mersenneforum.org  

Go Back   mersenneforum.org > Great Internet Mersenne Prime Search > Hardware > GPU Computing

Reply
 
Thread Tools
Old 2020-04-22, 13:04   #1
perfectGrease
 
Apr 2020

2 Posts
Default AMD GPU instruction for beginners (request)

Good day Honourable Mathlords!

I hope this message finds you well. I wish not to offend anyone with this artless request, and would feel deep gratitude to anyone deigning to answer.

I'm using GIMPS for almost a year (~2k Ghz days), started from Prime95 on two Windows laptops, then went for linux64 Prime95 on Ubuntu, finally got stuck when trying to add some GPUs.

I'm new to linux and I'm really overwhelmed with the amount of information on this forum. I have a kind-of mini crypto-rig (mainboard, cpu, 3x RX580-8GB, USB stick as HDD, no OS) which I would happily engage in GIMPS if only given a simple instruction how.

I would be endlessly grateful for anyone to make a simple to-do list, containing information on:
- which linux to choose (if simplicity is concerned),
- which GPU mersenne software to use (to be the easiest),
- what else do I have to install (and how?),
- how to configure it for multiple GPUs?

similar to the following 10-step instruction on how to run Prime95 on Ubuntu:

1. Create a bootable Ubuntu USB stick on my Win PC, as described here.
2. Boot the target device from it and make a new minimal Ubuntu installation.
3. Update (sudo apt update).
4. Install SSH (sudo apt install openssh-server).
5. Download Prime95 to my Win PC from here.
6. Unpack and upload it to the target device with WinSCP.
7. Give it writing rights (chmod 777 *).
8. Access target device from Putty.
9. Install Screen (sudo apt install screen).
9. In "Screen" start Prime95 (./mprime -m).
10. Detach the screen (ctrl+a -> d) and voila!

The above takes 2 minutes to prepare for somebody who knows how. I expect the GPU part to be more complex, but not impossible to be summarized in such instruction. I'm already burned after many hours of failing, close to give up, would really appreciate some help.

perfectGrease
perfectGrease is offline   Reply With Quote
Old 2020-04-22, 20:41   #2
kriesel
 
kriesel's Avatar
 
"TF79LL86GIMPS96gpu17"
Mar 2017
US midwest

10010100011112 Posts
Default

Quote:
Originally Posted by perfectGrease View Post
similar to the following 10-step instruction on how to run Prime95 on Ubuntu:

1. Create a bootable Ubuntu USB stick on my Win PC, as described here.
2. Boot the target device from it and make a new minimal Ubuntu installation.
3. Update (sudo apt update).
4. Install SSH (sudo apt install openssh-server).
5. Download Prime95 to my Win PC from here.
6. Unpack and upload it to the target device with WinSCP.
7. Give it writing rights (chmod 777 *).
8. Access target device from Putty.
9. Install Screen (sudo apt install screen).
9. In "Screen" start Prime95 (./mprime -m).
10. Detach the screen (ctrl+a -> d) and voila!

The above takes 2 minutes to prepare for somebody who knows how. I expect the GPU part to be more complex, but not impossible to be summarized in such instruction. I'm already burned after many hours of failing, close to give up, would really appreciate some help.
Welcome. What you gave seems to me a pretty good start on directions for a base system on which to add some gpu GIMPS application(s). The application and directions would vary depending on whether you want to use NVIDIA or AMD gpus, and what computation type you want to perform.

It's not clear what is going wrong for you or what gpu application you may have attempted. You may find what you're looking for or close to it in gpu cloud computing directions. (Includes python scripts that set up applications in linux VMs.) https://www.mersenneforum.org/showth...906#post527906 Some more of https://www.mersenneforum.org/showthread.php?t=24607 may also be useful.

A few things that trip people up sometimes are:

using device number one on a system with one gpu and an application with a zero-based-numbering scheme (first gpu is device 0 or -d 0, not -d 1 there)

A driver or OpenCl installation that's not working yet. NVIDIA cards need NVIDIA's drivers, not the linux nouveaux driver, and some distros put up a valiant fight to keep that change from happening.

Dueling OpenCl drivers from different hardware vendors can be a problem. Sometimes it's just not worth the trouble to get an IGP working in mfakto.

Separate test utilities that show what driver version is installed, and test for function such as CUDA or OpenCl, can be very helpful in determining whether it's a gpu driver issue or Opencl install issue or GIMPS gpu application issue.

In your case, with RX580 x 3, gpuowl for P-1, PRP, or now LL, and mfakto for TF are what to look at. See the relevant lines in the attachment of https://www.mersenneforum.org/showpo...91&postcount=2 for download and discussion thread links, summary specs, etc. Both gpuowl and mfakto use OpenCl. Some hardware is capable of TF but not more. RX580s should handle both mfakto and gpuowl.

Gpu apps don't use the PrimeNet API, which is what allows mprime to get work and submit results automatically. I suggest beginning with manual assignment reservation and result reporting to get things working. https://www.mersenne.org/manual_assignment/ https://www.mersenne.org/manual_result/

Then maybe later looking into client management options like gpuowl's python script, mfloop for mfakto, etc.


Good luck!

Last fiddled with by kriesel on 2020-04-22 at 21:02
kriesel is offline   Reply With Quote
Old 2020-04-22, 21:53   #3
perfectGrease
 
Apr 2020

2 Posts
Default

Thank you kriesel for the fast answer and the warm welcome!
(Let me use my time now to analyze your answer and give it a try).
perfectGrease is offline   Reply With Quote
Old 2020-04-26, 23:30   #4
kriesel
 
kriesel's Avatar
 
"TF79LL86GIMPS96gpu17"
Mar 2017
US midwest

4,751 Posts
Default

When you get your process worked out, please post the sequence here.
You are approximately at where I was 3 years ago.
I'm Windows oriented, so linux coverage is thin.
kriesel is offline   Reply With Quote
Old 2020-08-31, 18:28   #5
M344587487
 
M344587487's Avatar
 
"Composite as Heck"
Oct 2017

3×233 Posts
Default

Can I put mathlord on my CV?
M344587487 is offline   Reply With Quote
Old 2020-08-31, 20:53   #6
moebius
 
moebius's Avatar
 
Jul 2009
Germany

461 Posts
Default

Quote:
Originally Posted by perfectGrease View Post
- which linux to choose (if simplicity is concerned),
- which GPU mersenne software to use (to be the easiest),
A Ubuntu installation on your SSD/HD should work with gpuowl the only gpu-software wich is able to do PRP (and Proofs). Have a look at https://mersenneforum.org/forumdisplay.php?f=171 for answers about requirements and installation issues. You can also use the newest version of Gpuowl (with proofs) with Windows 10 and the executables from user kriesel.

gupowl (with OpenCL) works relativly fast (f.e) with the following grafic cards.

Nvidia A100 [only in my wettest dreams]
AMD Radeon VII (Pro)
Nvidia Tesla-P100
RX Vega64


Last fiddled with by moebius on 2020-08-31 at 21:26
moebius is offline   Reply With Quote
Old 2020-08-31, 21:21   #7
ewmayer
2ω=0
 
ewmayer's Avatar
 
Sep 2002
Rep├║blica de California

3×7×19×29 Posts
Default

How different is the Ubuntu/gpuowl setup procedure expected to be for RX580 from that for Radeon VII? If the set of diffs between the 2 is not huge, perhaps I could change the title of the aforementioned thread to reflect multiple AMD GPUs, and add the RX580-specific info as a post-on-page-1 of the thread.
ewmayer is offline   Reply With Quote
Old 2020-08-31, 22:59   #8
kriesel
 
kriesel's Avatar
 
"TF79LL86GIMPS96gpu17"
Mar 2017
US midwest

128F16 Posts
Default

Appropriate config.txt -maxAlloc entry will depend on whether it's a 4GB or 8GB RX580, affecting P-1 runs and whether 2 instances of P-1 on the same gpu at the same time is wise, or limiting exponent & stage 2 bounds combinations that are feasible or reasonable.
I've run stage 2 to over 500M on an 8GB RX480, one instance, so not terribly limiting there.
Radeon VII, or 16GB Teslas, can do up to 1G exponent and probably higher.
Something else that can limit it is installed system ram; runs that are fine with 16GB or 10GB had memory allocation problems seemingly at stage 1 gcd time with 4GB of system ram.
(That's on Windows, except the Teslas were Google Colab so linux there.)
Memory requirements are highest for P-1 stage 2, lowest for trial factoring.

Last fiddled with by kriesel on 2020-08-31 at 23:09
kriesel is offline   Reply With Quote
Reply

Thread Tools


Similar Threads
Thread Thread Starter Forum Replies Last Post
RAID for beginners xilman Lounge 2 2009-08-17 17:32
Larrabee instruction set announced fivemack Hardware 0 2009-03-25 12:09
LMH for Beginners on V5 Bundu Lone Mersenne Hunters 3 2008-12-30 17:41
instruction for counting bits? bsquared Programming 7 2007-10-18 14:31
ECM Factoring for beginners BWetter246 GMP-ECM 5 2006-11-15 13:19

All times are UTC. The time now is 01:19.

Thu Dec 3 01:19:54 UTC 2020 up 83 days, 22:30, 1 user, load averages: 1.72, 1.82, 2.01

Powered by vBulletin® Version 3.8.11
Copyright ©2000 - 2020, Jelsoft Enterprises Ltd.

This forum has received and complied with 0 (zero) government requests for information.

Permission is granted to copy, distribute and/or modify this document under the terms of the GNU Free Documentation License, Version 1.2 or any later version published by the Free Software Foundation.
A copy of the license is included in the FAQ.