20210926, 14:39  #56 
"University student"
May 2021
Beijing, China
5^{3} Posts 
You can try NFS M1217/141065889786325378120502378437521368981420168194011324498831469066398800377270303571726687433864860259570077205499951967 which is much smaller than M1061.
Last fiddled with by Zhangrc on 20210926 at 14:43 
20210926, 14:56  #57 
Apr 2020
543_{10} Posts 
The difficulty of SNFS is determined by the size of the original number, not the composite cofactor, since it's the original number that determines the polynomial. M1217 has 367 digits, so it has SNFS difficulty 367, not 248 (the size of the cofactor). This would be a record size for an SNFS job: I believe the current record is M1193 which had difficulty 360.
(edit: M1217 might be faster by GNFS on the 248digit cofactor; the record is 250 digits so this is technically within reach if you have thousands of cores, but still far more difficult than M1061, which is comparable to 218 or 219 digit GNFS.) Last fiddled with by charybdis on 20210926 at 15:02 
20210926, 16:00  #58 
Bamboozled!
"πΊππ·π·π"
May 2003
Down not across
2^{4}×13×53 Posts 

20210926, 16:24  #59 
"6800 descendent"
Feb 2005
Colorado
1010101010_{2} Posts 

20211103, 22:47  #60  
"Seth"
Apr 2019
17E_{16} Posts 
Quote:
Using Merten's the calculation seems to simplify to 165/80 = 18.75% Using the Dickman function (Code) and asking what's the odds the 2nd largest factor is between x^.169 and x^.20811 (and that factor is not < x^.169) I also get 18.5% Looking at M1481 where the t65 could conceivable be increased to t70. Merten's suggests a 7.1% chance of finding a factor. Dickman suggests a greater 10.1% (as you exit the linear phase and start approaching sqrt(N)) Looking at M1217 where the t65 could also be increased to t70 (assuming Ryan hasn't already) Merten's suggests a 7.1% chance of finding a factor. Dickman suggests 18.2%! t70 would be roughly 0.4 gpu years of work (and $90 of power) or $500/factor t75 would be roughly 2.1 gpu years of work (and $450 of power) or $1400/factor with 34.8% of finding a factor. t75 would would be a fair bit past optimal (0.28 * digits) but more reasonable for a dedicated individual than factoring a C248. 

20211103, 23:11  #61  
Apr 2020
3×181 Posts 
Quote:
Also 0.28 is very low. Using a more normal figure of 0.31 would suggest running t77 on a c248. With GPU stage 1 the optimal amount of ECM may be even higher. [this number is likely a close call between GNFS and SNFS but that doesn't affect the amount of ECM it needs] 

20211103, 23:34  #62  
"Seth"
Apr 2019
2·191 Posts 
Quote:
I didn't take CPU cost into account for $ (it roughly doubles the cost) but I did take it into account for time. I used some numbers from a previous project. A single test curve at B2=105e12 took 1807 seconds while the B1 curves are estimated at ~116 seconds/curve. 1807/116 = 15.6 this would be fine if I had ~24 cores (my 12x cores giving ~9x single core perf). Instead I'd use B2=105e11 which takes 663 seconds and compensate by doing ~20% extra curves. This would give me a more comfortable 6/1 CPU:GPU ratio which I could satisfy with 810 cores. The hard part is that my system is sluggish/nonresponsive when I'm running ecmgpu I think that if I add 10% wait time that goes away (meaning I can still use my computer as my primary driver). 

20211104, 00:29  #63  
Apr 2020
543_{10} Posts 
Quote:
At the t75 level the high RAM usage may be an obstacle to running lots of stage 2 threads on a single machine. You can use the maxmem flag to limit it, but stage 2 will take longer as a result. 

20211104, 02:50  #64 
"Seth"
Apr 2019
2×191 Posts 
That's a great point about maxmem.
I updated my optimizer. New hypothetical numbers for a t70 on the C248 remainder of M1217: 147K curves at B1=3e9 B2=6.4e12 Stage 1 (on GPU) takes 120 seconds/curve and Stage 2 (on CPU) takes 285 seconds/curve Peak memory was reported as 3244MB The GPU would complete a batch every 2.5 days which is a little more than I'd like. 204 GPUdays + 484 CPUcoredays 111K curves at B1=4.3e9 B2=12.9e12 GPU 165s / curve, CPU 370 s / curve Peak memory was reported as 3249MB 217 GPUdays + 475 CPUcoredays 260K curves at B1=1.2e9 B2=22.4e12 GPU 48s / curve, CPU 474s /curve Peak memory was reported as 6596MB (so I might run into trouble with running more than a few instances) GPU batches would complete every day which is nice. 144 days GPUdays + 2100 CPUcoredays The 3rd cost slightly more (10 CPUcoredays = 1 GPU day) but finishes faster. Sadly the long batches from the first option mean I probably can't actually do this (I may try one batch just to validate numbers). 
20211104, 04:38  #65  
"Curtis"
Feb 2005
Riverside, CA
5×1,013 Posts 
Quote:
The 0.31 and 0.21 numbers for GNFS and SNFS are empirical results to the query "how much ECM should I do such that the time per expected factor of the last block of ECM is equal to the time to run NFS?" When I started paying attention in 2010ish, 1/3 and 2/9 were widely regarded as reasonable values with an observation that as jobs get bigger the ratios should drop a little. Then NFS kept getting faster, so the correct ratios dropped particularly GNFS, as poly select has taken 2030% off of sieve time compared to ten years ago. GPUecm shifts the ratio back upward a bit, since it speeds stage 1 so much. The bayesian ECM tool is a moreaccurate way to determine "optimal" ECM, but that tool also makes assumptions about how long NFS takes. Lucky for us, it barely matters in efficiency terms whether we do 2/3 of optimal ECM or 1.5x optimal ECM... so lots of us still use the ratios as "good enough". Last fiddled with by VBCurtis on 20211104 at 04:40 

20211104, 05:29  #66  
"Seth"
Apr 2019
2×191 Posts 
Quote:
I believe your sum is the number of expected factors (e.g. you expect to find 0.206 more factors by completing a t80), 0.175 for the first factor + 0.175^2 for finding two factors + ... this makes a little less sense when we know the number is a product of exactly 2, 3, or 4 primes but meh. Last fiddled with by SethTro on 20211104 at 05:31 

Thread Tools  
Similar Threads  
Thread  Thread Starter  Forum  Replies  Last Post 
Do you know this method to factorize?  Godzilla  Miscellaneous Math  28  20171031 18:14 
mathematica7.0 can easily factorize 10^67+1111  aaa120  Factoring  14  20081207 13:14 