Proposed DC assignment and recycle rules

[Prime95]

Assignment rules (top 3000 exponents):
	*1)  No v4 users
	*2)  No manual reservations
	*3)  Reliability is high.
	*4)  Speed is satisfactory.
	5)  "Days of work" setting is short.

Recycle rules (top 1500 exponents):
	*1)  If not a manual assignment and expected completion date is not updated for 60 days.
	2)  If assignment made before 2014-02-01 and:
	  2a) assignment is 12 months old and < 50% complete it is recycled.
	  2b) assignment is 15 months old it is recycled.
	3)  If assignment made after 2014-02-01 and:
	  3a) assignment is 6 months old and not started it is recycled.
	  3b) assignment is 9 months old it is recycled.
	4)  If DC is completed by another user, the assignment is cancelled.

Assignment rules (not top 3000):
	*1)  Speed is satisfactory.

Recycle rules (not top 1500):
	*1)  If not a manual assignment and expected completion date is not updated for 60 days it is recycled.
	2)  If assignment is one year old and not started it is recycled.
	3)  If assignment is two years old it is recycled.
	4)  If DC is completed by another user, the assignment is cancelled.

[Prime95]

Here is a proposed starting point for DC assignment and expiration rules. Discussion is welcome for improving these proposed rules. We are trying to achieve a happy balance between those that like to see a steady progress on the milestones page without annoying users by rendering their DC work useless by recycling an exponent prematurely.

[chalsall]

Originally Posted by Prime95

Here is a proposed starting point for DC assignment and expiration rules. Discussion is welcome for improving these proposed rules.

Thank you for this.

But could you please define the variable "Reliability"?

Is this simply the success rate (no matter how much time is spent), or does it also include the time spent working on a candidate?

If the former, than we will likely find many candidates assigned to machines which will take many years to complete.

[Prime95]

Reliability and confidence are the values the server keeps on each server. Reliability goes up when you report an LL result without any possible hardware errors. Confidence (in the reliability value) goes up with each result you report.

I think the server's current definition of a reliable computer is one that has reported at least 2 results and at least 90% error free results.

[tha]

Since all assignments will end up in the top 1500 soon enough regardless of when they started I think only the recycle rules for these really matter, the rest is just early recycling of seemingly hopeless assignments. To me these rules look 'aggresive' enough and still balanced.

The only thing I would like to see investigated is if the value of 3000 exponents should not be increased to 4000 because of the amount of available reliable clients. (and the halve of it to 2000)

[Prime95]

The current assignment rules are here:

http://www.mersenne.org/thresholds/

[chalsall]

Originally Posted by Prime95

I think the server's current definition of a reliable computer is one that has reported at least 2 results and at least 90% error free results.

So, then, what you are saying is that Primenet keeps no heuristics about throughput?

Thus, a machine which does one LL every three years is as "trusted" as another which does one every four days; and thus might be given a preferred candidate?

Seems rather short-sighted to me....

[Prime95]

Originally Posted by tha

The only thing I would like to see investigated is if the value of 3000 exponents should not be increased to 4000 because of the amount of available reliable clients. (and the halve of it to 2000)

What does everyone else think? I could see upping the 3000 to 4000, 5000, or 6000. This would keep newbies (more prone to abandoning) out of the lower zone. It would also push slower computers to the higher exponents where they would have more time to complete the assignment.

[Prime95]

Originally Posted by tha

Since all assignments will end up in the top 1500 soon enough regardless of when they started I think only the recycle rules for these really matter, the rest is just early recycling of seemingly hopeless assignments. To me these rules look 'aggresive' enough and still balanced.

The only real advantage to reclaiming exponents before they enter the critical top-2000 is that it lets users get slightly smaller assignments at the leading edge.

Do you think the reclaim rules in the non-top-2000 are too aggressive? Recommended changes?

[Prime95]

Originally Posted by chalsall

So, then, what you are saying is that Primenet keeps no heuristics about throughput?

The server currently uses the reported CPU speed and rolling average to assess throughput (the speed is satisfactory step).

There are circumstances when the rolling average can be reset. Also the CPU speed can be miscalculated by the client.

The confidence/reliability check makes sure the computer has returned at least 2 prior results.

It would not be hard to add a criteria such as "and computer has returned at least 2 * num_cores results in the last 3 months".

[chalsall]

Originally Posted by Prime95

It would not be hard to add a criteria such as "and computer has returned at least 2 * num_cores results in the last 3 months".

If it's not hard, then may we please ask that this is implemented?