Folding on Cloud instructions (3 mil ppd)

[Foldbert]

This is awesome, thanks Knish for taking the time to write out such detailed instructions for us cloud neophytes. I found the GCP instructions to be slightly out-of-date, for example I had to execute Step 3 by not just clicking on Compute Engine, but proceeding one step further in (Create an instance?) before it initialized instancing on my account. Before that, I was not able to access "GPUs (all regions)" in Quotas. I think one other GUI thing had shifted, but it was easy to figure out. Azure was a breeze.

~200m points of folding compute at no required cost is great. Hopefully more people see this thread -- took me all of an hour to get both set up, and that was with the hitch on GCP quota set up.

[Knish]

Foldbert, thanks very much for your feedback! Another user brought up the GPU all regions thing, and I have fixed by changing that instruction to "regions" instead of typing in "GPU," and updating more details about the quotas. With regards to initializing for step (3), I took another look and it still seems to be fine. I believe you inadvertantly hopped ahead to step (14) in VM Creation. That's where u must click on 'create.' I verified that compute engine gets initialized at step (3) and that's what's necessary before asking for a quota increase. Hopefully the new edits make this more clear.

[ptchernegovski]

Have had GCP stop folding overnight. When checking the logs, it is telling me:
"No CUDA or OpenCL 1.2+ support detected for GPU slot 00: gpu:0:4 TU104GL [Tesla T4] 81$"

Anyone else getting this issue?

[Knish]

it's an easy fix. u just have to redo 2 steps:
sudo apt-get install gcc make linux-headers-$(uname -r)
and then
sudo ./NVIDIA ... (whatever version you downloaded. it should still be sitting in the VM's directory)

then sudo reboot and you're back in business

why this happens I do not know... I don't think it happens more than once per the lifetime of your VM, which is good news, but when the VM is preempted, something gets fouled up with the gpu driver so it has to be reinstalled.

[ptchernegovski]

Awesome, thanks. Will give it a try now

[will.allison]

do you know why you need to use Sudo for each instruction? I am afraid that create some permission problems when someone without root permissions whats to make use of those drivers...

[ajm]

You use sudo to launch a command as an admin, which your normal user account is not.
When you do, the system asks for your password (and keeps track of that command).
Someone without a proper password can't use sudo.

[will.allison]

I know what is the purpose of Sudo, but my point is that if you use Sudo, all the execution is going under that mode, and my concern is that, if some files are generated from there (an installation normally does), then I am not sure if a normal user can make use of the installed software.... logically thinking I would say no

[ajm]

Sorry! I don't know, but if someone is keen to try and test that, I'd be interested in the result.

[Joe_H]

From the commands shown, sudo is needed because they are writing or modifying system level files. A typical user account does not have that permission. Typical installs of system level files include execute status and sometimes read status to either all users or all users in specific groups.

[Foldbert]

I spent 45 minutes this morning trying to figure out why GCP suddenly stopped working, and ended up in the same place of installing kernel source headers and getting the Tesla drivers re-installed. Guess I should have checked here first!

[PantherX]

I was looking for the PPD of the P4.

~300k ppd

300k PPD for the Pi4?
Last time I ran a 6-core I5, it ran well below 100k PPD.
What sorcery is this?

It isn't Raspberry Pi 4, it is a type of instance that is available on GCP (Google Cloud Platform): https://cloud.google.com/compute/docs/gpus which would be the Tesla P4 GPU.

[toTOW]

Does anyone know the typical PPD we could expect for a T4 GPU in GCP ? Also, did you notice consistent PPDs, or do they vary ?

[Knish]

it's been a pretty steady 800k to 1 mil ppd for several months, but checking now it's got a santa clause P13437 WU so it's projecting 1.4 mil

[toTOW]

It is very disappointing for a GPU that has the same number of SP as the 2070 Super ...

By the way, I have never been able to create a spot instance on Azure, so I've given up.