TF Limits to Release At

[Dubslow]

Can I request that we start releasing exponents at lower bounds?
At mersenne.org/various/math.php, it gives the following bound:

47450000 2^68

meaning that exponents less than 4745... only be factored to 68 bits. Now, before GPU272 went up, P95 had increased the limits two bits on the PN assignment page. As it is, I think 72 bits for 45M exponents is excessive, being four bits above various/math and I could make a decent argument that even at 50M 72 bits is excessive. Thus I propose that GPU272, instead of aiming for 72 across the board, aims for three bits above the levels posted at mersenne.org/various/math.php . That would mean 72 bits between 47M and 58M, and 71 bits below that, which is more reasonable than what we've been doing. It would also increase our throughput, I think, specifically with an immediate gap.

[diamonddave]

Originally Posted by Dubslow

Can I request that we start releasing exponents at lower bounds?
At mersenne.org/various/math.php, it gives the following bound:
meaning that exponents less than 4745... only be factored to 68 bits. Now, before GPU272 went up, P95 had increased the limits two bits on the PN assignment page. As it is, I think 72 bits for 45M exponents is excessive, being four bits above various/math and I could make a decent argument that even at 50M 72 bits is excessive. Thus I propose that GPU272, instead of aiming for 72 across the board, aims for three bits above the levels posted at mersenne.org/various/math.php . That would mean 72 bits between 47M and 58M, and 71 bits below that, which is more reasonable than what we've been doing. It would also increase our throughput, I think, specifically with an immediate gap.

While I agree that we should release the low bound early until we catch up with the wavefront. I'm not convinced with your assessment that factoring to 72 is excessive at 50M.

[Dubslow]

Well, while I did say that, my final proposal did include 50M to 72 bits, more specifically everything above 47450000 to 72 bits.

[diamonddave]

Originally Posted by Dubslow

Well, while I did say that, my final proposal did include 50M to 72 bits, more specifically everything above 47450000 to 72 bits.

I was talking about this point:

Originally Posted by Dubslow

I could make a decent argument that even at 50M 72 bits is excessive

[Dubslow]

It takes me about 3h30 with a GTX 460 to go 69-72.

[Dubslow]

This same argument also implies that exponents above 58,520,000 should be taken to 73 bits.

[davieddy]

Originally Posted by Dubslow

This same argument also implies that exponents above 58,520,000 should be taken to 73 bits.

And before LLs above that are assigned for the first time.

Unfortunately there is FAT CHANCE of that happening, because of the chaotic manner in which TF by GPU is being "organized".

A year ago the wavefront was at 53.5M. I argued that it would be comfortably within the reach of available firepower to ensure that all exponents above 54M could be GPUed to 72 before ever being assigned.
What has happened? Only about 10% of them have been.

Now suppose 54M-58M had been TFed to 72, and smaller "grabbed" exponents to say 70. Chalsall's spider would now be redundant (few expos left below 54M, and no more TF needed for expiries >54M). We might now be able to consider 73 bits above 58M.

I put this to George, (who found my original suggestion "very persuasive"), Oliver, Garo, Chalsall and Christianson via PM.
Oliver suggested that we could TF more than 200 a day to 72.
Chalsall & Christianson bleated, and Garo banned me for a week for heresy/blasphemy.

This is what we are up against.

How does the average Joe CPU get a nice "low" exponent to LL or DC these days? Expect decreased throughput and more expiries/returnees in the near future.

The current strategy is counterproductive.

David

[axn]

Originally Posted by davieddy

Expect decreased throughput and more expiries/returnees in the near future.

Explain

[diamonddave]

Originally Posted by davieddy

The current strategy is counterproductive.

I'm very curious what you understand is the current strategy and why it's counterproductive.

Because in the last week we:

1) 2000 exponent (40M to 59M) were TF to 72, While only 1160 were LL tested in the same range.

2) 1400 exponent (20M to 29M) were TF to 69, While only 849 were DC tested

[davieddy]

Originally Posted by diamonddave

I'm very curious what you understand is the current strategy and why it's counterproductive.

Because in the last week we:

1) 2000 exponent (40M to 59M) were TF to 72, While only 1160 were LL tested in the same range.

2) 1400 exponent (20M to 29M) were TF to 69, While only 849 were DC tested

First of all, due respect to Chalsall for making this sort of data so easy to come by.
Your 1) merely justifies my assertion that the 200 new LL assignments per day can comfortably be GPUed to 72, with something (albeit insufficient) to spare for the ~1000/day (regrettable) expiries/returnees. So settle for 70 bits ATM.

For 5 years I have monitored the Primenet Summary hourly.
This may sound sad, but I have learned much from it, although Chalsall's spider has rendered this summary even more opaque than it used to be:(

Big mystery: when (if ever) do his grabbed exponents get released to the masses? This is part of the counter-productive business.

As I anticipated, 26000 primenet TF assignments 50-60M have ground to a halt, presumably lower than 72 bits. Time these were expired, and I would suggest Chalsall allocates them and the 58M - 60M range to GPUers.

At least most expos > 53M are TFed to 71 ATM.

David

[davieddy]

Originally Posted by axn

Explain

Well I'm not tempted to complete a 58M LL TFed to 71.