TFing 100M numbers
 

Sorry if this has been answered, a search didn't turn up info.

When a 100M number is selected, why are there three successive jobs TFing to 2^75, 2^76, and 2^77 on only one core? It seems like a lot of wasted time for what mprime or p95 reports as only having a 6% chance of finding a factor for.

I assume that you are talking about 100M digit numbers, those with exponents 332,192,831 and above.

In order to save the most work possible (and not do LL tests for numbers that we can factor), we want to eliminate numbers with 'easy to find' factors. Doing factoring from 29 -> 65 bits is fairly quick on modern equipment and so is very much worth doing. Doing TF from 65 -> 75 is takes longer, but is still worth doing. Above 75 may seem like it takes a long time (75 -> 77 is about 69 GHz-days of work). But when compared with the LL test (~4,940 GHz-days of work), it is still a good bet (4,941 * 0.06 = 296, not 69). The same thing with P-1 tests, the idea is to prevent the need for the LL test, which is more work.

The point where it makes sense to stop TF and start a LL test is encoded in Prime95.
If we use GPU's to do trial factoring, that point is 2 or more bits higher than with a CPU.

There have been several cases where people have skipped some of the TF and P-1 on these big numbers, then I have come along and found factors for them. The person could have saved a lot of time if they had done the TF.

Why do them on a single core? Depending on the set-up on your machine that may be also how your LL test will run (on a single core.) It does not make much sense to start 76->77 before 75->76 is done (unless you are in an awful rush).

I'm not doing them on a single core. The TFing runs on a single core regardless of how many threads I give it, which is a bit of a waste, especially with those 8+ core xeons I've got.

[QUOTE=aurashift;399242]I'm not doing them on a single core. The TFing runs on a single core regardless of how many threads I give it, which is a bit of a waste, especially with those 8+ core xeons I've got.[/QUOTE]

You're TF'ing on a CPU?

Why bother?

[QUOTE=chalsall;399249]You're TF'ing on a CPU?

Why bother?[/QUOTE]

Because PrimeNet assigned him LLs that were insufficiently TFd.

OP, I'd recommend posting in one of the GIMPS forums asking for GPUs to TF your exponents. You can run P-1 in the meantime, which Prime95 will multithread, and when the TFers report back (and P-1 completes) then start the LL.

+1 for Dubslow.

I was a bit in hurry in the morning because I had to go to the immigration dept to "punch in" (Thai laws: farangs have to go there 4 times a year to say they are still here and stay at the same address or moved around, the officers look to you, smile, and stamp your paper, see you again in 3 months, but the queues are sometimes very long and you have to be early in the morning).

What I wanted to say by my short post:

I am following aurashift since he/she started talking about millions of cores here around :smile:.
This user jumped over 1000 positions (one thousand) in [URL="http://www.mersenne.org/report_top_500_custom/?team_flag=0&type=1003&rank_lo=1&rank_hi=100&start_date=2000-01-01&end_date="]lifetime top[/URL] in the last ~70 days, see his "stars" at position ~60. He is big iron!
We should "[URL="http://hallo.ro/search.do?d=en&l=ro&type=both&query=a+face+pe+dracu-n+patru"]split the devil in four[/URL]", as we Romanians say (or was is "[URL="http://hallo.ro/search.do?d=en&l=ro&type=both&query=a+face+luntre+si+punte"]to make both a boat and a bridge[/URL]"? hehe) to keep him in the project. Give him "well done" TF-ed expos for his LL cannons. Make lists of exponents assigned to him and put them out for GPU TF, etc.

@op: Hey buddy, don't waste your time and CPU resources doing TF. Let the GPUs to do that, they are 200 times faster than the CPU. One "third world" GPU can easily supply many of your cores with well TF-ed exponents. You or someone else can make a list with exponents you got which are not TF-ed enough, and post it here. We will work through it with our GPUs, as the posters above said already.

well, right now I've only got two servers and my laptop doing 100M digit numbers because of the trial factoring thing. The servers are already done with TF, and my laptop is about 70% done so no big deal.

I had another thought though. Since computing power inevitably increases across the board over time, why not hand numbers out on a first come, first served basis? I could crunch through smaller numbers quicker with these AVX cores, and doing LL on a P3 800mhz is a bit of a waste of electricity anyway.

(p.s, a ton of my 12 core blades are being repurposed, probably this month. Maybe half of my computer power on here. I'll try to get the assignments done before that, though)

I feel like I might be mis-understanding. Assignments ARE handed out on a first-come first-served basis with a small buffer at the front reserved for faster computers. Unless you're talking specifically about 100 million digits which doesn't have very many people anyway.

But yeah for sure if you're planning on continuing with the big LL tests post whatever exponents you want to work on here and I'm sure some people will give you a hand TF'ing them for you with GPU's. I for one would be happy to help out.

Eh the last server worker thread is just about done(have four going, two with 13 cores and two with 10), but my laptop is still churning M332421413 on 2^76. (should I pause it?)

I've been getting mostly assignments in the M70,000,000+ range (on the xeon e5v2's w/ AVX) with a few of my slower boxes getting smaller exponents.

Pause that. Is a minuet job for a GPU, I will schedule it right now on one of mine.

Edit: scheduled 75 to 78, ETA ~8 hours.