mfaktc: a CUDA program for Mersenne prefactoring

[TheJudger]

Originally Posted by S34960zz

Was this via cut-and-paste or by browsing to your results.txt file and allowing that file to upload?

File upload

[James Heinrich]

Originally Posted by TheJudger

Warning, don't upload huge result files with the manual result form. I've just tried to upload 1920 results from one file... seems to be too much for one request.

On the other hand, anyone is welcome to submit as many results as they want to mersenne-aries.sili.net

[Christenson]

Originally Posted by James Heinrich

On the other hand, anyone is welcome to submit as many results as they want to mersenne-aries.sili.net

Unfortuneately, it's *just* a cheap GPU...so a few per day is all I get.....going to have to build an new PC (or at least upgrade PS) before I can support a thousand at a time....

[James Heinrich]

Still, doesn't matter -- all stats are welcome, whether you have many or few, from mfaktc or Prime95, TF / P-1 / LL / ECM...

[Christenson]

Originally Posted by James Heinrich

Still, doesn't matter -- all stats are welcome, whether you have many or few, from mfaktc or Prime95, TF / P-1 / LL / ECM...

I know you mersenne hunters are taking all the help you can get...I've assigned some of my CPUs to P-1 recently, and have some new CPUs helping out too.

Is there any real difference to submitting my manually-assigned TF results through mersenn-aries as opposed to mersenne.org?

[James Heinrich]

Originally Posted by Christenson

Is there any real difference to submitting my manually-assigned TF results through mersenn-aries as opposed to mersenne.org?

YES! If you submit to mersenne.org, your assignment will be complete, PrimeNet will know about it, and you will have done useful work. mersenne-aries.sili.net simply displays pretty stats and graphs and such about your work done, it does NOT relay it to PrimeNet in any way. You must make sure all your completed work is submitted to PrimeNet (whether automatically, or manually). But you are then also welcome to contribute your results to my site. Caveat: Prime95 work is listed under your own username, but any mfaktc work is compiled under the username "mfaktc".

[TheJudger]

Hi Christenson,

Originally Posted by Christenson

Unfortuneately, it's *just* a cheap GPU...so a few per day is all I get.....going to have to build an new PC (or at least upgrade PS) before I can support a thousand at a time....

actually this was a test. That was the result of a couple of days of work.
A GTX 470 with 3 i7 CPU cores running at ~3.5GHz can take 9 exponents in the M54.xxx.xxx range from 2^68 to 2^69 per hour (~210-220 per day).

I would bet that a few per day on a cheap GPU is still more than a single core of a current highend CPU can do... And since you've a lot spare cycles on your system (high average wait) you might want to try to run Prime95 with low priority on the same CPU(-core) once I've released mfaktc 0.17.

Oliver

[nucleon]

Originally Posted by TheJudger

Warning, don't upload huge result files with the manual result form. I've just tried to upload 1920 results from one file... seems to be too much for one request.

Oliver

I think it maxed out for me at around 200-500 lines.

I got jack of manual submissions, and so I created a bash script (using lynx) to submit results automatically. (Max rate is 1 per 10secs so I avoid overwhelming the server).

-- Craig

[drh]

On my i5 dual-core and a 3100M GPU, my CPU is only running at 25%. I tried another instance of mfaktc, but my throughput was cut in half, so I stopped that and now I'm running a P-1, and my mfaktc throughput is exactly the same, so I'm leaving it like this until something better comes along.

[James Heinrich]

Originally Posted by drh

On my i5 dual-core and a 3100M GPU, my CPU is only running at 25%. I tried another instance of mfaktc, but my throughput was cut in half, so I stopped that and now I'm running a P-1, and my mfaktc throughput is exactly the same, so I'm leaving it like this until something better comes along.

25% reported CPU usage is 100% one of four virtual cores on your i5. This may drop once TheJudger releases v0.17 which has better CPU-idling code. You're very much GPU-limited; starting multiple mfaktc instances only helps if you're CPU-limited. Even on my 8800GT, which is roughly 6x faster (~30-35M/s), I'm still definitely GPU-limited (not CPU-limited), so one instance of mfaktc is plenty to get maximum throughput.

[drh]

Originally Posted by James Heinrich

25% reported CPU usage is 100% one of four virtual cores on your i5. This may drop once TheJudger releases v0.17 which has better CPU-idling code. You're very much GPU-limited; starting multiple mfaktc instances only helps if you're CPU-limited. Even on my 8800GT, which is roughly 6x faster (~30-35M/s), I'm still definitely GPU-limited (not CPU-limited), so one instance of mfaktc is plenty to get maximum throughput.

True, totally agree. I'm using mfaktc in the 76M range, taking them to 71 bits. Would it be more beneficial to take them to 72 bits? It takes 16 hrs for 70 to 71, would take 32 hrs for 71 to 72. I'll make that change if it helps.