COMPLETE!!!! Thinking out loud about getting under 20M unfactored exponents

[LaurV]

Yay! Nice one that 4M7! We found 4 factors last night and closed the range (overshot by one).

In fact, you know, for a long while now, we only found factors during the night, so, paraphrasing that politician we had, who, reading that in a train crash the last car is the most destroyed, wanted to pass a law to remove the last car from all the trains , we should also switch to do P-1 only during the night, hehe... During the day the factors stay hidden, you can't find them...

Anyhow... We have now 190 ranges left, due to good people clearing some other ranges too. Time for a new party, hehe.

I moved to 11.3M indeed, but my contribution will decrease for the coming weeks/months, etc., due to two reasons. One is increasing the ETH price, which caused moving some of my local toys to mining. The second is Gugu releasing "Colab Pro +", in fact just some trickery to get more money (i.e. 5 times more!) from the paying subscribers, for offering the same hardware. Before, a "free" tier was getting the half k80, or full P4, T4, and occasionally P100, and they use to interrupt every 4 to 8 hours, while the "Pro" would get mainly V100 and P100, with occasional T4, and they would get disconnected every 24 hours. Right now, they added "Pro +" which costs 5 times more, it is kept running for about 48 hours, and it sucks all the V100 cards (and occasional, you can get A100 which is better, or P100 which is worse - for records, the P-1 performance for a V100 is double than a P100, and the TF performance is about 4 times a P100, and about 2.5 times a T4), so, now the "Pro" subscriber mostly gets T4 for 24 hours, or rarely P100 which lasts about 8-12 hours. This also depends on your past activity (more activity in the past -> lower cards, because the users who didn't get them in the past have priority).

All in all, my P-1 output decreased to about a quarter of what I was doing. Today I only got T4, and I switched to Chris' TF (T4 gets 1800GHzDays/Day TF, as opposite to P100 which only gets about 1200, while for P-1 the T4 is few times slower than P100). So, I am finding TF factors in 9.xM (or whatever Chris serves me, I already found some).

[firejuggler]

23.8M is done. last stretch was done by TF to 72 or 73. Many thanks to whoever that was.

[lycorn]

Originally Posted by LaurV

The second is Gugu releasing "Colab Pro +", (...), so, now the "Pro" subscriber mostly gets T4 for 24 hours, or rarely P100 which lasts about 8-12 hours.

Ok, I reckon that´s why my free tiers are only getting K80s lately . Still an infinite benefit vs cost ratio, tho .

[chalsall]

Originally Posted by axn

Yes. Perhaps chalsall and rebirther can coordinate and push that range to the top of the queue?!

Super busy... But quickly...

1. Reb and his BOINC crew like to work breadth-first.

1.1. Completing (most of) everything "of interest" to 71 is about three to four weeks out.


2. Not quite sure for the reason for the fixation on 11.xM.

2.1. 6.xM is almost entirely TF'ed to 71 "bits".

2.2. Same with 8.xM. Many candidates in this 1M range with a B2<2.5M just asking to be worked by a P-1'er.


Please remember that TF is very expensive at the low ranges, while P-1 becomes reasonably cheap (even with extreme bounds).

Thanks for all the interest everyone. It's cool dropping in a few times a day, to see what's going on.

[firejuggler]

Is there another range I can explore? (PM1- wise)


edit : 11M?

[masser]

Originally Posted by firejuggler

Is there another range I can explore? (PM1- wise)


edit : 11M?

6.2M?

[petrw1]

Originally Posted by masser

6.2M?

Any 6.xM would be good choices.
They need good sized B1/B2 but not outrageous.
if you give me a day I can recommend some.

[firejuggler]

Sure thing, i can wait a day or two.

[lycorn]

This morning (UTC+1) the last 1M range (15M) of the Primenet search space (0 - 999 M ) was cleared to less than 21,000 exponents to factor. It´s certainly a significant milestone on the way to achieve the original goal stated in the title of this thread. There were just a few dozen left and the SRBASE team gave as usual a great push. The last one was cleared 3 hours ago by a much smaller team of only 3: me, myself and I .
There are only 20,504,150 to go. We are nearly there. Ah, ah, ah....

[petrw1]

Originally Posted by lycorn

This morning (UTC+1) the last 1M range (15M) of the Primenet search space (0 - 999 M ) was cleared to less than 21,000 exponents to factor. It´s certainly a significant milestone on the way to achieve the original goal stated in the title of this thread. There were just a few dozen left and the SRBASE team gave as usual a great push. The last one was cleared 3 hours ago by a much smaller team of only 3: me, myself and I .
There are only 20,504,150 to go. We are nearly there. Ah, ah, ah....

Good news!
But for this effort there are only 504,151 to go.
Better yet this team only needs 16,290.
The rest will eventually be found by the mainstream.
The other great news is that we should now finish our side first.
When I started this it was just a whimsical hope.

[petrw1]

Originally Posted by petrw1

Any 6.xM would be good choices.
They need good sized B1/B2 but not outrageous.
if you give me a day I can recommend some.

If you really don't want to read all my technical blah blah below just do this:
1. https://www.mersenne.org/report_fact...99&tftobits=72
2. Sort by lowest B1 then B2
3. P1 starting from the lowest B1/B2 using new bounds of about: 3M/90M
4. Each run should take about 2.0GhzDays with about a 6.64% (net 2.73%) success rate.
5. Statistically you'll need to do about 1600 exponents to get 45 factors.

Thanks and good luck.

For an in-depth discussion you can use yourself for future ranges read on:

In a nutshell this is the process I use ... I'm open to anyone who cares to simplify this:
I'll use 6.2 in this discussion below.
1. How many factors are remaining to be found: 45
--- This: https://www.gpu72.com/account/twok/
--- Or: https://www.mersenne.ca/status/tf/0/0/5/620

2. Do I want to use TF or P1 or a little of each?
--- Depends on what hardware you have: CPUs: P1; GPUs TF or maybe P1 with certain GPUs
--- Depends on how deep TF is vs how deep the P1 currently is and which is more "efficient" considering expected factors per GhzDay or per Day for your gear.
--- You asked for P1 work so lets assume for now you hope to find all 45 factors with P1.
--- Because I am CPU heavy I start with this assumption but after analyzing as described below I may have to concede P1 can't effectively find them all and may have to TF another bit later.

3. What success rate do I need from the exponents to get 45 factors?
--- At the highest level 2,044 exponents for 45 factors is a 2.2% success rate.
--- So you could use process all 2,044 exponents with bounds 2.2% above current.
--- That is in most cases NOT the most efficient.
--- Another possibility is taking the 1,000 best candidates up 4.5%.
--- It all depends on what the current bounds/success rate look like; more-so what percentage are currently poorly P1'd vs aggessively P1'd. Focus on the poor only. (Hmm do I sound generous?)
--- Point 4 helps you make an educated guess.

4. Get a sense for how much P1 has already been done.
--- This range is easier to calculate than most since there are very few current P1 bounds.
--- Here: https://www.mersenne.org/report_fact...99&tftobits=72 show the current P1 efforts.
--- Eyeball some of the more typical B1/B2 (maybe sort by those columns first)
--- There are a 160 of these: 385000/9528750
--- Plug them here to get a sense for the current factor-found percentage: https://www.mersenne.ca/prob.php?exp...000&b2=9528750 = 3.51%
--- Similarly 1,397 of 600000/12000000 = 3.90%. These are most of them.
--- Similarly 463 of 700000/15000000 = 4.16%

5. So here are some scenarios.
--- You could choose to find the bound for each of the above 3 sets that are 2.2% higher than current
--- Here using the entry fields on the right side:
--- https://www.mersenne.ca/prob.php?exp...s=71&prob=5.71
--- The yellow 30X row shows 5.71%(3.51+2.2) requires bounds of 1,672,790/50,183,700 with 1.109040 GhzDays per run.
--- 160 exps x 2.2% = 3.5 Factors. 160 runs x 1.10904 = 177.5 GhzDays = 50.7 GhzDays per factor
--- Similarly for 6.10%: 2,148,050/64,441,500 with 1.424132 GhzDays per.
--- 6.36%: 2,285,915/80,007,025 with 1.675555 GhzDays per.

6. IMHO because this range is so heavily loaded with current runs of 600000/12000000 I would use the 6.1% bounds padded up slightly because I don't want bad luck to cut me short ... but start with the 160 better candidates first hoping for an extra factor here.
--- Alternatively is you are not averse to a little bad luck and having to TF another bit for part of this range then you could choose to not pad the bounds.