Embedded machine

[millsie]

Hello,
This is an odd one. How difficult would it be to build a custom OS for GIMPS?

I am talking about a stripped down, streamlined OS. Probably UNIX-like.

Would have to be simple to install, have little hardware support (why use up memory for USB drivers etc, when they aren't gonna be used).

I am thinking something along the lines of Download-install-input username and password for GIMPS-run. It just does that.

Any thoughts?

[R.D. Silverman]

Originally Posted by millsie

Hello,
This is an odd one. How difficult would it be to build a custom OS for GIMPS?

I am talking about a stripped down, streamlined OS. Probably UNIX-like.

Would have to be simple to install, have little hardware support (why use up memory for USB drivers etc, when they aren't gonna be used).

I am thinking something along the lines of Download-install-input username and password for GIMPS-run. It just does that.

Any thoughts?

What possible purpose would it serve? What would it provide that is
not already provided now?

Disabling drivers is easy. Memory is cheap.

[millsie]

Originally Posted by R.D. Silverman

What possible purpose would it serve? What would it provide that is
not already provided now?

Disabling drivers is easy. Memory is cheap.

Purpose wise it would mean the utilisation of cheaper hardare for a dedicated device. It could also squeeze more processing out each CPU. It could form the basis of a system that would link multiple boxes running this to make them work as one. Possible to process 64-bit or larger number over multiple 32-bit CPU's.

Memory is only cheap if you can afford it. Same as food, ask a Zimbabwean.

[TimSorbet]

Originally Posted by millsie

Purpose wise it would mean the utilisation of cheaper hardare for a dedicated device. It could also squeeze more processing out each CPU.

You could just get a Linux distro that's as barebones as possible. What benefit would something you're suggesting have over that?

Originally Posted by millsie

It could form the basis of a system that would link multiple boxes running this to make them work as one.

Trying to multithread over separate boxes would mean a bandwidth limitation that would probably make it much slower than any of the individual CPUs.

Originally Posted by millsie

Possible to process 64-bit or larger number over multiple 32-bit CPU's.

What?

[millsie]

Originally Posted by Mini-Geek

You could just get a Linux distro that's as barebones as possible. What benefit would something you're suggesting have over that?

Trying to multithread over separate boxes would mean a bandwidth limitation that would probably make it much slower than any of the individual CPUs.

Hmm, I didn't think about the bandwidth issue, though I am sure somebody can figure out how to compensate for that (maybe a way of pre-splitting the task ahead of time?)

Yes a cut down Linux distro is something I am thinking along the lines of, but at a very low-level altered to run the software for GIMPS. Sort of integrated into the kernel.

Take the old Acorn RISC machines for example. Basic software was included in ROM with the OS. It ran extremely fast (for the day).

[retina]

Originally Posted by millsie

Hmm, I didn't think about the bandwidth issue, though I am sure somebody can figure out how to compensate for that (maybe a way of pre-splitting the task ahead of time?)

Sure. How about we work the LL from each end. You know, one starts with the squarings and the other starts with squareroots, and they meet in the middle. Sounds simple enough. I'll let you work out how to do the squareroot part.

[jasonp]

Originally Posted by millsie

Memory is only cheap if you can afford it. Same as food, ask a Zimbabwean.

Memory is only expensive if you haven't bought any lately. 1GB of it is $10-15 now.

[TimSorbet]

Originally Posted by millsie

Hmm, I didn't think about the bandwidth issue, though I am sure somebody can figure out how to compensate for that (maybe a way of pre-splitting the task ahead of time?)

No, not really. You obviously don't really know anything about what's going on behind the scenes, so please don't talk like you do.

Originally Posted by millsie

Yes a cut down Linux distro is something I am thinking along the lines of, but at a very low-level altered to run the software for GIMPS. Sort of integrated into the kernel.

How would integrating GIMPS into the Linux kernel make it run any faster? Prime95/mprime already uses all the assembly code optimizations and stuff it can to speed up as much as possible. The only things left to do are have the fastest hardware possible and have the fewest other processes stealing CPU time and other resources from mprime as possible.

[cheesehead]

Originally Posted by millsie

Hmm, I didn't think about the bandwidth issue, though I am sure somebody can figure out how to compensate for that (maybe a way of pre-splitting the task ahead of time?)

You may not have read the threads where that sort of thing has been discussed in the past, long before you joined.

The basic nature of the Lucas-Lehmer test is such that the iterations must be done in sequential order.

(In case you're wondering about a shortcut: Iteration values grow rapidly. Without modular arithmetic to cut back the size of each iterated value, only the first few hundred full iteration values could be stored within the size of the known universe. It is only by use of modular arithmetic that values of iterations past the first few can be computed. But since the modulus is different for each different Mersenne number tested, there's no way to use the iteration values from one test to speed up another test.)

Paralleling could be done, to some extent, within each FFT transform, but all those results have to be put back together at the end, before the next transform could be parceled out for parallel execution. That is where bandwidth comes in.

Yes a cut down Linux distro is something I am thinking along the lines of, but at a very low-level altered to run the software for GIMPS. Sort of integrated into the kernel.

Take the old Acorn RISC machines for example. Basic software was included in ROM with the OS. It ran extremely fast (for the day).

The trouble is that streamlining an OS for GIMPS would speed execution by only a fraction of a percent. It's just not worth the effort, in regard to efficiency of GIMPS execution.

It could have educational value for the developer, as a project for learning about operating system internals, but not as a noticeable contribution to GIMPS progress.

- - -

I applaud your inquiry and questioning. Just be prepared to get some curt responses from veterans when your question turns out to have been answered, or your suggestion, to have been refuted, long before you joined GIMPS. :-)

I keep promising to extend our FAQs to include topics such as this, but ... (* sigh *) ... Feel free to PM me with a polite prod about my progress there.

[millsie]

Thank you cheesehead for you reply. It was informative. I shall go and mull this over.

Thank you again.