GPU to 72 status...

[Prime95]

Some DB queries show something is wrong with either our scripts or Ryan's scripts. We'll figure it out and correct the excessive P-1 assignments. May take a week.

[petrw1]

Originally Posted by Prime95

Some DB queries show something is wrong with either our scripts or Ryan's scripts. We'll figure it out and correct the excessive P-1 assignments. May take a week.

Whoa....that was quick.
Almost all but about 1,000 of the P-1 are now released.

[KEP]

Originally Posted by Prime95

Chris, should the appearance of Ben Delo change our TF or P-1 strategy?

According to the Top Primality Producers page it appears Ben Delo is contributing 800000 GHz-days/month. About double curtisc's output. I assume this has greatly eaten into the reserve of available LL/PRP exponents that have been TF'ed and P-1'ed.

It's all about balancing TF resources with the LL/PRP resources. I'm wondering if it is time to reserve further ahead of the wavefront and do one bit less of TF. You are the one with the best handle on the capabilities of the TF'ers and prime testers.

If P-1 resources cannot keep up, then I should change PrimeNet to save a pool of P-1'ed exponents for manual reservations. Non P-1'ed exponents given to custisc, Ben Delo, or other prime95 users will get some P-1 done by the assignee.

I have just moved some resources to P-1, actually I'm going to do 51/p-1 tests a day in the 92.3M range, but this "sticky" situation made me remember - back in the days (2004 or something) when trial factoring were not done very deeply, the user had to first do some trial factoring and P-1 factoring before the actual primality test was begun - maybe the time has come to force trial factoring and/or P-1 before the actual test start

[chalsall]

Originally Posted by KEP

...but this "sticky" situation made me remember - back in the days (2004 or something) when trial factoring were not done very deeply, the user had to first do some trial factoring and P-1 factoring before the actual primality test was begun - maybe the time has come to force trial factoring and/or P-1 before the actual test start

Indeed. I took some pride in being a small part of the "team" getting to the situation such that when someone was given a LL test to do, it had already been "optiomally" TF'ed (a *huge* thanks to James for coming up with an analysis telling us where that actually was ), and P-1'ed -- usually "well" (read: run with several GB of RAM).

And so everyone knows... Now that George or Ryan has released several thousand P-1 assignments which have already been TF'ed to 77, we're not in that bad of a situation.

More of the TF'ers will need to spend some time doing the "last bit" (read: TF'ing to 77), but we do have some breathing room. And, as always, this is entirely a volunteer effort -- if people prefer to do "breadth-first" (read: only going to 74 or 75), that's fine.

Even if someone is given an LL assignment not yet P-1'ed, the worst case is they spend a day or so first doing a P-1. Most likely they won't have set the memory to be the appropriate 8 GB or so, so only "Stage 1" of the P-1 test will be done. Sub-optional, but not the end of the world.

Lastly, my understanding is that George et al already have Primenet preferring to hand out work with P-1 already done. There's some kind of a "sort by [P-1 done] desc, [Exponent] group by int([Exponent] / 1000000)" clause in the assignment code.

But, since we're talking about this...

For any GPU72 TF'ers currently doing DCTF, it would be useful if you'd switch to LLTF -- at least for a month or two. Ideally "Pledging" to 77, or choosing "Let GPU72 Decide". (For historical reasons, "What Makes Sense" will not ever assign work to a factoring depth beyond the Pledge; LG72D, on the other hand, sets the pledge level.)

[Uncwilly]

Originally Posted by chalsall

For any GPU72 TF'ers currently doing DCTF, it would be useful if you'd switch to LLTF -- at least for a month or two.

Is there a way yet to do that in Colab instances, or should we just generate a new assignment key/notebook name?

[chalsall]

Originally Posted by Uncwilly

Is there a way yet to do that in Colab instances, or should we just generate a new assignment key/notebook name?

You'll be able to do that within two hours...

[KEP]

Originally Posted by chalsall

...but we do have some breathing room...

Whew, it appears that I have been to long away from doing P-m1, to really know how long it takes. The Xeon (12 cores) completes 9.2 P-m1 each day, using ~35500 MB RAM in Stage 2 of 110000 MB available. The i5-4670 is (fingers crossed) gonna do about 5 a day. I'll upload new results once a week, and hopefully this will keep the breathing room intact for a little longer

[storm5510]

Originally Posted by chalsall

...For any GPU72 TF'ers currently doing DCTF, it would be useful if you'd switch to LLTF -- at least for a month or two. Ideally "Pledging" to 77, or choosing "Let GPU72 Decide". (For historical reasons, "What Makes Sense" will not ever assign work to a factoring depth beyond the Pledge; LG72D, on the other hand, sets the pledge level.)

LG72D? Sorry!

I've been running 2⁷⁵ LLTF for a while now and it is almost always in the 95M area. With all due respect, there seems to be some stagnation going on. Looking at the Work Distrubition Map, it appears that the majority of P-1 work is happening in the 92M range. Is there an urgency here that I am not seeing?

[chalsall]

Originally Posted by storm5510

LG72D? Sorry!

Sorry... Didn't define the acronym. LG72D == "Let GPU72 Decide".

Originally Posted by storm5510

Is there an urgency here that I am not seeing?

Up until early this morning, there was...

There continues to be a bit of an issue with people coming in and reserving *very* large quantities of P-1 work manually. For reasons I don't fully understand, this is allowed without constraint -- even for Anonymous users (read: those who haven't even bothered to log into Primenet). And Manual Assignments remain for 180 days (if not "intervented").

But as I said above, if people enjoy doing "breadth-first", that's cool. Every "bit" helps...

Edit: Oh... Also... We TF'ers /do/ now have ~60,000 candidates already P-1'ed but not yet "optionally" TF'ed to "work". I consider it our responsibility to keep ahead of the LL wavefront, even the "Cat 4 Churners"...

[James Heinrich]

Originally Posted by chalsall

if people enjoy doing "breadth-first", that's cool

Although some of us take that to extremes... all exponents <10000M have been fully TF'd to 57-bit now.

[storm5510]

Originally Posted by chalsall

...There continues to be a bit of an issue with people coming in and reserving *very* large quantities of P-1 work manually. For reasons I don't fully understand, this is allowed without constraint -- even for Anonymous users (read: those who haven't even bothered to log into Primenet). And Manual Assignments remain for 180 days (if not "intervented").

But as I said above, if people enjoy doing "breadth-first", that's cool. Every "bit" helps...

Edit: Oh... Also... We TF'ers /do/ now have ~60,000 candidates already P-1'ed but not yet "optionally" TF'ed to "work". I consider it our responsibility to keep ahead of the LL wavefront, even the "Cat 4 Churners"...

First, let me correct something I wrote above. I wrote I was running to 2⁷⁵. That should have been 2⁷⁴. Sorry.

Taking large amounts of P-1 work makes no sense. It's simply asking for a rollover after a given span of time because of a person having far more than they can run. I suppose they think they can run them just as fast as TF.

A reservation limit can be gotten around with a recursive batch file. They run a bunch of loops, and there you are. Thousands of assignments in just a few minutes.

Thank you for your replies.