Modifications to DC assignment rules

[Prime95]

The LL assignment rules change has been implemented and seems to working OK. Obviously, we need months of experience before we know if the self-adjusting features are working as envisioned.

Now it's time to look at the current DC rules. Feel free to weigh in.

I imagine the goals should be similar to LL assignment goals with one major difference. We need to be far more reluctant to expire exponents. When we erroneously expire an LL test, it becomes a DC, which we need anyway -- no wasted work. When we expire a DC erroneously, and both the new assignment and expired assignment send results, then we have wasted work.

But, first I'll start gathering some data....

In the last year, 81761 LL tests have been completed on exponents between 28M and 47M.

In the last 48 hours, we've assigned the following number by category:

cat 1 139
cat 2 12
cat 3 150
cat 4 3403

DC results reported in the last 30 days grouped by category:

cat 1 1649
cat 2 273
cat 3 2431
cat 4 5176

[chalsall]

Originally Posted by Prime95

We need to be far more reluctant to expire exponents.

Perhaps for DC all we need to do is adjust the boundaries a bit, and perhaps introduce a "Cat 0" (although stragglers seem to be far less frequent).

Further, to your point about not wishing to recycle unless absolutely nessesary, stick with the "Must promise" requirement for Cats 0 through 2.

Auto-promotion from Cat 4 might be useful, although I suspect that most users/machines who are currently getting DC Cat 4 will be getting LL Cat 3 once they've proven themselves.

[Madpoo]

Originally Posted by chalsall

Perhaps for DC all we need to do is adjust the boundaries a bit, and perhaps introduce a "Cat 0" (although stragglers seem to be far less frequent).

Further, to your point about not wishing to recycle unless absolutely nessesary, stick with the "Must promise" requirement for Cats 0 through 2.

I have a feeling that exponents originally assigned as cat 0-2 are going to finish, as long as the steps to make sure "good" machines are getting them is working out well.

There's likely going to be issues with exponents assigned as cat 4 that take WAY longer than they should and wind up getting promoted all the way down to cat 1 area, even though the person is still working on it. I think that's going to happen sometimes no matter how hard we try to avoid it.

But hey, even if so, the assignment rules apply to cat 4 as well and if they wind up taking too long and get expired, well... too bad I guess. It happens.

[Siegmund]

Aside possibly from the small size of category 2, I'm quite happy with how DC rules are working now. (Of course I was quite happy with the LL tests too.)

It seems, compared to the LL anyway (maybe just because the tests dont take as long), that we have been progressing through the cat 1 exponents in a steady fashion and not having many huge outliers.

I am not too worried about lost work from people really making slow progress on cat 4 exponents after a whole year.

[chalsall]

Originally Posted by Siegmund

I am not too worried about lost work from people really making slow progress on cat 4 exponents after a whole year.

Excellent. Good to know your opinion.

[Prime95]

Originally Posted by chalsall

Perhaps for DC all we need to do is adjust the boundaries a bit, and perhaps introduce a "Cat 0" (although stragglers seem to be far less frequent).

Further, to your point about not wishing to recycle unless absolutely nessesary, stick with the "Must promise" requirement for Cats 0 through 2.

I also think the current rules are not seriously deficient. The only "must have" items are the addition of cat 0 and banning any computers that have recently expired or bad/suspect results from that cat 1 area.

I also agree that since we have good cat 1 participation now, we should continue limiting cat 1 assignments to those that have signed up for smallest exponents.

I'll put together some changes consistent with my thoughts above and that make the rules similar to LL testing and keep y'all posted.

[Dubslow]

As before, my major complaint is that the throughput bounds on Cat 1 are not anywhere near tight enough. The current throughput should be Cat 2, and you could easily triple the throughput bound for Cat 1 and still not even come close to what a five year old Sandy Bridge desktop core can put out.

[Prime95]

See the new thresholds page: http://mersenne.org/thresholds/

The cat 3/4 boundary is going to move to 52000 exponents. I'll keep an eye on how many assignments are going to each category and post the results here.

As always, let me know if you see anything suspicious -- no guarantees I implemented it all correctly.

[Prime95]

Originally Posted by Dubslow

As before, my major complaint is that the throughput bounds on Cat 1 are not anywhere near tight enough.

In the last 24 hours, we've assigned the following number by category:

cat 1 48
cat 2 98
cat 3 46
cat 4 1494

So it looks like the rules for cat 1 are now a bit tighter. I did not break it down as to why -- it could be the no expireds, no bad/suspect, days-of-work <= 5, or GHz-days/day/worker requirement.

There is a big jump in cat 2 assignments as there is no longer a requirement that the user sign up for smallest exponents.

At the rates above, cat 1/2 boundary should be 6000 exponents (120 day supply @50/day) and the cat 2/3 boundary should be increased to 32000 exponents (240 day supply @150/day).

We didn't discuss this earlier: are we happy with the 60/120/240/360 day expirations for cat 1/2/3/4 DCs?? My random thoughts: The generous timeframes do decrease the chance we waste work due to early expirations. A downside is generous timeframes increase the cat boundaries which ends up giving users bigger exponents. A DC test takes about 1/4 the CPU time of an LL test and we need a way for GIMPS' slowest contributors to participate without fear of poaching, but is 360 days too generous?

I wonder if I can conjure up a SQL query that would tell me how many DC results come in that take 240 to 360 days. If the answer is very, very few then maybe a 60/120/180/240 day time limit for cat 1/2/3/4 would be better.

[Mark Rose]

I think we should err on the side of extra time with DC. If an LL assignment expires or is poached and then finishes, the work is not wasted. With DC, the work is usually wasted. I think we should aim for all cat 2/3/4 assignments to be finished before they become cat 1, for the extra time margin.

[Syntony]

There may be an unintended consequence of using 'suspect' results as a reason for downgrading a machine in double-checks. A double-check which disagrees with the first LL check is apparently counted as 'suspect', regardless of whether it is later verified as correct by a third check. As I've found a number of faulty first-time checks (later verified by another user) in the past year, I now find I'm considered unreliable and excluded from the higher categories...