Mini-Projects

[MrHappy]

Hi folks!

I want to propose some well-defined mini-projects or tasks and a combined approach to finish/achieve it, some selfmade milestones... with the overall goal of speeding up GIMPS in the present or near future...
like: no DC should be handed out without TF-attempts up to the 69 bit level.
or: same with LL up to the XX bit level
but not: Let's do the 600M range up to 80 bit! (no offense! that is fun too!)

It's the same we do now, but with a team and a goal...

[cheesehead]

Originally Posted by MrHappy

... with the overall goal of speeding up GIMPS in the present or near future...
like: no DC should be handed out without TF-attempts up to the 69 bit level.
or: same with LL up to the XX bit level

Since the default TF limits are already set to maximize GIMPS throughput if CPUs are used for the TF and LLs, are you referring to establishing a new set of GPU-TF default limits for the case of GPUs performing TF?

There could be one set of GPU-CPU-TF defaults for the case of GPU performing TF and CPU performing LLs, plus another set of GPU-GPU-TF defaults for the case of GPUs performing both TF and LLs. (I'd recommend a different, more readily-distinguishable set of designations than mine, however.)

(Yes, I know that the GPU/CPU LL speed ratio is not expected to be as large as the GPU/CPU TF speed ratio. Perhaps the GPU-GPU-TF defaults will be only partly or not at all different from the GPU-CPU-TF defaults, but until someone actually does the computations I'd rather they not be assumed to be the same.)

[MrHappy]

I don't want to change any server side values (such as default trial factoring depth) but would like to see

(first task)
- all exponents that are to be handed out as DCs for the rest of this year should be TFed to 68 bit at least

So I reserve DCs, TF them 1 bit (by GPU) and release them and do DCs with available CPUs.

Does anyone want to join or does someone have another (better) idea? Feel free to say so...

[Christenson]

ckdo has me TF'ing another bit on exponents before they go to DC. It's been successful, in that I've been finding factors, and doing so more quickly than I can complete LL-DCs. If you want to participate that way, PM to ckdo. There's lots more exponents than I can effectively TF; I only have about 80% of a GTX440 on standard GIMPS TF duty. (the other 20% is on operation Billiion Digits, and special requests, and the GTX480 does cudaLucas and Billion Digits, if I can ever get back to it to retrieve the results...)

Also, as I have done for daviddy in the past, anyone that thinks their LL exponent needs a few more bits of TF, let me know and I can put it in my work queue.

[Dubslow]

What ckdo is doing actually sounds like an excellent use of TF resources instead of extending the TF wavefront so far out in from of the LL wavefront. Any possible PrimeNet implementation?

[davieddy]

Originally Posted by MrHappy

No DC should be handed out without TF-attempts up to the 69 bit level.

ABSURD

and in so many ways.

TF is obviously best done before the first LL test, and to 4 bits
more than a DC (assuming the exponent is twice as big).

Endorsed by George and Jacob (to name but two), I've suggested that
a few GPUS could easily TF the exponents currently being dished out for
first time LL to 72 bits. Hardly any of them have been.

David

[davieddy]

Originally Posted by Christenson

Also, as I have done for davieddy in the past, anyone that thinks their LL exponent needs a few more bits of TF, let me know and I can put it in my work queue.

Yep. ETA for the next Mprime is 7th October

Not sure whether Eric or his ilk will have the patience for the
75 bits he did on my 47M exponent, but this simple sort of co-operation
enhances the feeling of a "community project".

David

[Christenson]

Originally Posted by Dubslow

What ckdo is doing actually sounds like an excellent use of TF resources instead of extending the TF wavefront so far out in from of the LL wavefront. Any possible PrimeNet implementation?

The TF wavefront is getting extended so far in front of the LL wavefront because one measure of TF progress, factors found, is being maximised. (Every new bitlevel doubles the effort required; TF is definitely a brute-force algorithm). Yes, TF'ing is going on on stuff that won't be LL'ed for awhile, but they will run out and come back and do another bit level.

As for the primenet implementation of mfaktc, that is on my desk and taking much longer than I thought it would, not because of the intrinsic difficulty of the project, but because of distractions like updating the worktodo file format, a job that pays $$, paying creditors, keeping house, and consulting for a First Robotics team.

In the meantime, I just run mfaktc and turn in the results manually once a day (or in this case, after a few days).

[davieddy]

Finding "small" factors of big Mersenne numbers is dirt cheap.

Why bother?

Your good mate,
David

[Christenson]

Originally Posted by davieddy

Finding "small" factors of big Mersenne numbers is dirt cheap.

Why bother?

Your good mate,
David

But double the price of dirt a few times, and suddenly it's no longer cheap...

[Dubslow]

Originally Posted by Christenson

The TF wavefront is getting extended so far in front of the LL wavefront because one measure of TF progress, factors found, is being maximised. (Every new bitlevel doubles the effort required; TF is definitely a brute-force algorithm). Yes, TF'ing is going on on stuff that won't be LL'ed for awhile, but they will run out and come back and do another bit level.

As for the primenet implementation of mfaktc, that is on my desk and taking much longer than I thought it would, not because of the intrinsic difficulty of the project, but because of distractions like updating the worktodo file format, a job that pays $$, paying creditors, keeping house, and consulting for a First Robotics team.

In the meantime, I just run mfaktc and turn in the results manually once a day (or in this case, after a few days).

Well, actually I meant PrimeNet assigning DC exponents TF assignments, not mfaktc. Also, my HS has a FIRST team :) (which I was sort of kind of not really on)