbait title would suggest, NASA's new moon-landing supercomputer is more powerful and more eco-friendly, it will not be landing on the moon. - If you need more than 1,500 reasons for justification, I'd assume its the capability to make any conceivable calculation in real time. $\endgroup$ – Mazura Aug 26 at 23:24
  • 8
    $\begingroup$ @uhoh Well, presumably some of the power is being used for Artemis, unless the claim quoted in the question is just false. And, as the answer explains, it's impossible to estimate the actual proportion from the information available. Your question is like "Does talking to your grandmother really require gigabytes of storage, and more CPU power and graphics resolution than a desktop PC had pretty recently?" "No. People use their cell phones for other stuff, too." "But what proportion of time do people spend talking to their grandmothers?" $\endgroup$ – David Richerby Aug 27 at 12:11
10
$\begingroup$

Does the "modeling and simulations of the entry, decent and landing to the moon" really need 46,000+ cores, 3.69 petaflops and 221 TB of memory?

No. A machine with the power of a Raspberry Pi is good enough for solving some simulation problems. A nice desktop machine will have more than enough oomph for solving many, many more simulation problems. On the other hand, there are a small number of problems that require boatloads of computational power.

Coupling the multi kilohertz simulation that is needed to investigate flex and slosh issues with the 2159x2159 EGM2008 Earth gravity model, the highest Earth rotation mode, the highest precision Earth atmosphere model, and then use that to perform tens of thousands of Monte Carlo runs from launch to landing, and yes, you might find you need a supercomputer. But doing so is beyond stupid.

That said, some parts of NASA love overkill and have a one-size-fits-all mentality. This leads to a single simulation capable of modeling flex, slosh, uses a 2159x2159 Earth gravity, and uses a ridiculously expensive Earth atmosphere model with a plate model of the vehicle. In this case one might well need a supercomputer to solve the simplest of problems. Or one could get qualitatively the same results using a far simpler simulation that runs quite nicely on a Raspberry Pi.

$\endgroup$
  • 5
    $\begingroup$ Computational scientists, as a rule, try to use the most computationally efficient model available that accurately reproduces the phenomenon of interest. Sometimes this is hard to do and scientists err on the side of accuracy. However, you seem to imply that they NASA is wasting supercomputer time using models that are unnecessarily expensive for the purpose. Do you have any evidence of that these supposedly "overkill" simulations have been performed (publications by NASA scientists)? Have other scientists pointed this out in the literature? $\endgroup$ – WaterMolecule Aug 27 at 13:45
  • 3
    $\begingroup$ @WaterMolecule this likely comes from direct, first-hand knowledge. You can check people's profiles and sometimes find stuff out. $\endgroup$ – uhoh Aug 28 at 2:49
8
$\begingroup$

They will help Artemis, but not how you might think

Experience with the Apollo program tells us how much computing power it takes to get to the moon. The spacecraft itself was controlled by the Apollo Guidance Computer, which had an instruction set, instruction speed, and memory space somewhat comparable to the microcontroller inside a microwave oven today.

So you don't need a supercomputer to operate the spacecraft.

There seems to be a misconception that the AGC was autonomous. In fact, most of the programs required dozens of parameters to be pre-calculated back in Houston and then entered into the AGC. These parameter calculations were performed by IBM mainframes back in Houston. According to the Apollo Program Summary Report, section 7.3.1 "Command Systems":

The computer had matured, as had the use of digital communications. While radio-frequency modulation remained the same, modified 642B computers replaced the digital command system. Up-link commands were no longer transmitted directly from Houston. Each remote station computer was programmed with unique command words, and the execute decision from the Mission Control Center became requests for the pre-programed command words. Additionally, the Mission Control Center was no longer limited to command execution through three stations because 13 prime range stations within the Manned Space Flight Network were linked to a 494 computer in Houston by a similar 494 computer at the Goddard Space Flight Center in Greenbelt, Maryland.

Although today's dumbphones are more powerful than those mainframes, they were state-of-the-art at that time. Not did Houston have an advantage over the spacecraft in terms of better computers, but it also had more expertise and better tracking of the spacecraft's position and velocity.

So you don't need a supercomputer to calculate spacecraft trajectory.

So why does NASA have supercomputers? As I explained in my Retrocomputing answer to "Did a shuttle launch take most of the world's computing power?",

The division performs calculations for fluid dynamics, aircraft design, weather modeling, and ocean current prediction. They also have resources for data visualization and massive data storage. This is all for research, and isn't actually needed for a shuttle launch.

Artemis will need rocket engines. If new ones are being designed, you will want to simulate the flows inside the engine. You'll need a supercomputer for that.

Artemis will need a vehicle to re-enter Earth's atmosphere. You will want to simulate the re-entry of various shapes and sizes. You'll need a supercomputer for that. (In fact, the final shape of the Space Shuttle was developed through supercomputer simulations.)

Launches and landings need to know the weather, many days in advance. You'll need a supercomputer for that.

So you'll need a supercomputer, but not the way you expected.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.