Newer milestone thread

[Brian-E]

Originally Posted by richs

I also do what petrw1 mentions in post no. 1065. I carefully manage my DC assignments, and I've already lost some work in this latest move to organized poaching. So once again I respectfully ask everyone not to touch my assignments. Thank you.

And I, too, frequently still have an in-progress assignment which is older than 90 days. Two reasons:

- Sometimes assignments actually take me longer than 90 days to carry out when working on them: my machine is only switched on for a fraction of the time

- I get my assignments from GPUto72 when they are sometimes already a few weeks old according to PrimeNet. For this reason I suggest also subtracting a fixed time, say 20 days, from the age of the assignment before calculating its expected finishing time.

[chalsall]

Originally Posted by Brian-E

And I, too, frequently still have an in-progress assignment which is older than 90 days.

I /think/ what is being discussed here is not to automatically re-assign candidates just because they're older than 90 days, but rather that after 90 days enough progress should have been done to be able to calculate with reasonable accuracy when the work on the candidate will be completed. This is instead of trusting the client's estimate, which is often very inaccurate (because the machine is not on as much as the configuration is set to -- by default 24 hours/day).

If no work has been done, or the completion date is past a year or so, then Primenet would reassign it to someone else's machine with a better history of productivity.

[chalsall]

Originally Posted by Brian-E

I get my assignments from GPUto72 when they are sometimes already a few weeks old according to PrimeNet. For this reason I suggest also subtracting a fixed time, say 20 days, from the age of the assignment before calculating its expected finishing time.

I just realized what you were saying here; excellent point.

I'll modify my Perl script to have it check if the assignment was made through GPU72, and if so use its assignment date rather than when the candidate was reserved from Primenet.

[tha]

Originally Posted by richs

I also do what petrw1 mentions in post no. 1065. I carefully manage my DC assignments, and I've already lost some work in this latest move to organized poaching. So once again I respectfully ask everyone not to touch my assignments. Thank you.

This is exactly what we wanted to prevent from happening.

In the years that the v5 server has been in operation the tail end of the DC front has not moved noticeable. Now, in just a few weeks it moves by about 5.000.000. Neither situation indicates a healthy condition.

So, during a walk in the forest today I came up with Yet Another Metric to let the server recycle assignments in time. I hope this one is the most simple to implement, easy to understand and fairest to all.

The treshold value for preferred DC exponents is saved by the server on a day to day basis. All the exponents lower than or equal to the threshold value and higher than the treshold value of the previous day get a year from that day on to be completed. If a client reports to the server an 'estimated completed by date' of more than 6 months prior to the final day the server sends an informative warning message to the client about the remaining time available. If the client reports a date beyond a year than the server takes the assignment away from the client.

So a client can lose an assignment if:

a: 60 days go by without a report.
b: it is estimated to take more than a year to complete since the treshold value passed the exponent value.

For first time LL test we may have to look for a similar value of the 'one year since the treshold value passes the exponent' limit.

Maybe it is even possible to assign exponents below the preferred treshold value that have less than a year to complete only to clients with an average rolling value high enough to complete the assignment still in time.

[TheMawn]

I'm finding it extremely difficult to sympathize with people hoarding their preferential numbers for over ninety days. There is an option to tell your program how many days of work to stockpile, i.e. 5, not 180.

I don't think know why people should be are getting batches of work for their CPU from GPU72. The Prime95 client is just so good at managing the exact cause of some of the problems we have. Typical GPU work gets done so fast that over-aged assignments are a complete non-issue. Prime95 can contact the server even on a 2-week basis to give it a reasonably accurate estimate of how long something is going to take.

Instead, people do their assignments manually and give absolutely no progress reports until the assignment gets finished, which leaves the percent complete at 0 for however many months, and flags the assignment for infinity days to completion.

[TheMawn]

I just want to further stress:

Report progress once in a while!

Don't take more assignments than you can complete get started on in ninety days!


Do these two things and these two things only and you will never get "legitimately" poached upon. The Prime95 client can do this very, very well with about thirty seconds of internet once every two months.

[TheMawn]

Originally Posted by TheMawn

"Anonymous" needs restrictions on their ability to get and extend assignments. Perhaps they too could be restricted from taking bottom-of-the-pile assignments entirely and be not allowed to extend them past 90 days for DC, 180 for LL?

Originally Posted by petrw1

Except that ANONYMOUS is many different people who choose not to be named.

Some are very well behaved in this regard ... some ???

The anonymous users who are well behaved do not need their DC assignments extended past ninety days. I'm not saying we should force users to declare themselves. I'm saying they should accept some restrictions in highly preferential assignments and some restrictions in assignment extension past a very generous 90 or 180 days.

[TheMawn]

The problem is with everyone coming in and telling us how they're an exception to every single proposal we give, no progress is going to get done. On the other hand, people could accept that their cherry-picked DC is going to get done by someone else and move on with their lives.

So far:

• We can't kill off all 90-day-old assignments with zero progress because of the GPU72 delay (this I can accept is fine because Chris is working on a fix)
• Even after the fix, we can't kill all 90-day-old assignments with zero progress because some people like hoarding six months worth (or potentially years worth and constantly extend)
• We probably can't kill any assignment with more than 0% progress because they're being worked on (possibly with a microwave oven).

Originally Posted by petrw1

Safer but probably much more difficult might be to look for NO progress and NO reduction in Days to complete over time.

I'm glad petrw1 is actually proposing an alternate criteria. However, his 90-day-no-progress assignments do actually fall into the criteria he suggested; his assignments have made no progress in ninety days. 92, 93, and 97, as a matter of fact.

[chalsall]

Originally Posted by tha

So a client can lose an assignment if:

a: 60 days go by without a report.
b: it is estimated to take more than a year to complete since the treshold value passed the exponent value.

I agree with you, but would extend the criteria thusly:

1. Any candidate held for 120 days without reported progress is automatically recycled.

2. Any candidate held for more than 365 days with less than 50% completion is recycled.

3. Any candidate held for more than 730 days is recycled regardless of the reported completion.

4. Manual assignments are valid for the same 60 days without updates as are regular assignments, and can't be extended except by putting them on a Prime95 / mprime instance and reporting them properly.

4.1. It continues to blow my mind that any random agent (human or bot) can reserve assignments from Primenet with no prior authentication, and they're then held up for 180 days.

5. Any work currently assigned which is no longer needed (e.g. TFing or P-1'ing) is released back into the pool. Several people are holding up candidates for LL assignment while they "claim" to being doing work.

In all honesty, and to be quite frank, it almost seems like George is tacitly approving this current "poaching" activity.

I personally, and I suspect many others in the community, would welcome his thoughts on this discussion.

[chalsall]

Originally Posted by TheMawn

The problem is with everyone coming in and telling us how they're an exception to every single proposal we give, no progress is going to get done.

I don't think this is actually the case. Some very valid points are being raised here.

"richs" is correct that he very carefully manages his assignments. And he did lose some work because of a mistake I made -- "petrw1" agreed to let me take back some "milestone" assignments to process on my own machines, but I made a mistake in my SQL which resulted in three of richs' assignments also being transferred. Even though his workers had not yet reported progress to Primenet, they had just started the work.

(For the record, when he brought this to my attention I immediately stopped work on them, and offered to simply hold them for him to complete.)

Surely we, as a community, can come to some reasonable compromise here. Ideally Primenet and GPU72 will only assign the low LL and DC candidates to those workers who are reliable -- I've already removed two "horders" from being able to reserve low candidates from GPU72.

I'm hopeful George will enter this discussion, and give his perspective and thoughts.

[petrw1]

Originally Posted by TheMawn

I'm glad petrw1 is actually proposing an alternate criteria. However, his 90-day-no-progress assignments do actually fall into the criteria he suggested; his assignments have made no progress in ninety days. 92, 93, and 97, as a matter of fact.

Only 1 of the 2.

1. Reporting progress: No because they are in the queue
2. Reducing completion days: Yes because they are advancing in the queue.

Only if both are No can you reasonably conclude they are essentially abandoned.

Then for those that are progressing the question becomes: Is the total days to completion acceptable? I propose that is a forum decision. And my suggestion would be to allow 90 days of work (because and as long as that is the limit in the entry field) plus up to 60 days to complete. I further agree that if no report and no reduction in days for 60 days it is also fair game.

Finally why might I grab 90 days or more of low-end work?
I was doing this before GPU72 but for similar reasons:
- I could ensure it was going to get focus and get done; maybe in months but not years and not held and abandoned.
- And when there were more than I could handle I offered to share with others of good standing.
For example about 200 DC in the low 20s came available some years back. I took 80. I didn't get many takers for the rest and they went back to the general public...some of which got stuck again. In hindsight I wish I would have grabbed them all. Might have taken me more than a year but that would have been quicker than what happened.