Treading LL P-1 for 100M TF

[Aramis Wyler]

Originally Posted by Aramis Wyler

Progress so far:

That third LMH finished to 79 w/o a factor.

[kladner]

Originally Posted by kladner

80-81 progress:

Apr 12 08:47 |  540  11.8% | 183.58   1d19h |    361.14    82485    n.a.%

@ 61% w/no factor. Should finish early evening (UTC-06:00) 4/14/13.

Out of curiosity, will this exponent be retained by GPU72 for P-1?

[chalsall]

Originally Posted by kladner

Out of curiosity, will this exponent be retained by GPU72 for P-1?

No -- currently everything taken above 74 in the 332M range is returned to Primenet.

P-1 up there would take a very long time! Let's at least wait until there's a GPU P-1 solution before coordinating that work.

[kladner]

Originally Posted by chalsall

No -- currently everything taken above 74 in the 332M range is returned to Primenet.

P-1 up there would take a very long time! Let's at least wait until there's a GPU P-1 solution before coordinating that work.

I was not at all sure of the magnitude of the task.

On the example cited above, the first P-1 completed about 4-1/2 months after TF to 79.

History no factor for M332250257 from 2^78 to 2^79 [mfaktc 0.18 barrett79_mul32] by "James Heinrich" on 2011-12-20 
History B1=2800000, B2=47600000 by "Sergey Nosov" on 2012-04-05

From this I surmise that Sergey did not pick it up immediately after James finished. I say this because the second P-1 at higher bounds finished 18 days after TF 83-84. Since Sergey did both, I assume that he proceded immediately from one to the next.

History no factor for M332250257 from 2^78 to 2^79 [mfaktc 0.18 barrett79_mul32] by "James Heinrich" on 2011-12-20 
History B1=2800000, B2=47600000 by "Sergey Nosov" on 2012-04-05

If I am viewing this correctly (a dicey proposition!), does the sequence suggest something in the 2-3 week vicinity for P-1 in this range (with similar CPU power)?

[chalsall]

Originally Posted by kladner

If I am viewing this correctly, does the sequence suggest something in the 2-3 week vicinity for P-1 in this range (with similar CPU power)?

I can't speak immediately to this question (just heading to beddies), but these links should give you an idea. The first is a 60M P-1 I just finished using 3GB of RAM...

P-1 M60899011 (B1=575000, B2=11500000) -- 3.76954387 GHz-Days
P-1 M332250257 (B1=2800000, B2=47600000) -- 113.03810000 GHz-Days
P-1 M332250257 (B1=5000000, B2=139855708) -- 266.31606386 GHz-Days

[kladner]

Originally Posted by chalsall

I can't speak immediately to this question (just heading to beddies), but these links should give you an idea. The first is a 60M P-1 I just finished using 3GB of RAM...

P-1 M60899011 (B1=575000, B2=11500000) -- 3.76954387 GHz-Days
P-1 M332250257 (B1=2800000, B2=47600000) -- 113.03810000 GHz-Days
P-1 M332250257 (B1=5000000, B2=139855708) -- 266.31606386 GHz-Days

First off, thanks for pointing out resources which I could have accessed to get a decent approximation of an answer. Those resources say that the two P-1's respectively took 30 and 70 times as long as a 60M P-1 (if done on the same hardware.)

I also realize now that there was nothing to keep Sergey from running the higher-bounded P-1 concurrently with his GPU TF work. This adds unknown variables to the mix, but knocks down the idea that he necessarily had a phenomenal amount of CPU power to throw at the project.

In addition, I have now noticed that he ran different stages of TF in parallel, so all bets are off. That TF work took him ~43 days to do 80-81 bits instead of the ~2 days it is taking me on a similar exponent. This would seem to make it even more unlikely (though not impossible) that he ran the higher-bounded P-1 in ~19 days. Source for above:
http://www.mersenne.org/report_expon...p_lo=332250257

Thanks again for suggesting ways to form better conjectures about the situation.

Finally- Dang it! I pasted the same data twice in the previous post.
Intended:

History no factor for M332250257 from 2^83 to 2^84 [mfaktc 0.19 barrett92_mul32] by "Sergey Nosov" on 2012-12-28 
History B1=5000000, B2=139855708 by "Sergey Nosov" on 2013-01-16

[kladner]

! 26 minutes to go on 332M to 81 bits!


It will be good to get back to the more familiar timings of the low 60Ms.

[kladner]

In addition, I have now noticed that he ran different stages of TF in parallel, so all bets are off.

In looking at some of my own results, the above is a misinterpretation. For the first time in a long time I went a few days without reporting results. When I dumped all the results, several bit levels appeared as happening on the same day, which I very well knew was not the case. Of course, the recorded times only reflect the date of reporting. I made silly assumptions about what I was seeing.

[Aramis Wyler]

Whenever I do the big ones, I throw the top bitlevel (78-79 say) on one card and everything else (74-78) on another. They take about the same amount of time, the top level vs everything before it. I have yet to find a good way to divide the work over more than 2 cards though, because the last step is always going to take as long as everything before it, so 2 cards or 10 cards, it will get done in the amount of time it takes to do the last step on 1 card.

[TObject]

You can also split a single level among multiple cards by manipulating the checkpoint files. The checkpoint files are protected by the checksum, but the code to generate the checksum is available with the rest of the mfaktc source code.

[kladner]

Originally Posted by Aramis Wyler

Whenever I do the big ones, I throw the top bitlevel (78-79 say) on one card and everything else (74-78) on another. They take about the same amount of time, the top level vs everything before it. I have yet to find a good way to divide the work over more than 2 cards though, because the last step is always going to take as long as everything before it, so 2 cards or 10 cards, it will get done in the amount of time it takes to do the last step on 1 card.

Thanks for telling that. It is good to know that the idea was not too outrageous. Still, I don't see a way to tell the difference between levels run in parallel which finish close together, and ones just jammed together by longer reporting periods.