GPU LL Testing FAQ

[retina]

Originally Posted by ewmayer

In looking at the GPU subforum through a n00b user's eyes, it strikes me what a mess it is. I want to be able to get to the best practices for my target GPU/OS in post 1 of a "GPU how-to" thread. This thread has 2 problems in that regard:

1. Title, "GPU LL Testing FAQ", reflects the old days when nVidia was the only game in town - it should be retitled to reflect that it's specific to CudaLucas on nVidia gfx cards, and "users of ___ cards should visit the ____ thread(s)".

2. Whatever was initially posted in Post #1, as a new user I expect to see either a list of, or link to, a Best Practices guide right there, and to have same updated on a running basis to reflect changes in Best Practices and/or new editions of hardware and software of the particular family covered by the thread. The PDF in Post #1 of this thread was last updated in 2012, based on the post edit history.

Your user name is painted in red so I think you just volunteered yourself for a job.

[kriesel]

Originally Posted by ewmayer

1. Title, "GPU LL Testing FAQ", reflects the old days when nVidia was the only game in town - it should be retitled to reflect that it's specific to CudaLucas on nVidia gfx cards, and "users of ___ cards should visit the ____ thread(s)".

Well, unless you want to refer new users to GpuOwl v0.6 or earlier, for LL testing on AMD gpus with 4M fft, which is not adequate for current first-test wavefront assignments, CUDALucas still is the only game in town for LL testing. GpuOwl from v0.7 up does PRP, with some later versions introducing various flavors of P-1, and briefly there was TF with linux & ROCm.
But why not refer new users to https://www.mersenneforum.org/showpo...91&postcount=2, which is
a) as comprehensive a picture of what's useful and where to find it as I know how to make, while keeping it somewhat concise
b) full of links for ready to run code, source, and discussion threads about each,
c) already being actively maintained in a fixed location,
d) only a small part of a large accumulation of reference material laid out at https://www.mersenneforum.org/showthread.php?t=24607

Originally Posted by retina

Your user name is painted in red so I think you just volunteered yourself for a job.

I second the nomination, although if Uncwilly or someone else wishes to volunteer, that's fine too.

[ewmayer]

Originally Posted by retina

Your user name is painted in red so I think you just volunteered yourself for a job.

Happy to do it, just don't want to step on anyone's toes. My aim is this: Once in the GPU subforum, a user with hardware X under OS Y should need no more than 2 clicks to get to a comprehensive how-to guide for their platform:

Click 1: GPU testing FAQ, post #1 has list of hardware/code-to-use links;

Click 2: Once in relevant thread (or online readme) for their platform, getting to an up-to-date Best Practices guide should need no more than 1 added click on a link in Post #1 of the thread.

Re. Ken's LL-vs-PRP note, I consider that a red herring - if a user has an AMD card, they get pointed to that how-to thread and find they'll be doing PRP testing, which offers the same "should you happen to win the lottery" credit and is more reliable, to boot.

Ken, we greatly appreciate the effort you put in to the how-to-guide thread you link, but note that violates my above 2-click criterion:

Click 1 into the thread. In Post #1 (good!) I see a link to a PDF (not so good);
Click 2 on the PDF link. My browser displays a "what should firefox do with this file?" pop-up. Since I expect I may be consulting the guide repeatedly, I want "Save file", which happens to be the default on my FF setup, otherwise there would be one more click. Hit <return>, that counts as
Click 3 - Browser throws up a "where to save" directory tree pop-up. Luckily for me it happens to be pointing to Downloads, which is fine for now. Hit <return>, that counts as
Click 4 - Now I need to open the PDF. On my Mac, mouseover to Finder in the sidebar,
Click 5 - Finder opens, it's pointing elsewhere than Downloads. Scroll up to Downloads,
Click 6 - double-click to open Downloads folder. Luckily it's sorted by date, so the just-DLed M-prime Hunting Software PDF is the top item.
Click 7 - double-click to open PDF. Whoa, that font is tiny. Scroll up to View options pulldown, scroll down to Sidebar, right-arrow, scroll down to Hide Sidebar, hit <return>. Lucky for you I'm only count all that mousing, scrolling and right-arrowing as
Clicks 8,9 - Sidebar closed, page auto-resizes, but is still way to small to be readable. Hit Ctrl+ to zoom in, that counts as
Click 10.

I hope the problem-as-I-see-it is evident by now. :) Long story short, the contents of that PDF should form the body of Post #1 in the thread.

[nomead]

Originally Posted by ewmayer

1. Title, "GPU LL Testing FAQ", reflects the old days when nVidia was the only game in town - it should be retitled to reflect that it's specific to CudaLucas on nVidia gfx cards, and "users of ___ cards should visit the ____ thread(s)".

Just a quick note regarding this. After the recent speedups, GpuOwl has been the way to go on nVidia cards too, unless there's a specific need to run LL (DC, for example). But for first time tests, GpuOwl all the way.

All the years of FAQs and discussions of course still point to CudaLucas and it is likely to take a while for the established mindset to change according to disruptive changes in relative performance.

[kriesel]

Originally Posted by nomead

Just a quick note regarding this. After the recent speedups, GpuOwl has been the way to go on nVidia cards too, unless there's a specific need to run LL (DC, for example). But for first time tests, GpuOwl all the way.

All the years of FAQs and discussions of course still point to CudaLucas and it is likely to take a while for the established mindset to change according to disruptive changes in relative performance.

Note that some older NVIDIA gpus are incapable of running Gpuowl. It requires OpenCL 2.0 support. So they default to CUDALucas and perhaps LL DC. (They are generally not good candidates for P-1 due to memory size and CUDAPm1 characteristics.)

And there's a need for early PRP DC to verify expectations of a low error rate in practice, for first-test PRP with GEC, as implemented, in each of mprime/prime95, Gpuowl, and Mlucas. With an expected very low error rate, it will likely take a large number of verifications.

[kriesel]

Originally Posted by ewmayer

Happy to do it, just don't want to step on anyone's toes. My aim is this: Once in the GPU subforum, a user with hardware X under OS Y should need no more than 2 clicks to get to a comprehensive how-to guide for their platform:

Click 1: GPU testing FAQ, post #1 has list of hardware/code-to-use links;

Click 2: Once in relevant thread (or online readme) for their platform, getting to an up-to-date Best Practices guide should need no more than 1 added click on a link in Post #1 of the thread.

Re. Ken's LL-vs-PRP note, I consider that a red herring - if a user has an AMD card, they get pointed to that how-to thread and find they'll be doing PRP testing, which offers the same "should you happen to win the lottery" credit and is more reliable, to boot.

Ken, we greatly appreciate the effort you put in to the how-to-guide thread you link, but note that violates my above 2-click criterion:

Click 1 into the thread. In Post #1 (good!) I see a link to a PDF (not so good);
Click 2 on the PDF link. My browser displays a "what should firefox do with this file?" pop-up. Since I expect I may be consulting the guide repeatedly, I want "Save file", which happens to be the default on my FF setup, otherwise there would be one more click. Hit <return>, that counts as
Click 3 - Browser throws up a "where to save" directory tree pop-up. Luckily for me it happens to be pointing to Downloads, which is fine for now. Hit <return>, that counts as
Click 4 ...
I hope the problem-as-I-see-it is evident by now. :) Long story short, the contents of that PDF should form the body of Post #1 in the thread.

Ernst:
If you just open the pdf in your browser, that's 3 clicks. Saving and futzing with it is your option, adding numerous clicks.
It covers several applications, so divide by however many applications you intend to use.
(Saving the pdf is ok for the short term, but my updates will be on my blog, not in a static saved copy or a designated thread's post 1 that I would need moderator status I don't have, to update.)
If someone can't bear more than two or even 5 clicks to get to such info, I doubt they'll have the patience for long term GIMPS participation.
But efficiency is good for everyone. I have no objection to someone pasting a copy of what I've written into the relevant Post #1s, as long as it includes a link to the maintained source. Who that is and how current they'll keep it is TBD.

AND:

your post implicitly raises a valuable question: what ARE the best practices? See https://www.mersenneforum.org/showth...225#post535225 which will hopefully generate some discussion and agreement on that.

[nomead]

Originally Posted by kriesel

Note that some older NVIDIA gpus are incapable of running Gpuowl. It requires OpenCL 2.0 support. So they default to CUDALucas and perhaps LL DC. (They are generally not good candidates for P-1 due to memory size and CUDAPm1 characteristics.)

Yeah... even there, extra clarification could be useful. For example, the Wikipedia OpenCL page states about OpenCL 2.0 support that ...

(2017+) incomplete Evaluation support: Nvidia Kepler, Maxwell, Pascal, Volta and Turing GPU's (GeForce 600, 700, 800, 900 & 10-series, Quadro K-, M- & P-series, Tesla K-, M- & P-series) with Driver Version 378.66+

Then, running clinfo on Linux, recent card and fresh drivers, gives info like this (only a few relevant lines pasted in) :

  Platform Version                                OpenCL 1.2 CUDA 10.2.109
  Device Name                                     GeForce RTX 2080
  Device Version                                  OpenCL 1.2 CUDA
  Driver Version                                  440.44
  Device OpenCL C Version                         OpenCL C 1.2

So this confuses the heck out of me. 2.0 needed but it's still version "1.2" somehow, and yet it works.

[henryzz]

Much better than pdfs would be a page on the wiki. pdfs as attachments are annoying. They never seem to open in my browser when I want them to. They force me to download them.

[kriesel]

PDFs are a readily available form of output for both spreadsheets and word processing software used to organize and format the various information. Working in the tiny window of the forum thread replies, basic or advanced, and trying to get content and format right for a nontrivial document is tedious and frustrating and limiting by comparison. A wiki is yet another location to maintain after getting authorized, learning its edit tools, etc and a time expenditure I choose not to make.

[kriesel]

Originally Posted by nomead

2.0 needed but it's still version "1.2" somehow, and yet it works.

Conversely, my Quadro K4000 does not work.

2019-12-19 23:55:22 ostrich-qk4000 <kernel>:13:9: warning: GpuOwl requires OpenCL 200, found 120
#pragma message "GpuOwl requires OpenCL 200, found " STR(__OPENCL_VERSION__)

And that is with NVIDIA driver v392.58

[ewmayer]

Originally Posted by kriesel

PDFs are a readily available form of output for both spreadsheets and word processing software used to organize and format the various information. Working in the tiny window of the forum thread replies, basic or advanced, and trying to get content and format right for a nontrivial document is tedious and frustrating and limiting by comparison. A wiki is yet another location to maintain after getting authorized, learning its edit tools, etc and a time expenditure I choose not to make.

Ken, we know PDF is your preferred format, and I suggest still retaining as an attachment to the modified forum-thread Post #1.