Assignments?

[R.D. Silverman]

The range {10330000 12830000} has had 3 unfinished candidates for ages.
Isn't anyone doing these?

[nfortino]

Originally Posted by Bob Silverman

The range {10330000 12830000} has had 3 unfinished candidates for ages.
Isn't anyone doing these?

Yes, they are all assigned by primenet. One seems to be close to expiring, and will most likely be finished within a couple days of expiration. The other two are making progress, however, only one has a chance of finishing in a semi-finite amount of time.

[garo]

I think we need to ask George to do a formal release on at least one of the exponents which shows over 1000 days to completion. Otherwise it is becoming an attractive target to poachers.

[Matthias C. Noc]

I don’t understand why there are no faster re-assignments for such exponents. We are not even loosing something. Should the result of the original come in it would be used as double-check. So we only would run the double-check before the first LL is in.

Matthias Noch

[garo]

well, these exponents were assigned to the respective users first and if any one of these exponents is prime, we do not wish to e unfair to them and take away their chance of discovering the prime. However, fair does not mean waiting forever and George usually releases exponents that have not had any work done or are being worked on very very slowly, once every 6 months or so. He will probably be doing a release of these numbers too. But in the meanwhile I hope no one wil take it upon himself/herself to complete them as that is "poaching" and we all know that "poaching is very bad" :)

[cheesehead]

Will one or more of the complainants please refresh my understanding of precisely what harm is being done by these slow-completers?

[cheesehead]

Apparently some of the complainants feel impatience when they view the progress reports of the slow-completers. But those are only internal feelings generated by the complainants themselves, not some real damage done by the slow-completer. So, what actual harm is done in reality by any slow-completer?

I've seen the discussion of errors caused by cosmic rays, and that the likelihood of such errors is proportional to the length of the L-L run. In those discussions it is apparently assumed that slow-completers pose a greater risk of turning in erroneous reports due to such errors.

But the likelihood of cosmic-ray-induced errors is also proportional to each computer's sensitivity to such errors. The slower computers are likely to be those with larger mask sizes used in printing their circuits, thus reducing their sensitivities to cosmic-ray damage, relative to the faster systems with thinner circuitry in which a cosmic-ray hit can affect a larger fraction of the area of a component.

- - - - - - -

Bob, Matthias, anyone else -- can you demonstrate actual real harm (not just an effect on someone's feelings) done by a slow-completer?

[Uncwilly]

< advocate >

Cheesehead,

Perceived lack of progress of the whole project can make people feel like nothing is happening and that the project is stalling. If people feel that the project is stalling out, they may decide that their cycles have a better impact else where. The rats-jumping-off-a-sinking-ship syndrom is of no help to the project.

A smooth progress milestone production is a good thing. People don't like to see big spikes and then slow speed. Whiplash ain't fun.

< / advocate >

[cheesehead]

Okay, that's another feelings example.

Perhaps some of our mathematically-simple progress indicators are ill-suited (in terms of feelings generated by impatient observers :) to the reality of what they're measuring. GIMPS progress, depending on the efforts of human volunteers using systems with a wide range of capabilities, is necessarily always going to have some stragglers. That doesn't mean GIMPS isn't progressing smoothly.

Proposal for the GIMPS status page: (1) Instead of always showing the single lower bound of DCs, which progresses in jerks, show only the average exponent of the trailing 100 (or 41 or whatever) DCs. This average would, I think, increase relatively smoothly and not give the misleading impression of jerky progress. (2) Stop showing the countdown of exact number of tests needed to prove Mxx is Mxx -- anyone who wants to know these numbers can derive them from the various report files. When the countdowns reach the single digits, they're not really informative -- they're just provocative to the impatient.

[Matthias C. Noc]

Originally Posted by cheesehead

Bob, Matthias, anyone else -- can you demonstrate actual real harm (not just an effect on someone's feelings) done by a slow-completer?

Hi,

you asked what harm slow-completer are causing. I have three points:

1.) The project is not only about finding isolated new Mersenne primes, but although about proving that these new found are the only ones in the range the current GIMPS is covering. We want to find all Mersenne primes in the data range and prove that they are the only ones. That means that we need to cover all exponents below a given new Mersenne prime to prove that this one is the number 36, 37 or 41. If it takes years to finish one small exponent, which under today’s normal performance would take only 4 of 5 days to calculate, it is holding up the whole process without need. That doesn’t help to get a better understanding of the distribution of those Mersenne primes.

2.) It makes a bad impression about the organisation of the project when we can’t finish a data range in an acceptable amount of time. We are crunching very large numbers, but fail to finish much smaller exponents. That doesn’t look good.

3.) We are doing a very abstract stuff here. To keep people motivated one needs to reach small and big milestones from time to time. As we don’t find a new Mersenne every month other milestones have to keep people interested. Like :

All exponents below 8,715,700 have been tested and double-checked.
All exponents below 12,441,900 have been tested at least once.
Countdown to testing all exponents below M(13466917) once: 4
Countdown to testing all exponents below M(20996011) once: 9,549
Countdown to testing all exponents below M(24036583) once: 42,739
Countdown to proving M(13466917) is the 39th Mersenne Prime: 39,154
Countdown to proving M(20996011) is the 40th Mersenne Prime: 204,395
Countdown to proving M(24036583) is the 41st Mersenne Prime: 271,351

If we loose one normal cruncher with an up-to-date PC due to frustration about no progress we are loosing much more power then when we loose 10 angry slow-crunchers because their exponents were reassigned in time to keep the project rolling.

It will close to be impossible to keep the project running in the much higher areas with only slow-completers. We need as much high-power as we can get now that we really working big numbers.

Best,

Matthias C. Noch

[S00113]

Originally Posted by Matthias C. Noc

If we loose one normal cruncher with an up-to-date PC due to frustration about no progress we are loosing much more power then when we loose 10 angry slow-crunchers because their exponents were reassigned in time to keep the project rolling.

Each time you loose one slow cruncher, there is a great cance you'll loose the rest of his machines too. (I have >200, a few of them are rather slow, but I don't want to see cycles wasted.) It is very discouraging when you find that the exponent you have been working on the last four months was poached, and I don't think you'll find that many people leaving the project because of a couple of exponents testing very slowly.

I think we should take another approach to handing out work. Let us decide how large exponents we want by the number of days it will take to complete it, and hand out exponents in lots of different ranges at the same time. This will let everyone make steady progress and see exponents complete regularly at the speed he or she decide. And let the server deny requests for exponents which will take more than 120 days or so to complete. We will make progress in all ranges at the same time, and if you believe that there is a good chance to find a prime in a specific range, you can choose exactly that range. Machines will last longer in the project (remember that every cycle count as long as the exponent is completed) and produce results longer. Shure we won't complete the lower ranges as fast, but we will have perfect work for all kinds of computers for longer. When there are no more exponents left to test at 120 PIII days, the PIIIs will get double checks from the higher ranges, while there is still work left for Pentium IIs in the lower ones.