Proposed DC assignment and recycle rules

[Prime95]

It is a short-term rolling average. The computer must still pass the reliability/confidence/minimum-speed tests to get preferred assignments.

If the wording on the assignment rules page can be made more clear, please make suggestions.

[henryzz]

Originally Posted by Prime95

2 results means 2 LL results (LL or DC).

There is no "penalty" for having been doing LL. Yes, it is harder to complete 2 results in 90 days, but I think a Core2 computer (4 or 5 year old technology) could meet that requirement.

These rules are a first draft to see how well they work. The penalty for getting a category 4 vs. a category 3 is only 2 million (a 34M exponent instead of a 32M).

What happens if a pc runs 1 core LL and 3 cores on TF/ECM/P-1?
It seems to me that the 1 core on LL would need to be finishing 8 LL per 90 days. Either that or it needs rewording. It should be 2 IMO.

[Prime95]

Originally Posted by henryzz

What happens if a pc runs 1 core LL and 3 cores on TF/ECM/P-1?
It seems to me that the 1 core on LL would need to be finishing 8 LL per 90 days.

Good point. I'll see if I can hack the SQL to do something better.

[Prime95]

Originally Posted by henryzz

What happens if a pc runs 1 core LL and 3 cores on TF/ECM/P-1?
It seems to me that the 1 core on LL would need to be finishing 8 LL per 90 days. Either that or it needs rewording. It should be 2 IMO.

Fixed both the code and the web page.

[petrw1]

Just noticed this comment on the DC rules page
"Your account is set to get the smallest exponents. All your computers must meet the requirements for returning results in a timely manner as outlined below."

This concerns me. Among my dozen or so PCs most will pass but some certainly not.
I suspect many others are so too. This suggests to me I am completely ineligible, though I have about 20 fast and full time cores assigned to what I had hoped would be preferred DC and LL. I expected these rules to be PC based rather than account based.

[Prime95]

I'll change the wording.

[Prime95]

The server applied the new rules for recycling DC assignments in the top 1500.


Assignments over a year old and < 60% complete were recycled. Assignments over a year-and-a-half old were recycled.


The actual SQL query was:

      ((dt_when_assigned < '2014-03-01' AND       -- Grandfathered assignment  
         exponent < @exp1 AND		           -- exponent is in the most critical category  
         dt_when_assigned < DATEADD (DAY, -365, GETDATE()) AND     -- and assignment is over a year old  
         percent_done < 60 + (DATEDIFF (DAY, dt_when_assigned, GETDATE()) - 365) / 3) OR -- plus a grace period if close to finished

[chalsall]

Originally Posted by Prime95

Updated web page for review:

http://www.mersenne.org/thresholds

I'll be tweaking it off and on, so don't be surprised if there are bugs

Just putting this out there for thought and discussion...

Modeling this a bit more, I'd like to suggest that the ranges for categories 1 and 2 might perhaps be optimally increased, assuming our goal is to "compress the 'wave'", reduce "poaching" and to try to ensure that work is completed appropriately, rather than having a slow machine waste its time and energy only to have its assignment be recycled by the system after it had already invested cycles...

As an example, I personally can do about 15 DCs a day on my machines alone (and I'm by no means the fastest DC'er), so given the 1,500 cut-off for "Category 1" which must be completed within 60 days, I can do about 900 of these myself (if assigned), and thus those assigned low "Category 2" work will quickly become "Category 1" and be recycled after the 100 day limit.

Meanwhile, my and other fast computers will be assigned Cat 2 or even Cat 3 assignments. Now, maybe this isn't a bad thing, and I'd have no problem taking on such work.

My primary point is once the legacy assignments (more than a year old assigned before 2014.03.01) are recycled and processed, we're likely going to again find ourselves with a situation where the slower machines are once again holding up milestones (for up to 240 days) and possibly have their work wasted.

Personally, I'm happy to take on whatever assignments make the most sense. But since this exercise was to help prevent hold-ups (and discourage poaching), perhaps the cut-offs should be increased. Such as (just throwing numbers out there) 5,000 for Cat 1, and 10,000 for Cat 2 (basically a bit less than our daily production rate times the number of days they are valid).

Thoughts?

[LaurV]

Although what you propose makes totally sense (in my opinion), my take would be more of "let's run for a while with what we have" and see how the things progress. The limits may be adjusted later, and generally, not everybody have 96 high-profile CPU cores on his hand, as you do (talking only about the six R720 servers, and not counting the "normal" computers and workstations you control )

[tha]

Originally Posted by chalsall

Thoughts?

As far as I am concerned the objective of the rules is to make sure that low end exponents are handled by fast reliable machines in due time. So the main thing we have to agree on is what constitutes a 'low end' exponent. I agree that the amount of available fast reliable machines influences that.

However we are not trying to feed the fast reliable machines with low end exponents to keep them busy. We should use these machines when needed, not when available.

[chalsall]

Originally Posted by tha

However we are not trying to feed the fast reliable machines with low end exponents to keep them busy. We should use these machines when needed, not when available.

OK... My main concern is not that the fast machines are only given the lowest candidates (and one of my machines is intentionally configured to get Cat 2 assignments), but rather that the slower machines may have their work wasted because of recycling once they move into the Cat 1 range.

But whatever is communally decided I'm fine with.