A supercomputer cluster wouldn’t care. A cluster of supercomputer clusters might.
Know those moments in movies where they put “all the computers in the world” onto the task of cracking some encryption before it’s too late? This could be used for the technical part of that (once you’ve got all the HPC labs’ cooperation.)
Could an abstraction of "single Linux computer" be actually useful with Earth-scale latencies?
Think of it in a "Compute at home", "folding at home", blockchain-sense then latency doesn't matter as much, I do not know if anyone has tried solving things with big clusters separated by 40ms-60ms latencies.
Maybe if you had some kind of amazing cache system that organized everything into one big tree so that data needed for different computations is as local as possible. It would be very challenging though.
Re:Popcorn Linux
by convolvatron on Sunday May 03, 2020 @04:04PM
(#23061461)
at Cray, at one point we had custom vector processors, custom SMT(for lack of a better term) processors, FPGAs on the roadmap, GPUs, and commodity AMD processors...and a strong desire to ship them in the same frame
that may have been the only time in history though
Re:Popcorn Linux
by marcosdumay on Sunday May 03, 2020 @03:46PM
(#23061342)
Some ISAs can be way more effective at some parts of the problem than others, while still not being an optimum for the entire problem.
That's why your computer has a GPU.
Can it run on a gpu so to abstract always cuda
Re:Popcorn Linux
by marcosdumay on Monday May 04, 2020 @01:57PM
(#23071169)
I imagine it can't (that would be way too good, everybody would be using it already). I also don't think there are many examples of supercomputers with diverse ISAs today.
But a system like this is a requisite for non-uniform computers to exist. Thus it must exist without the other, at least for a while.
Why would servers/supercomputers care about heterogeneous hardware (including multiple ISAs, even!)? If I'm playing around at home, that's great; I'd love to make a cluster out of a couple of old laptops, an old phone, and a Raspberry Pi. But at work, I'd just... buy a dozen identical servers. Even across generations, x86 is mostly uniform enough to just ignore the changes (with some minor gotchas around extensions, but if you only migrate forward that should be fine too).