20221113, 21:10  #34 
Jun 2003
The Computer
401 Posts 
Hi Ken, if necessary I can post my server specs again to get it qualification completed, or at least note that it is also an ECC system and all current P1 assignments are using ECC RAM.

20221113, 22:13  #35  
"TF79LL86GIMPS96gpu17"
Mar 2017
US midwest
17060_{8} Posts 
Quote:


20230302, 19:42  #36 
"TF79LL86GIMPS96gpu17"
Mar 2017
US midwest
1111000110000_{2} Posts 
Current TF & P1 status is shown at https://www.mersenne.ca/obd. Current indicated Level is 24.05.
3321928171 TF completed to 92 bits, 20220205 by mersenne.ca user johnny_jack, reserved for P1, stage 2 5% complete 3321928307 TF completed to 92 bits, 20220223 by kriesel; reserved for P1, stage 2 ~14% complete on an ECC ram system 3321928319 TF completed to 92 bits, 20220226 by kriesel; reserved for P1, stage 2 ~16% complete on an ECC ram system 3321928373 TF completed to 92 bits, 20230223 by kriesel; right to reserve P1 by kriesel extends to 20230323, and has NOT been ceded. (reserved anyway on the https://www.mersenne.ca/obd page by Rafael R for P1 ~20230302) 29 other exponents have been completed to 90 bits, and are available for reservation to a higher TF level. These are in the exponent range up to 3321929987, as part of an effort to go to ~30 exponents eventually remaining with no known factors after TF and P1 factoring completion to recommended levels and bounds, for eventual PRP/GEC/proof & CERT when hardware and software development advance sufficiently. Anyone with a sufficiently fast GPU, or fast CPU with at least 64 GiB ram, is welcome to help with TF or P1 respectively. (But please follow the rules!) At the moment: OBD TF completed to 92 bits, reserved for P1 and in progress in P1 stage 2: 3 OBD TF completed to 92 bits, by Kriesel 20230223, so restricted for reservation to kriesel until 20230323: 1 OBD TF completed to 92 bits, ready for reservation by anyone with qualified hardware for P1: 0 OBD TF completed to 91 bits: 0 OBD TF completed to 90 bits, reserved to 91 bits: 0 Systems with qualification(s) completed & posted for OBD P1: 3 (1 is unconditionally) Exponents completed thru stage 1 P1: 3 Exponents completed thru stage 2 P1: 0 kriesel is the only user that has yet posted system qualification results. (This status summary is specific to OBD Mersennes. See also Ernst Mayer's F33 P1 effort, mentioned in the Xeon Phi hardware thread.) Last fiddled with by kriesel on 20230302 at 19:44 
20230302, 22:58  #37 
"TF79LL86GIMPS96gpu17"
Mar 2017
US midwest
2^{4}×3×7×23 Posts 
OBD P1 coordination proposal revised
OBD P1 is now feasible. For now, Mlucas v20.x is the only known software capable of OBD Mersenne P1 factoring, with sufficiently large fft lengths and several months of QA testing and revision (patching) accomplished. The latest version is recommended, which is V20.1.1 patch 20220706. See http://www.mersenneforum.org/mayer/README.html and https://mersenneforum.org/showthread.php?t=27295
I propose the following:
Last fiddled with by kriesel on 20230302 at 23:18 
20230305, 17:54  #38  
"TF79LL86GIMPS96gpu17"
Mar 2017
US midwest
2^{4}×3×7×23 Posts 
Since the mersenne.ca pages have been changed, this is now called the target row.
Quote:
A proper P1 on it requires ~21663. GHD, which would have been comparable effort to 3/4 of the total mersenne.org P1 GHD credits for him in the preceding year. (Which I think were produced mostly by the more efficient new stage 2 code of mprime, while OBD P1 requires Mlucas for sufficiently large fft length.) OBD work produces no GHD credits on mersenne.org, whether performing TF or P1. Exponent status update attached below. Last fiddled with by kriesel on 20230305 at 18:00 

20230420, 19:26  #39  
Aug 2002
Buenos Aires, Argentina
2×761 Posts 
Quote:


20230420, 21:50  #40 
"TF79LL86GIMPS96gpu17"
Mar 2017
US midwest
2^{4}·3·7·23 Posts 
Why? Odds of a factor at the full indicated bounds after TF to 92 bits are only 3.98%: https://www.mersenne.ca/prob.php?exp...&b2=1000000000
Please, before your response, see https://www.mersenneforum.org/showpo...9&postcount=20, and R.D. Silverman's comments on the matter and the rest of that thread, and consider that for OBD P1, Mlucas v20.x will be required, which does periodic GCDs during stage 2, lowering stage 2 cost in the case of a stage 2 factor found. (Which I think improves further the case for full bounds from the start, compared to only a single GCD after completing stage 2.) We expect to do dozens of OBD P1 in the course of taking OBD to "level 28" or higher, by relatively few very determined participants, so what is statistically expected to be most efficient in the long run is relevant. (Footnotes: currently, there is no PrimeNet row or GPU72 row, only "target" row, which corresponds to the older GPU72; ECM as as a followup to prior P1 as suggested in https://mersenneforum.org/showpost.p...0&postcount=31 is not an option, since there is no implementation for OBD exponents.) Last fiddled with by kriesel on 20230420 at 22:17 
20230420, 22:55  #41  
Aug 2002
Buenos Aires, Argentina
2×761 Posts 
Quote:
So about 39% of the positive results you will find with P1 in OBD would have been found a lot faster by using lower bounds. For the other 61%, there is almost no difference on running both instances of P1. 

20230420, 23:49  #42 
"TF79LL86GIMPS96gpu17"
Mar 2017
US midwest
7728_{10} Posts 
So for the average case, and taking your 20:1 as a given without question, we have for your proposal, since for 1.54% of the exponents we run only the short small bounds and find a factor, while for the other 98.46% of exponents, we run both a smallbounds set and a fullbounds set, giving (ignoring the small chances of finding a factor in full bounds stage 1 that eluded the small bounds both stages, or finding a factor in smallbounds stage 1 that eluded trial factoring) a cost for P1 effort of roughly
.0154 * .05 + 0.9846 * (0.05 + 1) = 1.0346, instead of 1. * 1 = 1 for running full bounds the first time every time, giving on average, an estimate that lowbounds first costs a few percent MORE P1 effort than full bounds initially. Mlucas IIRC does not have B1 extension later from a saved file. Mlucas does periodic GCD during stage 2, so we may as well aim for the full B2, and if we get lucky and it finishes early with a factor found, great, some cycles saved. To prepare for levels 29 and 30 (PRP, and proof or DC, respectively), we'll TF ~3334 exponents to 92 bits, finding factors for 12 of them; then P1 to full bounds ~3132, to have 30 PRP candidates surviving the full complement of optimalcomputeeffortlevel factoring attempts. https://www.mersenne.ca/obd shows 33 at 90 bits TF or higher now. (1.03461) * 32 ~ 1.107 extra P1 effort from doing smallbounds P1 first. .0154*32 = 0.4928, so it's ~5050 we find NO factors with smallbounds P1. Last fiddled with by kriesel on 20230421 at 00:20 
20230421, 01:24  #43 
Aug 2002
Buenos Aires, Argentina
2×761 Posts 
Yes, you are right. It is better not use small P1, especially if the software performs GCD in the middle of a step.
Another way it helps using small p1 is if there are memory errors during a very long run. It is possible that doing P1 with large bounds skips the factors that could have been found by P1 with smaller bounds because (in our case) 95% of the errors would be introduced by the long P1 run. 
20230421, 01:55  #44 
"TF79LL86GIMPS96gpu17"
Mar 2017
US midwest
2^{4}·3·7·23 Posts 
Hence, why ECC ram, time limits on run lengths, and qualifying runtime scaling with refinding known factors are proposed. And better software error detection may be added to P1 in various software titles later;
see https://www.mersenneforum.org/showpo...37&postcount=3 and from the Announcements page, "There's a paper describing a generalization of GEC and VDF, which appears to allow adding the same excellent error detection and verification now available for PRP, to P1 stage 1 computations. Discussion starts here in https://mersenneforum.org/showthread.php?t=24654." Last fiddled with by kriesel on 20230421 at 02:01 
Thread Tools  
Similar Threads  
Thread  Thread Starter  Forum  Replies  Last Post 
A couple of 15e candidates  fivemack  NFS@Home  1  20141130 07:52 
How to calculate FFT lengths of candidates  pepi37  Riesel Prime Search  8  20140417 20:51 
No available candidates on server  japelprime  Prime Sierpinski Project  2  20111228 07:38 
Adding New Candidates  wblipp  Operation Billion Digits  6  20110410 17:45 
new candidates for M...46 and M48  cochet  Miscellaneous Math  4  20081024 14:33 