Got first batch of TF to 75

[chalsall]

Originally Posted by TheMawn

I've got my stuff working on 75 bits too and it's soooo sloooowwwww. Maybe it's bad luck but maybe not. I can only clear about 6 per day per card so to hit 75 attempts would take almost two weeks.

How many LLs could you you do in that time in this range?

I'm serious with that question. I'm entirely directed by James' analysis; we're "just at the edge" doing LLTF at 66M (or, at least, should be).

[TheMawn]

I've never tried CudaLucas with this card but my CPU can LL one in about two weeks, per core.

[Chuck]

Actually I've gotten a little more luck lately and now have 3 factors in the 75 bit range.

[chalsall]

Originally Posted by TheMawn

I've never tried CudaLucas with this card but my CPU can LL one in about two weeks, per core.

It's entirely a function of what that card can do; CPUs are to GPUs as apples are to oranges....

[Chuck]

I have received a couple of "results not needed" to 75 errors back from MISFIT this afternoon.

One is 69059299, which is showing up as a completed assignment by me in GPU72, but mersenne.org indicates it was processed by ktony. The other error I had earlier today was also completed previously by ktony.

[chalsall]

Originally Posted by Chuck

One is 69059299, which is showing up as a completed assignment by me in GPU72, but mersenne.org indicates it was processed by ktony. The other error I had earlier today was also completed previously by ktony.

Hmmmm... OK, looking at the GPU72 Assignments table, it appears that ktony was assigned this 2015-02-11 07:13:46, but then unreserved it 2015-02-11 07:26:18. You were then assigned it 2015-02-14 01:09:00, and then ktony completed it sometime around 2015-02-15 20:00.

Sorry about this, but there's not much I (nor Primenet) can do in cases where someone unreserves work, but then completes it.

The reason GPU72 is crediting you with the completion is you were the "Owner" when it completed -- it doesn't actually check who completed the work, solely on who owned it at the time.

[kladner]

Originally Posted by chalsall

Hmmmm... OK, looking at the GPU72 Assignments table, it appears that ktony was assigned this 2015-02-11 07:13:46, but then unreserved it 2015-02-11 07:26:18. You were then assigned it 2015-02-14 01:09:00, and then ktony completed it sometime around 2015-02-15 20:00.

Sorry about this, but there's not much I (nor Primenet) can do in cases where someone unreserves work, but then completes it.

The reason GPU72 is crediting you with the completion is you were the "Owner" when it completed -- it doesn't actually check who completed the work, solely on who owned it at the time.

Apologies to all. I had also gotten a "Not needed" response, for one Chuck had finished. Now I am suspicious of a string of remaining assignments.

[PM sent to chalsall and Chuck.]

I have cleared my worktodo files and sent the contents. I will work on something else until this is resolved.

[casmith789]

I got a weird result.

I got the assignment 66203861 from GPU72 18 days ago (6th Feb) and finished it today.

To my surprise, looking at the assignment page it was assigned for LL to curtisc on the same day (6th Feb)

Does that mean GPU is giving away assignments it doesn't actually own?

[chalsall]

Originally Posted by casmith789

Does that mean GPU is giving away assignments it doesn't actually own?

Hmmmm... It definitely shouldn't, but looking at the records in this case it clearly did. Sorry about that.

Bdot reported a similar issue a while back; clearly I've got a bug somewhere. Looks like a race condition.

Do you see any other examples of weirdness? Specifically, results you've completed and submitted and show up on Primenet, but not at GPU72?

I definitely want to find and squash this -- no parallel assignments are ever supposed to happen.

[Xyzzy]

Here are some factoring statistics for 1,915 tests.

Do these numbers reflect a reasonable allocation of resources or is going to 75 too expensive? (Are ~2,000 tests enough data?)

[Mark Rose]

Originally Posted by Xyzzy

Here are some factoring statistics for 1,915 tests.

Do these numbers reflect a reasonable allocation of resources or is going to 75 too expensive? (Are ~2,000 tests enough data?)

The question to ask is if you could have LL'ed 18 exponents on the same hardware in the same amount of wall time. Could you have?