Slashdot Mirror
SpaceX Will Deliver The First Supercomputer To The ISS (hpe.com)
Slashdot reader #16,185, Esther Schindler writes:
"By NASA's rules, not just any computer can go into space. Their components must be radiation hardened, especially the CPUs," reports HPE Insights. "Otherwise, they tend to fail due to the effects of ionizing radiation. The customized processors undergo years of design work and then more years of testing before they are certified for spaceflight." As a result, the ISS runs the station using two sets of three Command and Control Multiplexer DeMultiplexer computers whose processors are 20MHz Intel 80386SX CPUs, right out of 1988. "The traditional way to radiation-harden a spacecraft computer is to add redundancy to its circuits or by using insulating substrates instead of the usual semiconductor wafers on chips. That's expensive and time consuming. HPE scientists believe that simply slowing down a system in adverse conditions can avoid glitches and keep the computer running."
So, assuming the August 15 SpaceX Falcon 9 rocket launch goes well, there will be a supercomputer headed into space -- using off-the-shelf hardware. Let's see if the idea pans out. "We may discover a set of parameters with which a supercomputer can successfully run for at least a year without errors," says Dr. Mark R. Fernandez, the mission's co-principal investigator for software and SGI's HPC technology officer. "Alternately, one or more components of the system will fail, in which case we will then do the typical failure analysis on Earth. That will let us learn what to change to make the systems more reliable in the future."
The article points out that the New Horizons spacecraft that just flew past Pluto has a 12MHz Mongoose-V CPU, based on the MIPS R3000 CPU. "You may remember its much faster ancestor: the chip that took you on adventures in the original Sony PlayStation, circa 1994."
So, assuming the August 15 SpaceX Falcon 9 rocket launch goes well, there will be a supercomputer headed into space -- using off-the-shelf hardware. Let's see if the idea pans out. "We may discover a set of parameters with which a supercomputer can successfully run for at least a year without errors," says Dr. Mark R. Fernandez, the mission's co-principal investigator for software and SGI's HPC technology officer. "Alternately, one or more components of the system will fail, in which case we will then do the typical failure analysis on Earth. That will let us learn what to change to make the systems more reliable in the future."
The article points out that the New Horizons spacecraft that just flew past Pluto has a 12MHz Mongoose-V CPU, based on the MIPS R3000 CPU. "You may remember its much faster ancestor: the chip that took you on adventures in the original Sony PlayStation, circa 1994."
If you look at the ISS webcam when it switches to the interior cam, there's a few laptops (one running Ubuntu) tied to the sides of the walls.
The part you seemed to have missed is: This is an experiment to learn whether an alternative approach to hardening can be developed. If it's successful, the benefits would be obvious.
Experiments are the raison d'etre for the ISS... so why is this a problem?
#DeleteChrome
Whenever something inexplicable happened while testing a video game, I've always put down "gamma radiation" on the bug report. The developers hated that term but they couldn't explain why it happened either.
At a guess it's because sending data back to earth for processing isn't great when you're a long way away - the latency between Earth and Mars, for example, can get up to about 21 minutes. If your lander has to adjust for local weather systems, or your orbital station needs to make corrections due to local changes in EM fields, or if you're just operating in an environment where you can't predict exactly what conditions you're going to find, you need to do a lot of calculations to correct.
Of course this isn't an issue for the ISS, with a latency shorter than my ping to Google (seriously, my internet sucks). But if we're going to look at landers on Europa, exploring Ganymede etc it'll be easier if we can do some heavy computing on the fly. So test now in a controlled environment, and get it right for when we send stuff on 20 year missions.
I'd instead go with a RAIA -- a horde of off-the-shelf ARMs. Within the power budget of a single 20MHz 80386 you can fit nine 2GHz SoCs. Have them vote -- there's no way every single of them gets hit by a ray within a time slice. Periodically, resync their memory (especially when the vote disagrees). A 2GHz machine can take quite an overhead while doing the work previously done by a 20MHz one...
This assumes the 386 was alone -- it was at least doubled or tripled. So if you don't need 18x or 27x redundancy, you can do something else with the extra power.
But let's assume you do want that 27x redundancy. It's still a two orders of magnitude speed boost, and that's assuming same speed clock-to-clock. Which is wrong, as 386 timings were downright scary. Especially in floating point, with a hundred or more clock cycles per instruction. Modern ARM on the other hand includes a vectorized FPU...
The creatures outside looked from Alt-Right to Antifa; but already it was impossible to say which was which.
The approach is interesting, but putting it in the ISS is only slightly more demanding than putting it on your desk. Both remain well under the protection of the Van Allen belts. The real test is out beyond the Van Allen belts where the radiation really gets tough.
Who needs radiation hardening? Just send a Proliant rack server up there and call it good! That's why we're SpaceX and they're luddites!
Unless I've missed something and HPE has been sold to him, Elon Musk is just the owner of the company that will deliver this computer to the ISS. He did not design the experiment.
Why not do the heavy computing down here on the ground, where it is so much easier?
Bandwidth. ISS has 3 megabit upstream, 10 megabit downstream. Yes, megabit, not gigabit. And that's a massive upgrade over what it had for years, which was 2400 baud. There's any number of science experiments people would like to run that would benefit from beefy local processing handling large amounts of data. So much data that neither transmitting it off station nor storing it and physically transporting it off station is currently feasible. The bandwidth isn't available or the storage is too expensive.
That may change in the 2020s. I'd bet a pizza that SpaceX will be including upward-facing antennas in their satellites, not just Earthward-facing, in order to talk to their own rockets at high bandwidth regardless of where they are in their trajectories. Still, it's going to be quite some time before that option exists, so experiments to determine the feasibility of local processing are worth conducting.
Triple redundancy is what you want if you're running the operations in parallel and looking for consensus. If you're running them in series, you run them twice and only repeat if the two results don't match. The chance of two SEUs happening that disrupt the same operation in the same way, twice in a row is very unlikely.
If you want a vision of the future, imagine a youtube comments section scrolling - forever.