GPU to 72 status...

[davieddy]

Originally Posted by Dubslow

I would rather get the lower expos at 71 bits than the higher expos at 69 bits. Make sure everything it thoroughly done.

Totally bonkers.

In the days when you could follow the daily LL assignments
on the "Primenet Summary" page, you would have seen that
there were ~1000 per day, but only ~200 of these were "brand new",
thereby advancing the wavefront.

72 was determined simply because 200 a day could be done
to 72 AHEAD of the wavefront.

Rate of "saving work" goes as exponent^3.

Spidy grabs returned expos, and could take 69 to 70,
starting with the highest and working down.
But the priority should be taking expos just of the wavefront
straight to 73, I'd judge by now, working just ahead of the wavefront.
The reason this didn't start back when the wavefront was 53M is
that Primenet embarked on a half-hearted "one bit at a time"
strategy which (as I predicted) fizzled out at 70 or 71 bits.

David

[chalsall]

Originally Posted by davieddy

Totally bonkers.

Sigh... Here we go again...

Originally Posted by davieddy

In the days when you could follow the daily LL assignments on the "Primenet Summary" page, you would have seen that there were ~1000 per day, but only ~200 of these were "brand new", thereby advancing the wavefront.

You always talk about "brand new", but never explain how such an candidate is any more important to TF than some poor little orphaned candidate which has been abandoned. Many argue it is more important to process the lower candidates first -- eliminate what we can and get the rest back into PrimeNet's pool for processing -- since they might be prime, and doing the LL test is less expensive than higher candidates.

Originally Posted by davieddy

72 was determined simply because 200 a day could be done to 72 AHEAD of the wavefront.

Then isn't it impressive that we're currently doing about 600 a day WITHIN the wave, where it's computationally more expensive?

And to be a little more empirical (rather than hysterical), over the last month there have been on average 197 LL tests completed. During the same period, 28 candidates per day were eliminated thanks to Trial and P-1 factoring.

Originally Posted by davieddy

But the priority should be taking expos just of the wavefront straight to 73, I'd judge by now, working just ahead of the wavefront.

That is your opinion, which isn't shared by most people.

The end of the day David there is only one person who could convience us to change our stategy. And just in case it isn't clear, that person isn't you....

[petrw1]

Originally Posted by chalsall

The end of the day David there is only one person who could convience us to change our stategy.

Then it has to be:
http://www.chucknorrisfacts.com/search/node/prime

Note especially the 3rd fact.

[davieddy]

Originally Posted by chalsall

Sigh... Here we go again...



You always talk about "brand new", but never explain how such an candidate is any more important to TF than some poor little orphaned candidate which has been abandoned. Many argue it is more important to process the lower candidates first -- eliminate what we can and get the rest back into PrimeNet's pool for processing -- since they might be prime, and doing the LL test is less expensive than higher candidates.



Then isn't it impressive that we're currently doing about 600 a day WITHIN the wave, where it's computationally more expensive?

And to be a little more empirical (rather than hysterical), over the last month there have been on average 197 LL tests completed. During the same period, 28 candidates per day were eliminated thanks to Trial and P-1 factoring.



That is your opinion, which isn't shared by most people.

The end of the day David there is only one person who could convience us to change our stategy. And just in case it isn't clear, that person isn't you....

I love a good logical argument.
More anon, but in the meantime I shall watch "Pointless".
David x

[Dubslow]

Originally Posted by chalsall

And to be a little more empirical (rather than hysterical), over the last month there have been on average 197 LL tests completed. During the same period, 28 candidates per day were eliminated thanks to Trial and P-1 factoring.

Does this mean we are only getting about half as much P-1 as necessary? There are those like Never Odd or Even on PrimeNet, but I don't think PrimeNet is also getting 85/day not counting us... is there any way for mersenne.info to get a daily P-1 count?

(I should also point out that NOoE has a bizzare attempted/success count.)

[axn]

Originally Posted by chalsall

You always talk about "brand new", but never explain how such an candidate is any more important to TF than some poor little orphaned candidate which has been abandoned.

While the "brand new" rhetoric is distracting, there is a certain logic to this. The aim of GPUto72 is to reduce/eliminate "Unwanted LLs" -- defined as any completed LL test that shouldn't have run in the first place because it has a factor b/w PrimeNet limits and extended limits (but the factor was never found). Perfection is achieved when there are zero "unwanted LLs". However, if we have to choose b/w preventing a 45M unwanted LL and a 55M unwanted LL, we should favor the 55M one because preventing the latter would lead to greater reduction of "waste". Thus leading edge of LL assignment should be the focus.

Having said that, I believe this is only a temporary problem, and that the rate at which GPUto72 is progressing, it should soon overtake the LL wave (any projections on when this might be?).

Originally Posted by chalsall

Many argue it is more important to process the lower candidates first -- eliminate what we can and get the rest back into PrimeNet's pool for processing -- since they might be prime, and doing the LL test is less expensive than higher candidates.

From an individual cruncher's perspective, smaller is better. But since this is an open-ended project, from the project's perspective, they are all the same. They all need to be crunched.

[petrw1]

Originally Posted by Dubslow

(I should also point out that NOoE has a bizzare attempted/success count.)

This one?
1552 NoOne 10.911 3 0
3 and 0 does not seem bizarre?
What am I missing?

[chalsall]

Originally Posted by axn

Having said that, I believe this is only a temporary problem, and that the rate at which GPUto72 is progressing, it should soon overtake the LL wave (any projections on when this might be?).

I would have to write a fairly involved query to answer this with extreme accuracy (taking into account those candidates at 71 bits which "Spidy" is currently throwing back, the fact that TF gets cheaper the larger the candidate, etc), but a back-of-the-envelope calculation says at our current rate we can take everything we currently "own" to 72 bits in just short of two months.

And, for some time now, we have been processing many more candidates than are being LLed, which means while we haven't yet "pulled ahead of the wavefront" like we have in the DC range, the "wave" is now compressing.

Originally Posted by axn

However, if we have to choose b/w preventing a 45M unwanted LL and a 55M unwanted LL, we should favor the 55M one because preventing the latter would lead to greater reduction of "waste". Thus leading edge of LL assignment should be the focus.

The assignment pages have always allowed individuals to choose the range they wish. For those who agree that higher is better the system is more than willing to facilitiate.

[Dubslow]

Originally Posted by petrw1

This one?
1552 NoOne 10.911 3 0
3 and 0 does not seem bizarre?
What am I missing?

Never Odd or Even, rank 1.

[petrw1]

Originally Posted by Dubslow

There are those like Never Odd or Even on PrimeNet.

Never Odd or Even is doing very large P-1 at 26 GDs per.
On the other hand James H. is doing much smaller at 0.5 per.

Originally Posted by Dubslow

Is there any way for mersenne.info to get a daily P-1 count?

FYI: I did a query first full 20 days in March.

GhzDays      Attempts  Successes  Per Day
27459.439	14582   632        729

If you are interested in the P-1 activity near the LL line:
Depending on your definition of, there are between about 2,000 and 7,450 that are small P-1 and 40ish that are very large P-1 and could be excluded.

I would exclude all 7500 making the Per-Day average somewhere in the DC/LL ranges of about 354.

[petrw1]

Originally Posted by Dubslow

Never Odd or Even, rank 1.

Ah... Bizarre because he is doing very large P-1....maybe even in the 332M range.