[QUOTE=chalsall;362144]What are your orders, Admiral?[/QUOTE]

I don't give orders! I'm happy to help work out our best strategy.

It seems that at any point in time we want to TF the exponent that gives the biggest payoff. So let's define the payoff function:

profit = chance-of-finding-a-factor * LL cost / TF-cost

TF-cost is proportional to 2^TF-bit-level / exponent
chance-of-factor is 1/TF-bit-level
LL-cost is exponent * per-iteration-time

I think we can (mostly) ignore P-1 since we will end up doing P-1 on all LL candidates. Putting it together we want to maximize exponent^2 * per-iteration-time / 2^bit-level * TF-bit-level.

Back to our 58M / 64M example. We need a table from James detailing the per-iteration times for the most common GPUs in play. I guess that would be a 580, but the actual choice probably makes little difference.

Say the per-iteration time for 58M and 64M is 10ms and 12ms respectively. Then GPU72 should hand out 58M since 58M^2*10/2^73*73 > 64M^2*12/2^74*74.


One wrinkle: You don't want the GPU72 server to get too far ahead of the LL wavefront. If you added the 70M to 71M range to the server today, we would immediately start factoring that range to 2^72 as those exponents would have the maximum payoff. I'm guessing the server should be adding the same number of exponents to the pool each day as the Primenet server hands out for LL testing.

Two small observations again:

Subjective one: The "profit" name is already biased :P Your function formula already suggests that is more profitable to do LL (numerator) than TF (denominator), without any calculus, hehe.

Objective one: for LL, the time is "1x" when DC, and "2x" when LL. Also, ignoring P-1 is not right, because in the DC range, P-1 is already done, and it does not have to be done again. The 10/12 timing is "off" from the real values, but it is proportionally right (about) and we are only interested in proportion here. I only talk for GTX580 (the 7970/7990 are much better doing TF for another 2 bits more :razz:, this does not mean that they have to do it - it might not be optimally in context: other people will be able to clear the exponents faster).

Generally, we all agree here, that going to 74 is a bit "costly", expecially since P-1 became a cuda-enabled stuff. One will be much better in clear a range by doing P-1 in it, in any situations.

My argument is only "74 against 73 bits in the actual LL front range".

For high DC range (55M+), we both agree that is is more profitable to raise them to 73, doing two of them in about the same time it would take to do a 64M+ to 74. What we don't agree is the numbers for the "amount of profit", you say two times (i.e. ~100%more), I say 30% more, due to the P-1 (already) done into the range. Of course, the "profit" gets higher as "55M" gets higher (to 58M) and 64M gets lower, it may get almost doubled in the middle, but it will never be double, even if you have the exponents in the same range, because those counted as "DC" had already been "sieved" to some extent by the P-1, and those counted as "LL" are fresh, unsieved.

Edit: the "wrinkle" is common sense, we all agree with that. Exhausting one type of work "because we can", it would bankrupt many other participants which can [U]only[/U] do that type of work, and it is not nice, neither wanted.

The wrinkle can be accounted for easily enough by multiplying the result by a sigmoid function, with a percent based on the distance from the wave front, where 100% would be the next available candidate in the wavefront or less (because numbers can appear behind the wavefront as they are returned incomplete) and 0% would be some distance out in front of the wave front, say 2 months worth of work.

[QUOTE=Prime95;362178]I don't give orders! I'm happy to help work out our best strategy.[/QUOTE]

That was a joke, George. :smile:

And one final note. Never give out a TF assignment for a bit depth greater than recommended by James' TF vs. LL break-even chart.

Thus, with unlimited GPU firepower we'd factor everything up to James' limits. With less than unlimited firepower, we'd always do the TF with the biggest payoff.

Heck, even davieddy might agree with such a strategy.

In practice, will implementing such a strategy make much of an impact? Are there significant numbers of exponents in the 55M-64M range factored to only 2^72 being released by PrimeNet each night?

[QUOTE=chalsall;362144]What are your orders, Admiral?

I'm serious.[/QUOTE]

[QUOTE=chalsall;362196]That was a joke, George. :smile:[/QUOTE]

:ermm:

[QUOTE=TheMawn;362221]:ermm:[/QUOTE]

:jokedrum:

[QUOTE=chalsall;362223]:jokedrum:[/QUOTE]

I'll see your :jokedrum: and raise you one :weirdo:

[QUOTE=TheMawn;362226]I'll see your :jokedrum: and raise you one :weirdo:[/QUOTE]

I'm comfortable with that.

"Normal is not something to aspire to, it is something to get away from." -- Jodie Foster.

[QUOTE=Prime95;362208]And one final note. Never give out a TF assignment for a bit depth greater than recommended by James' TF vs. LL break-even chart.[/QUOTE]

The [URL="https://www.gpu72.com/account/getassignments/lltf/"]web-interfaces[/URL] for TFing assignments will pop-up a warning if the user "pledges" to anything higher than 75. But, if the user "says" they are sure they want to go higher, the system will give the assignment.

As always, the participant should be allowed to do anything they want to so long as it doesn't harm the project or others, even if it Doesn't Make Sense [SUP](TM)[/SUP].

[QUOTE=Prime95;362208]In practice, will implementing such a strategy make much of an impact? Are there significant numbers of exponents in the 55M-64M range factored to only 2^72 being released by PrimeNet each night?[/QUOTE]

A [URL="https://www.gpu72.com/reports/current_level/"]bit less than 11,000 LL candidates[/URL] below 57M are only TFed to 72 (read: most of them). With the average "abandonment rate" of ~ 80%, we could process most of them over time. Although, I imagine this rate will be lower because of the "preferred" status of the very lowest range.

I've started the configuration changes needed for this. However, unfortunately, my main monitor died on me this morning, so I'm a bit disabled at the moment (how do people work with only one monitor these days?!?!?). I should have a replacement tonight.

[QUOTE=chalsall;362289]I've started the configuration changes needed for this. However, unfortunately, my main monitor died on me this morning, so I'm a bit disabled at the moment (how do people work with only one monitor these days?!?!?). I should have a replacement tonight.[/QUOTE]

Well I personally only use the one monitor, but it has 27 juicy inches of 1080p goodness. Still I find myself badly wanting a second.