GPU to 72 status...

[chalsall]

Originally Posted by Prime95

Not sure what assignments you are seeing and what their story is.

Take a look at this query, then search for "Ben Delo". Currently I'm seeing 91 assignments, all assigned within the last two weeks.

[Prime95]

Originally Posted by chalsall

Take a look at this query, then search for "Ben Delo". Currently I'm seeing 91 assignments, all assigned within the last two weeks.

I mis-remembered. I wrote special server code that exempted Ben from mandatory first-time double-checks --- Amazon instances are super-reliable.

Since you think GPU72 will run more smoothly if his new instances are assigned cat 1, I've changed the get-assignment code to make that happen.

[chalsall]

Originally Posted by Prime95

Since you think GPU72 will run more smoothly if his new instances are assigned cat 1, I've changed the get-assignment code to make that happen.

Thanks, George.

And with regards to "run more smoothly", we're at a situation where LL Cat 3 and 4 are going to be given work seriously "sub-optimally" TF'ed, nor even P-1'ed. But many of those assignments are going to expire.

Cat 2 will be OK for another few months, getting work taken to at least 75, most likely 76 or even 77. We have enough of a buffer/firepower for Cat 1/0 for a good couple of years.

This is not the end of the world. It's "sub-optimal" for overall throughput, but rarely do we live in an optimal world.

And let's not forget: this is GIMPS, not GIMFS.

Thanks to Ben (et al), we're likely to find the next MP much sooner than previously estimated.

(Now if Ben would like to invest some GPU TF'ing firepower, that would be /really/ cool... )

[ixfd64]

Any chance GPU to 72 will start giving assignments above 100m soon? I'm starting to get LL assignments in the 100m range that have only been factored up to 74 bits.

[chalsall]

Originally Posted by ixfd64

Any chance GPU to 72 will start assignments above 100m soon? I'm starting to get LL assignments in the 100m range that have only been factored up to 74 bits.

Yeah... We really need to discuss this as a team, and figure out what's the best thing to do.

Basically, because of a certain individual, GIMPS LL throughput has more than tripled in the last three months (!). This is ***amazingly*** cool news! Thanks Ben!

However, this has completely messed with the goal of having all LL assignments "optionally" TF'ed, and P-1'ed (ideally "well", by P-1'ing "specialists").

Currently, it is "optimal" to TF to 77 "bits", but with the current TF'ing "firepower" we're only producing about 50 a day; to stay "steady-state" we would need to produce 900 a day!

We do have ~60,000 candidates already ready for LL assignment, but that's only going to last us two months. What I've currently got GPU72 doing is giving out work in such a way as we "chase" ahead of the Cat 2 assignments such that they are optimally TF'ed and P-1'ed. However, it won't take long until Cat 2 also gets into the 10xM ranges.

Then, I don't know... Should we start releasing at 75, hoping to occasionally get to 76?

And/or, should we bring in work in the 10xM ranges, and start bringing them up (many are still only at 72 bits).

I would really welcome suggestions as to what people want to see happen/thinks makes sense.

And, as always (but particularly now), if anyone has any GPU compute they could bring to bear, it would be much appreciated!

Thoughts?

[storm5510]

Originally Posted by ixfd64

Any chance GPU to 72 will start giving assignments above 100m soon? I'm starting to get LL assignments in the 100m range that have only been factored up to 74 bits.

Ditto!

In my humble opinion, This project should be doing only TF work. P-1's and LL's should not be done here. Those should be reserved from Primenet.

[chalsall]

Originally Posted by storm5510

In my humble opinion, This project should be doing only TF work. P-1's and LL's should not be done here. Those should be reserved from Primenet.

And that's pretty much the case currently.

For historical reasons, LL/DC work is available through a "proxy", but now-a-days it's simply a record-keeping function -- all work is actually fetched from Primenet.

P-1'ing is still available directly, but we only reserve a few hundred candidates at a time for that -- and in fact at a higher level (read: more computationally expensive) than what's available directly from Primenet.

[Prime95]

Originally Posted by chalsall

Yeah... We really need to discuss this as a team, and figure out what's the best thing to do.

Currently, it is "optimal" to TF to 77 "bits", but with the current TF'ing "firepower" we're only producing about 50 a day; to stay "steady-state" we would need to produce 900 a day!

Then, I don't know... Should we start releasing at 75, hoping to occasionally get to 76?

And/or, should we bring in work in the 10xM ranges, and start bringing them up (many are still only at 72 bits).

Thinking out loud. If you took all the GPU72 firepower what bit level could produce 800-900 results a day? Make that your target.

I'm guessing that pre-Ben you thought GPU72 could stay ahead of the wavefront TF'ing to 2^77. Post-Ben we have 4x the LL/PRP power means you need to drop down 2 bit levels and make 2^75 your target.

I'd say bring in the some of the 10xM ranges and get them up to 2^74 at a minimum.

It is all so complicated!
1) The daily cat 1/2/3/4 thresholds are changing faster than they used to. If you TF the very smallest cat 4 exponents and they don't get assigned, they become the highest cat 3 exponents and won't get assigned for quite some time.
2) TFing the highest exponents nets the most savings, but cat 4 churn means you may have time to find a potential factor at a later date.
3) Optimal is to maintain a supply of decently TF'ed exponents at 4 wavefronts that are moving targets.

Don't worry that we reaching optimal TF levels. The DCTF group will have years to achieve that with much better GPUs.

[chalsall]

Originally Posted by Prime95

It is all so complicated! ...

Don't worry that we reaching optimal TF levels. The DCTF group will have years to achieve that with much better GPUs.

Yeah; that goal was abandoned some time ago... I only today "did the numbers" with current heuristics, and realized just how little time we have. (And, based on the graphs, Ben appears to continue to add "kit"! Wow!)

My current plan is to continue keeping Cat 2 fed as best we can until they're up into the 100M range. We have a reasonable buffer up in 99M to keep them feed, but as you note that is climbing surprisingly quickly.

Then we'll start getting 9xM ready to at least 75, and over the next couple of days I'll bring in some 10xMs to bring up as best we can.

One thing which might be helpful is if Primenet would assign work sorted by TF level desc (as with P-1, grouped by 1M range). That way GPU72 workers could focus on the high end of each range, as the wavefronts race towards us.

As always, feedback and suggestions greatly appreciated.

P.S. A real pitty Kaggle is completely out of commission, and Colab appears to be seriously constraining us...

[petrw1]

We've lost some big TF users too.

The Judger...though he comes and goes
Air Squirrels
Flash JH
Jon Pace
LaurV ...maybe when he's back from Oz.
Ryan Propper


and a few others.

[chalsall]

Originally Posted by petrw1

We've lost some big TF users too. ... and a few others.

Yup. But even if they all returned, we still wouldn't be able to keep up with "optimal". As said before, this isn't the end of the world. And it does mean the next MP is going to be found /much/ sooner!

Oh, and just for the record, Jon Pace is continuing to contribute ~3.5 THzD/D of work.