Gpuowl v6 issue on free-tier Colab

[kriesel]

Originally Posted by moebius

-max Alloc 1000 is sufficient for PRP

For exponent < ~300M, yes; depends on exponent. See attachment of https://www.mersenneforum.org/showpo...3&postcount=24

[tdulcet]

Originally Posted by tdulcet

It does not look like anyone has reported the issue to Google yet

@danc2 was kind enough to report the issue to Google on Friday: https://github.com/googlecolab/colabtools/issues/3592, but we are still waiting for a response...

In the meantime, people who were using our GPU notebooks may want to temporarily switch back to the old version, which ran CUDALucas. CUDALucas is of course slower than GpuOwl and it only supports LL tests, but it uses CUDA, which currently still works on Colab. Here is a permalink: https://github.com/tdulcet/Distribut...ColabGPU.ipynb.

If the problem persists, I am considering restoring CUDALucas support to the latest version of our GPU notebook. For the long term, we also might want to look into updating CUDALucas to support PRP tests. The PRP algorithm is actually simpler with than the LL algorithm (no minus two each step), so this actually should not be too difficult to adapt the existing code. On the other hand, adding Gerbicz error checking (GEC) support would likely be prohibitively difficult, as it would require implementing a new fundamentally different algorithm from scratch.

I know many of the people in this thread wanted to do P-1 testing, so I would also be open to adding support for CUDAPm1, if there was sufficient interest. I recall there were issues in the past with CUDAPm1 on Colab, but as with CUDALucas (which my install script dynamically fixes), I am sure those issues could be resolved as well if needed.

[moebius]

I don't know if two many gpuowl instances are the reason or it is the possibility to generate cryptocurrencies like Ehtherum with OpenCl.
In any case, mersenne.org is purely a research project. and google colab research has become nothing more than a purely commercial product. Google has not managed to keep up with the state of the art, only with the current level of inflation. There will never be OpenCL at googlecolab for free again.
Whart about a Nvidia H100 instance that can be tested for free for a month or so?

[Mark Rose]

Originally Posted by moebius

I don't know if two many gpuowl instances are the reason or it is the possibility to generate cryptocurrencies like Ehtherum with OpenCl.

Ethereum mining ended last summer.

[kriesel]

Originally Posted by moebius

There will never be OpenCL at googlecolab for free again.

Mprime, mmff and mfaktc continue to run about daily for me (in the USA) on Google Colab Free, multiple instances each. So there are alternatives. The free GPU (Tesla T4) is better utilized in TF anyway.