GPU to 72 status...

[petrw1]

Originally Posted by Chuck

I'm using Chris' tool and getting 24 hour run times.

But paid ....

I'm getting 8 hours free and not using his tool.
Seems like most other are getting much, much less time including Chris.

[Flaukrotist]

I also have been using Chris' tool since May until four weeks ago. I liked it a lot for its simplicity but while I was getting 2.5 hours of GPU (free Colab user) last year consistently it was reduced to 0.75 hours this year. Worse are the intermediate CPU-only session times which lasted for 12 hours last year and stopped after 10 to 60 minutes this year.

Now I created my own Colab scripts for running mprime for P-1 and getting 12 hours session time (if I manage to click an "I am still there" checkmark after 6 hours) again. In parallel, I can run a second notebook (only if I get a GPU) with mfaktc, getting assignments from gpu72 by Misfit. These sessions have a very consistent starting time every day when the GPU starts being available and last for 4-6 hours. By the way, I serve two accounts with different email addresses from the same google drive.

The session times have started being longer and uninterrupted from the second that I abandoned Chris' tool. So I also think that they restrict Chris' tool somehow. I am now running a very similar setup with own scripts with very different runtimes.

[chalsall]

Originally Posted by Flaukrotist

The session times have started being longer and uninterrupted from the second that I abandoned Chris' tool. So I also think that they restrict Chris' tool somehow. I am now running a very similar setup with own scripts with very different runtimes.

I agree with you that constraining by Notebook source is likely part of the allotment algorithm.

When some people started saying they were getting less-and-less GPU (and then later, CPU) compute others were reporting no changes. Initially, only a few GPU72 Notebook users were reporting problems, then more and more.

Interestingly, others who had "rolled their own" reported no issues. Then sometimes even they started reporting constrained GPU resources.

Very, very strange (or, at least, a very complex "black box" with many dimensions).

P.S. Interestingly, there are several people who are currently getting good GPU72 Notebook utilization. I know one is paid, don't know about the others. I continue to get ~30 minutes per day of GPU per account each day; almost always a T4.

P.P.S. Yes Uncwilly, reducing the checkpoint file upload to every five minutes is modeled.

[chalsall]

Wow! Thanks a lot Oliver!

1.3 MILLION GHzDays in less than a week!

[Chuck]

I wonder what are Oliver's ongoing intentions for TF? He has cleared 104M, 105M and 106M up to 76 and is working on 107M.

[chalsall]

Originally Posted by Chuck

I wonder what are Oliver's ongoing intentions for TF? He has cleared 104M, 105M and 106M up to 76 and is working on 107M.

You would have to ask him... But, yeah... I've been watching with a slack jaw as he brings an incredible amount of compute forward!

Thanks, Oliver!!!

[Uncwilly]

Must be a slow week at the NSA.

[Flaukrotist]

Is there any reason why the factors in the 119M range are not included in the plot on my personal statistics page about GHs Days per Day of work saved? The last bar is for Feb 24 and after that I found 9 factors which are all in 119M range. They are included in the factors found list correctly though. The Trial Factoring Cost per Factor found page also doesn't list anything above 109M.

[chalsall]

Originally Posted by Flaukrotist

Is there any reason why the factors in the 119M range are not included in the plot on my personal statistics page about GHs Days per Day of work saved?

Probably because of some "tech debt".

GPU72 was supposed to be temporary. I never imagined we'd be into the >100M range when the code was started. There are many stupid artificial "where" statements sprinkled throughout the code.

I need to do a serious refactoring on the code stacks sometimes really-soon-now.

Unfortunately, I also have to do "real work" to pay for this. And the gods are currently a bit angry, so I'm having to deal with that as well...

If it helps at all, while the graphs and reports may not be sane, the work assignment and recording code-paths are sane.

[Flaukrotist]

Originally Posted by chalsall

Probably because of some "tech debt".

GPU72 was supposed to be temporary. I never imagined we'd be into the >100M range when the code was started. There are many stupid artificial "where" statements sprinkled throughout the code.

I need to do a serious refactoring on the code stacks sometimes really-soon-now.

Unfortunately, I also have to do "real work" to pay for this. And the gods are currently a bit angry, so I'm having to deal with that as well...

If it helps at all, while the graphs and reports may not be sane, the work assignment and recording code-paths are sane.

I don't consider this an urgent or important issue. It is just an observation I thought would be better to bring to your attention in case you didn't know and in case you have some spare time to spend.

I am truely grateful for you and all other people that keep this great project alive and improve it in the way you do with all the personal effort involved. Thanks a lot.

[chalsall]

Originally Posted by Flaukrotist

It is just an observation I thought would be better to bring to your attention in case you didn't know and in case you have some spare time to spend.

And bug reports which are more detailed than "it doesn't work" are greatly appreciated. Seriously.

Originally Posted by Flaukrotist

I am truely grateful for you and all other people that keep this great project alive and improve it in the way you do with all the personal effort involved. Thanks a lot.

I am equally grateful for this community. There's some very cool talent around these-here-parts. It's wonderful being able to contribute a very small gear "in the machine".

P.S. BTW, while I'm "radiating"... The GPU72 Colab TF/P-1 Notebook now uploads P-1 results every three minutes.