[QUOTE=cheesehead;222191]I've lost track of just what, specifically, defines "confidence" and "preferred" in PrimeNet nowadays. May we have the formulas or algorithms posted again?[/QUOTE]

Reliability is a rolling average of the last 10 results. Where you earn 1.0 for a matching LL result, else 0.0 for a submitted result that did not match a proven good residue (rare), else 0.98 for an error-free result, else 0.5 for an LL run with 1 error, else 0.3 for two or more errors.

Confidence is simply the number of LL runs used in computing the reliability number (max of 10).

[QUOTE=cheesehead;222191]I've lost track of just what, specifically, defines "confidence" and "preferred" in PrimeNet nowadays. May we have the formulas or algorithms posted again?[/QUOTE]

Although I sense this is a sarcastic question, it is a VERY good point. What will define the boundary between a machine we have confidence in and one we don't? Will it be based upon the time it takes to complete exponent N in comparison to how long it should take a computer of that speed to complete the same exponent? I am all for inclusion in GIMPS, but some of these machines contributing minimally do not qualify... unless there are extenuating circumstances in which the participant can only utilize his/her machine for a few precious hours per day/week etc.

Mr. Woltman, I strongly support your actions and given future actions. However, I think that we should define, more precisely, how much subjectivity we are going to introduce.

[quote=Prime95;222164]Yes, the current server is not enforcing the one-year assignment rule.

My new ideas for recycling assignments:

For LL tests on exponents < 80M and not manual testing:

If assignment is one year old and < 50% complete it is recycled.
If assignment is one-and-one-half years old it is recycled.

For non-LLs:

If assignment is 180 days old and < 50% complete it is recycled
If assignment is 270 days old it is recycled

I've coded up a sample query already. 3113 assignments are affected.

Comments?[/quote]

I'll bite.

If I had an LL test at 49% after a-year-and-a-day with 10 days ETC, that would be unassigned and then dealt out to someone with DaysOfWork=90 who'd finish it like 3 or 4 months after I would have? Doesn't exactly seem logical.

I actually [I]do[/I] have 103 P-1 tests between 39M and 40M that are 185-196 days old. All of the exponents have been LL tested at least once before. These tests would then be unassigned and not be dealt out again until the DC wave hits 39M, if at all? That's plain nonsensical.

Finally, I also have 1547 TF assignments between 90M and 91M that are over the 180 days mark (all at 0% and all likely done with this year). These would then be unassigned - making a machine I do not have access to on a regular basis "run dry" - and be dealt out again whenever everything below 90M is at 67 bits, or some such? That's also debatable, I'd say.

Oh, seems like I have half of the 3113 assignments you mention. Call me lucky.

[quote=ckdo;222205]I'll bite.

If I had an LL test at 49% after a-year-and-a-day with 10 days ETC, that would be unassigned and then dealt out to someone with DaysOfWork=90 who'd finish it like 3 or 4 months after I would have? Doesn't exactly seem logical.[/quote]
Hmm...I suppose that could happen if a computer fast enough to do an LL every 20 days has been sitting on a queue of at least 18 (!) of them and is now working on the 18th, but I don't imagine anybody seriously sitting on a queue of that many LL's due to their sheer size.

For TF and P-1 assignments, however, such a scenario is not entirely uncommon, as demonstrated by your current assignments as described. I would suggest, therefore, that non-LL/DC assignments be left alone by the new assignment recycling rules: after all, they don't have any affect on milestones, so nobody's going to care if there's a few TF or P-1 assignments "behind the curve" being worked on over a long span of time.

Thoughts?

[quote=mdettweiler;222206]Hmm...I suppose that could happen if a computer fast enough to do an LL every 20 days has been sitting on a queue of at least 18 (!) of them and is now working on the 18th, but I don't imagine anybody seriously sitting on a queue of that many LL's due to their sheer size.[/quote]

Well ... I'm getting [URL="http://mersenne.org/assignments/?exp_lo=35699737&exp_hi=35839163&execm=1&extf=1&B1=Get+Assignments"]close[/URL]. :wink:

[quote=ckdo;222205]I'll bite.

If I had an LL test at 49% after a-year-and-a-day with 10 days ETC, that would be unassigned and then dealt out to someone with DaysOfWork=90 who'd finish it like 3 or 4 months after I would have? Doesn't exactly seem logical.
...[/quote]
It seems that George's ideas for new rules for recycling assignments ought to be refined to also look at the recent rate of progress for the given assignment to take into account large contributors like you who reserve a lot of work in advance.

However, i'm puzzled: how do you justify reserving many exponents in advance when you are not yet in a position to work on them?

[QUOTE=mdettweiler;222206]Hmm...I suppose that could happen if a computer fast enough to do an LL every 20 days has been sitting on a queue of at least 18 (!) of them and is now working on the 18th, but I don't imagine anybody seriously sitting on a queue of that many LL's due to their sheer size.

For TF and P-1 assignments, however, such a scenario is not entirely uncommon, as demonstrated by your current assignments as described. I would suggest, therefore, that non-LL/DC assignments be left alone by the new assignment recycling rules: after all, they don't have any affect on milestones, so nobody's going to care if there's a few TF or P-1 assignments "behind the curve" being worked on over a long span of time.

Thoughts?[/QUOTE]

I like this idea.

[QUOTE=ckdo;222205]If I had an LL test at 49% after a-year-and-a-day with 10 days ETC, that would be unassigned and then dealt out to someone with DaysOfWork=90 who'd finish it like 3 or 4 months after I would have? Doesn't exactly seem logical.[/quote]

Is there some reason this computer is not set up in a normal manner - to get first-time LL tests in a 90 day queue.

I don't mean to pick on you, just raising a question: I suspect you've set it up this way because you like working on smaller exponents with a better chance of discovering a new prime (I do too). Is it "fair" to other GIMPS members to lock up these smaller exponents for a year? two years? five years?

[quote]I actually [I]do[/I] have 103 P-1 tests between 39M and 40M that are 185-196 days old. All of the exponents have been LL tested at least once before. These tests would then be unassigned and not be dealt out again until the DC wave hits 39M, if at all? That's plain nonsensical.

Finally, I also have 1547 TF assignments between 90M and 91M that are over the 180 days mark (all at 0% and all likely done with this year). These would then be unassigned - making a machine I do not have access to on a regular basis "run dry" - and be dealt out again whenever everything below 90M is at 67 bits, or some such? That's also debatable, I'd say.

Oh, seems like I have half of the 3113 assignments you mention. Call me lucky.[/QUOTE]

It seems like the suggested policy needs to be tweaked somewhat. Perhaps, it should only apply to exponents in or near the "preferred" ranges. That is, something like DC's under 24M, first LLs under 42M, P-1 under 50M, and TF under 60M.

The expired P-1's wouldn't be a problem (theoretically). You'll complete them and get credit long before they are handed out for double-checking -- though clearly it is better to keep the assignments for tracking purposes. I'll have to see how hard it is to query if this is a P-1 assignment on an already LL'ed exponent and thus not holding up anything.

I don't know if we need any LMH rules. Are there any LMH assignments that look abandoned? Probably best to worry about it in a few years -- plenty of exponents to go around.

[QUOTE=Prime95;222164]Yes, the current server is not enforcing the one-year assignment rule.

My new ideas for recycling assignments:

For LL tests on exponents < 80M and not manual testing:

If assignment is one year old and < 50% complete it is recycled.
If assignment is one-and-one-half years old it is recycled.

For non-LLs:

If assignment is 180 days old and < 50% complete it is recycled
If assignment is 270 days old it is recycled

I've coded up a sample query already. 3113 assignments are affected.

Comments?[/QUOTE]

If possible I would also suggest considering "recent" progress. That could possibly address ckdo's concern. So if I have a LL a year old and only 20% done but all 20% is in the last month that is a good indication it was waiting it's time most of the year but now is actively processing and is likely to finish soon. On the other hand if after a year it is 90% done but NO progress in the last 8 months is a good indication it has been abandoned.

I too would be less concerned about TF, P1 and ECM though considering these assignments are measured in hours or days I think the most likely reason for little or no progress over a long period of time is that it was abandoned OR is queued up as in ckdo's scenario again. Could you consider the recent activity of the CPU the "old" assignment is registered to. If it has been churning out 10 TF's a day for the last month then odds are good it will get to these old assignments too in due time.

Mind you, one could argue whether it is reasonable to reserve a large collection of assignments that you won't get to for over a year. In simple terms this is a debate between personal interest in a certain range or type of assignments VS the collective interests or progress of the project as a whole.

[QUOTE=Prime95;222174]Standard preferred assignment rules will apply. There are many TF assignments in the 3113 - they aren't all LL and DCHKs.[/QUOTE]

Does this mean that only fast machines will be assigned "milestone blocking" work?

If not, a sincere question... Is this a policy decision, or because of implementation issues (read: development required on the PrimeNet server)?

If the latter, may I please make an offer.

As many of you know, I have implemented a "mini-you" PrimeNet server which I configure as a Proxy on my "cluster". The machines talk to my server which assigns work, and passes the results back to PrimeNet in a slightly delayed process.

I could, with very little effort, make this available to selected individuals with fast machines to process those exponents which are holding up milestones. George would have to provide me with a dataset containing the work-type and official AID, which my server would hand out to those selected volunteers. Just to be clear, the work would be credited to the volunteer, not me.

I'd suggest that the selected volunteers agree that they must complete the work within, say, one month for a DC and two for a LL (or whatever is agreed is reasonable). And I pledge that my own cluster (which will begin doing serious DC work once all 63 -> 64 LMH work is complete) would not receive any preference (this would be independently verifiable on the PrimeNet server itself).

George -- if you're interested, and the GIMPS community agrees, you have my e-mail address.

(I make this offer in the hopes of avoiding any future poaching arguments. :smile:)

[QUOTE=chalsall;222254]Is this a policy decision, or because of implementation issues (read: development required on the PrimeNet server)?[/QUOTE]

George: I probably missed the updates when I was off but what ever happened to the thread discussion this Spring about getting a new PrimeNet Server? Is it already in place?

Thanks