100 computer system

If any one could figure out a way to design a software that could run on a server to over 100 computers simultaneously, I think I could run up to 100 suspect numbers per month. Please Respond.

Why do you want to do that?
Thats one BIG system dontcha know

Why do you want to do that?
Thats one BIG system dontcha know

[R.D. Silverman]

Originally Posted by MRiblett

If any one could figure out a way to design a software that could run on a server to over 100 computers simultaneously, I think I could run up to 100 suspect numbers per month. Please Respond.

It's been done. 100 computers is tiny.

Look up Grid Computing. I built such a system back in '85. Today,
there are a number of companies providing this capability. Look up
'Axceleon'

[Washuu]

Aloso google for such terms as:

"Beowulf"
"Cluster of Workstations"
"OpenMP"
"MPICH" - this one works for Windows NT!
"LAM"

BTW: Is someone here that can REALLY re-code a program to use MPI or similar environment?

ltsp.org
also ask prime monster for help he runs the prime monsters aka diskless these are very easy to manage after set up.

[ewmayer]

Originally Posted by Washuu

Is someone here that can REALLY re-code a program to use MPI or similar environment?

What would be the point? GIMPS work is what is referred to as "embarrassingly parallel," since each M(p) can be tested completely without any interprocessor communication. Distributing an FFT over multiple CPUs is always going to incur some kind of performance hit, and thus is only useful for e.g. double-checking an alleged new prime result. And there are already DC programs (e.g. Glucas) that are MPI enabled to the needed extent to do that.

If you've got a 100-CPU system burning a hole in your pocket and you want it to use it to do GIMPS work, no problem - just run an instance of whatever program is best for your CPU type on each CPU. That will maximize your system's contribution to the project.

If what you seek is instead the macho thrill of "I can do an LL test faster than the next guy by blasting it with N >> 1 processors," well, then you're not interested in the science anyway. And in that case a bottle of Viagra and a fast car will probably cost you less than a 100-CPU cluster.

[Washuu]

Originally Posted by ewmayer

What would be the point? GIMPS work is what is referred to as "embarrassingly parallel,"...

My question was general, not in the meaning of GIMPS. (I said "A program, not THE program"). I know that computing the FFT's is very hard to paralellize. What would be most paying IMO, is paralellisation of computing matrixes, because these computations demands A LOT of MEMORY and PROCESSING SPEED.

But my question was even more general - if there is someone on this forum who can re-program any not very complicated source code to use effectively such environments as OpenMP or MPI...

Washuu

[Wacky]

Originally Posted by Washuu

paralellisation of computing matrixes, because these computations demands A LOT of MEMORY and PROCESSING SPEED

Actually, in many cases, even this is no longer true. Individual processors with > 1GB of Ram are not uncommon. When it comes to connecting multiple processors, "the network is the computer" is actually the dominating factor.

If the many processors can share a common memory, the results can be amazingly better than those of a single processor.

If, on the other hand, you must rely on some network to connect the processors, it is probably going to be the network, rather than the processors, that becomes the limiting factor.

As an example. on a 32 processor (16x2) cluster of 1GHz processors,interconnected with Gigabit Ethernet and an expensive switch, Paul was actually able to run a siever "in the background" on each of the 16 boxes without affecting the rate at which the block Lanzcos matrix solution proceeded. Each node had "more than enough" real memory and processing speed to completely clog the network.

[ewmayer]

Originally Posted by Washuu

My question was general, not in the meaning of GIMPS.

Ah, I had in mind the thread author's sentence:

"I think I could run up to 100 suspect numbers per month."

...which refers specifically to GIMPS work. So he (or she), not you, wins the Viagra and sportscar.