Strategic Double Clicking

[airsquirrels]

Originally Posted by Mark Rose

I do wonder how many assignments you two do on a monthly basis.


Also, Madpoo, how many machines have now completed an LL test without having a DC currently assigned or completed? That would be a nice number to have a countdown of.

I would be an advocate of formalizing our goals around the various strategic check projects and hopefully getting others onboard to help reach them. Madpoo, if you want to reserve the ideal-range chunk Mark gave for me I'll put them on my new system for a quality burn in.

Over the past three months I've completed 1650 LL tests, 550 a month. Not all of those are on GPUs, but I estimate around 80% are. ~66% of those are verified DCs, 20% unverified LLs, and 10% open mismatches.

Two bad results so far in the past three months, one from a FuryX about to drown and 2 (counting one where I didn't match madpoo) from the GTX1080. Either it or CUDA8 RC are not 100% reliable.

I have 17 Verified, 2 bad, and 12 mismatches from that card. If anyone wants to TC. So far all the 7xM tests that card has done have mismatched.

38096281 M
38133923 M
38324009 M
38324521 M
39299047 M
39334423 M
73602083 M
73648273 M
37569533 M
38096281 M
73500677 Madpoo mismatch, 
74049323 M

[Mark Rose]

Yeah, I wish I were more familiar with Microsoft databases so I could just help write these things.

Tuning MySQL is basically a hobby of mine.

Originally Posted by airsquirrels

I have 17 Verified, 2 bad, and 12 mismatches from that card. If anyone wants to TC. So far all the 7xM tests that card has done have mismatched.

38096281 M
38133923 M
38324009 M
38324521 M
39299047 M
39334423 M
73602083 M
73648273 M
37569533 M
38096281 M
73500677 Madpoo mismatch, 
74049323 M

I grabbed the low exponents, except for 37569533 and 39334423 which were already assigned.

I've already been working on 73500677 for a couple of days. It should be done about Tuesday. MadPoo has 74049323 and I imagine he'll pick up the other two 73M exponents soon.

The 73M exponents were all done by a user with a crappy machine. Most of the exponents I've DC'ed by that user have mismatched.

[Madpoo]

Originally Posted by Mark Rose

...I've already been working on 73500677 for a couple of days. It should be done about Tuesday. MadPoo has 74049323 and I imagine he'll pick up the other two 73M exponents soon.

The 73M exponents were all done by a user with a crappy machine. Most of the exponents I've DC'ed by that user have mismatched.

Yeah, I picked up a few triple-checks recently where AirSquirrels was the most recent test. I don't remember why I picked those in particular, but I think there's just a couple left.

I looked at my past 3 months of LL results and I have 12,107. But the HUGE asterisk there is that includes a bunch of extra checks on smaller exponents... Excluding those, I have 1832 in the past 3 months, but working in different ranges.

Looking at it another way, over the past 90 days, I have 150208.30 of GHz-days and Airsquirrels has 164551.29 so he has me beat fair and square. LOL

[airsquirrels]

Our power output honestly comes from a pretty small number of systems. I know I'm only running 25 total systems, albeit stuffed to the brim with GPUs. I often wonder what would happen if someone like NVIDIA decided to just burn in a rack full of P100 systems on CUDALucas. They could easily turn in 100+ results a day.

[Madpoo]

Originally Posted by Mark Rose

Also, Madpoo, how many machines have now completed an LL test without having a DC currently assigned or completed? That would be a nice number to have a countdown of.

It's kind of a fuzzy metric... there are a *lot* of old machines that no longer participate, so, for example, they're not going to get DC work as part of the new assignment rules. We're on our own to go through those.

I should do more work on those systems and see what other things I can get out of them to see if they're good or bad. Right now with zero bad/zero good, the only other thing to look at is if they had suspect results (or mismatches for any other reason). Otherwise it's just a matter of doing a verification on one of them to see if it's good or bad, but even then it only helps find new bad machines if they have more than one solo-checked exponent.

The thing is, we'd be doing all of that testing and probably get a far lower success rate in finding bad tests than we currently can by going after the machines with some kind of known-to-be-bad track record.

At some point we'd exhaust that kind of work so we'd have to go searching for the undiscovered bad systems, I just don't know if we're there quite yet.

There are 17330 systems with zero good, zero bad and at least one result that has never been DC'd. That's a lot.

On the other hand, there are only 363 with zero good/zero bad but at least one mismatch or suspect result. That's more manageable as a start, and 318 of those have more than 1 solo-checked result, so at least if we mismatch the test run, there will be at least one more we can take advantage of our new found knowledge.

Systems with a mismatch in their history doesn't really mean they're bad...I'd have to look at their mismatch and see if their result was first or second or look at the other machine to get a better idea.

33 of those have a suspect result which also matches how many mismatches they have. Those 33 machines represent 287 solo checked exponents, so who knows... maybe?

[ATH]

Originally Posted by Madpoo

There are 17330 systems with zero good, zero bad and at least one result that has never been DC'd. That's a lot.

How many of these were active in the last 6 months or 12 months? Unless it is too complicated to search for.

[airsquirrels]

One more where the 1080 is likely to have been wrong. I did a driver update and hopefully stabilized the card, time will tell.

73381873

[Mark Rose]

David, did you queue that list of TC work? I don't see assignments for those exponents. If you have queued them, we should get them assigned so that we don't duplicate work.

[airsquirrels]

Originally Posted by Mark Rose

David, did you queue that list of TC work? I don't see assignments for those exponents. If you have queued them, we should get them assigned so that we don't duplicate work.

I think I've done this. That process was significantly more of a pain than expected since it appears to only let me reserve a few exponents per worker. Is there not a better way?

[Mark Rose]

Originally Posted by airsquirrels

I think I've done this. That process was significantly more of a pain than expected since it appears to only let me reserve a few exponents per worker. Is there not a better way?

That's odd.

Maybe try adding `UnreserveDays=99999` to prime.txt.

Unless someone else has ideas?

[LaurV]

When I used to have "tons of work" reserved in the past, I always had to add to prime.txt the two magic lines

MaxExponents=100
UnreserveDays=365

(or more, depending on your work list), before adding the work (using worktodo.add or alternative, stopping and exiting P95 and directly modifying the worktodo.txt file), to avoid P95 unreserving them at the first server check. Also, add them with the N/A key if you want to work them "unassigned". If they are not assigned, and if they are in assignable range, then add them without a key and P95 will try to reserve them at the first server connection (i.e. get a valid key for them). If you add them with N/A key, them P95 will not try to assign them, and it will not get a valid key, but work them in "unassigned" mode. However, at the end of each, the results will be properly reported to the server.