mfaktc: a CUDA program for Mersenne prefactoring

[James Heinrich]

Originally Posted by storm5510

My furnace has two 20 ampere heating elements. It is 240 VAC

For that power budget you could run 48x A100's and get 322 THz-days/d of TF done. And heat your house for free
(although at that point you probably don't have a house anymore because you sold it to buy the GPUs...)

[kriesel]

Originally Posted by James Heinrich

For that power budget you could run 48x A100's and get 322 THz-days/d of TF done. And heat your house for free
(although at that point you probably don't have a house anymore because you sold it to buy the GPUs...)

So, a well heated tent?

[pinhodecarlos]

Nothing is for free.

[chalsall]

Originally Posted by pinhodecarlos

Nothing is for free.

Second law of thermodynamics (paraphrased).

[xx005fs]

Originally Posted by TheJudger

Hi,

seems like mfaktc runs fine with CUDA 11 on Ampere (no specific changes for Ampere except Makefile).

mfaktc v0.22-pre8 (64bit built)
[...]
CUDA version info
  binary compiled for CUDA  11.0
  CUDA runtime version      11.0
  CUDA driver version       11.0

CUDA device info
  name                      A100-SXM4-40GB
  compute capability        8.0
  max threads per block     1024
  max shared memory per MP  167936 byte
  number of multiprocessors 108
  clock rate (CUDA cores)   1410MHz
  memory clock rate:        1215MHz
  memory bus width:         5120 bit
[...]
Starting trial factoring M66362159 from 2^74 to 2^75 (57.65 GHz-days)
 k_min =  142321062303420
 k_max =  284642124610180
Using GPU kernel "barrett76_mul32_gs"
Date    Time | class   Pct |   time     ETA | GHz-d/day    Sieve     Wait
Jul 19 21:19 |    0   0.1% |  0.829  13m15s |   6259.18    82485    n.a.%
Jul 19 21:19 |    4   0.2% |  0.779  12m26s |   6660.92    82485    n.a.%
Jul 19 21:19 |    9   0.3% |  0.780  12m26s |   6652.38    82485    n.a.%
[...]
Jul 19 21:31 | 4617 100.0% |  0.780   0m00s |   6652.38    82485    n.a.%
no factor for M66362159 from 2^74 to 2^75 [mfaktc 0.22-pre8 barrett76_mul32_gs CUDA 11.0 arch 8.0] 51D74917
tf(): total time spent: 12m 32.323s

New absolute performance champion and I guess best performance per watt, too!


Older benchmark data for Turing (RTX 2080 Ti): https://mersenneforum.org/showpost.p...postcount=2912

Oliver

That is blisteringly fast! Wonder if there will be GPUOWL benchmark coming up for this beast. The 1.5+TB/s memory bandwidth and its near 10TFLOPs of FP64 capabilities will be very interesting.

[DrobinsonPE]

I put together a computer with some old and new parts; i3-4130T, Linux Mint 20.1, and a GTX 1650 Super. I downloaded the compiled mfaktc for linux linked at post #3208 of this thread and discovered that I did not have CUDA on the computer so after searching the internet for a while I ran this.

sudo apt install nvidia-cuda-toolkit

Now mfaktc.exe starts but does not find the mfaktc.ini or worktodo.txt files so it errors out with (can't open "worktodo.txt"). Both mfaktc.ini and worktodo.txt are in the same folder as mfaktc.

I ran mfaktc.exe -st and after issuing a screen full of

"WARNING: Cannot read................. from mfaktc.ini, using default value........"

it proceeded to perform the selftest and passed all 26192 tests.

It seems I have stumbled my way to within inches of success but cannot figure out why mfaktc.exe cannot find mfaktc.ini or worktodo.txt. Is anyone willing to give helpful suggestions?

[kriesel]

Originally Posted by DrobinsonPE

I put together a computer with some old and new parts; i3-4130T, Linux Mint 20.1, and a GTX 1650 Super. I downloaded the compiled mfaktc for linux linked at post #3208 of this thread and discovered that I did not have CUDA on the computer so after searching the internet for a while I ran this.

sudo apt install nvidia-cuda-toolkit

Now mfaktc.exe starts but does not find the mfaktc.ini or worktodo.txt files so it errors out with (can't open "worktodo.txt"). Both mfaktc.ini and worktodo.txt are in the same folder as mfaktc.

I ran mfaktc.exe -st and after issuing a screen full of

"WARNING: Cannot read................. from mfaktc.ini, using default value........"

it proceeded to perform the selftest and passed all 26192 tests.

It seems I have stumbled my way to within inches of success but cannot figure out why mfaktc.exe cannot find mfaktc.ini or worktodo.txt. Is anyone willing to give helpful suggestions?

Check ownership and permissions on .ini and work files and on the folder. The process should be running with its current directory located in the folder where they reside. Must have read permission on ini, and the ability to modify worktodo file and the folder that contains it. If the process can't log to a file there also, by redirect or tee, there's an issue with the folder permissions probably. I'm a Windows guy mostly, and have never used Mint, so there are limits to how much I can help. But some of this is independent of OS.
Have you tried running mprime on it yet? Any issues there?
Good luck and have fun!

[DrobinsonPE]

Originally Posted by kriesel

Check ownership and permissions on .ini and work files and on the folder. The process should be running with its current directory located in the folder where they reside. Must have read permission on ini, and the ability to modify worktodo file and the folder that contains it. If the process can't log to a file there also, by redirect or tee, there's an issue with the folder permissions probably. I'm a Windows guy mostly, and have never used Mint, so there are limits to how much I can help. But some of this is independent of OS.
Have you tried running mprime on it yet? Any issues there?
Good luck and have fun!

Thank you for the suggestions. I have not run into permissions problems yet so had not thought of that. I checked and as a user, I have ownership and read/write permissions for the folder and files. I tested running as root in addition to running as user just to see if something needed root permissions.

When I downloaded mfaktc, I un-tared everything into the one folder/directory and created the worktodo in the same folder. I restarted the computer just now because that is a very windows thing to do and I figured it couldn't hurt anything. Mfakto is still not finding the ini or worktodo.

I already downloaded mprime and ran multiple benchmarks to test the CPU before I installed the GPU. I ran the benchmark again just now and then put one of my assignments in a worktodo file and started it up. Mprime seems to work fine.

I will keep noodling my mfaktc issue for a while and try a few more things but eventually if I cannot get it to work on Linux, I will load Windows and download one of the Windows versions. That is how I got mfakto working on my Ryzen APU. I tried for about 3 months to get it to compile on Linux before switching to Windows and finding success.

I have dabbled in DOS, Windows, Mac, and Linux but am far from an expert in any of them.

[kriesel]

Originally Posted by DrobinsonPE

Thank you for the suggestions. I have not run into permissions problems yet so had not thought of that. I checked and as a user, I have ownership and read/write permissions for the folder and files. I tested running as root in addition to running as user just to see if something needed root permissions.

When I downloaded mfaktc, I un-tared everything into the one folder/directory and created the worktodo in the same folder. I restarted the computer just now because that is a very windows thing to do and I figured it couldn't hurt anything. Mfakto is still not finding the ini or worktodo.

Linux unlike Windows is case-sensitive. Just a thought. Worktodo.txt != worktodo.txt etc. Occasionally I'll mistype and get filename..ext or filename.ext.
Maybe post ls-l output for your folder and any shell script or command line you use to launch mfakto.
In the meantime you could get some work done by specifying factoring work on the command line. See the -tf option in https://www.mersenneforum.org/showpo...90&postcount=7

[DrobinsonPE]

Originally Posted by kriesel

Linux unlike Windows is case-sensitive. Just a thought. Worktodo.txt != worktodo.txt etc. Occasionally I'll mistype and get filename..ext or filename.ext.
Maybe post ls-l output for your folder and any shell script or command line you use to launch mfakto.
In the meantime you could get some work done by specifying factoring work on the command line. See the -tf option in https://www.mersenneforum.org/showpo...90&postcount=7

Thank you again for helping. After my last post, I did the following:

the original download un-tared to user/mfaktc/mfaktc-0.21, I renamed the folder mfaktc-21 to make it easier to type (the renaming was done after the previous post). I then made a duplicate of the complete folder to user/mfaktc/mfaktc021, deleted the mfaktc.exe file, went into the src sub-folder and typed make. It did not work so I investigated and discovered I needed:

sudo apt-get install build-essential

after installing build-essential, I then typed make and it created a new mfaktc.exe file in the user/mfaktc/mfaktc021 folder. Then running mfaktc/mfaktc021/mfaktc.exe -st, I got the exact same results as before. A screen full of:

"WARNING: Cannot read................. from mfaktc.ini, using default value........"

and it proceeded to perform the selftest and passed all 26192 tests. I stopped there out of frustration.

Good comment on the case-sensitive and mistype. I did check that because when I was trying to get GPUOWL working on my Ryzen APU, I spent about 4 days looking at the screen wondering what was wrong before realizing I had typed worktodo.txt.txt (Windows issue...you do not need to add the extension when saving in Notepad).

here is the output of ls -l from the folder mfaktc/mfaktc021

-rw-rw-rw- 1 user user 18845 Feb 17 2015 Changelog.txt
-rw-rw-rw- 1 user user 35147 Feb 17 2015 COPYING
-rwxrwxr-x 1 user user 2719952 Aug 22 20:40 mfaktc.exe
drwxrwxr-x 2 user user 4096 Sep 4 2019 src
-rw-rw-r- - 1 user user 55 Aug 22 20:55 worktodo.txt

I am on a different computer so I typed the output in manually but double checked everything letter for letter against the output (only changing my user name to user).

I will try using the -tf option because it sounds interesting and to see if I can get it produce some work. However, long term I think I am also going to get a fresh download of mfaktc from the repository, put it in a completely different folder and try to compile from that.

[DrobinsonPE]

This is strange but I just figured out what was going on by accident. I ran mfaktc/mfaktc021/mfaktc.exe -h >>mfaktchelp.txt. When I went to mfaktc/mfaktc021, the text file was not there. I then ran a search for mfaktchelp.txt and it was in my user folder.

I copied mfaktc.ini and worktodo.txt to my user folder and then ran mfaktc/mfaktc021/mfaktc.exe. It works!

I am now going to get a real assignments from GPU72, load it up and see what happens. baby steps. I also need to look back on how to use nvidia-smi to monitor the GPU.

nvidia-smi output 93-100W, 77-78C, mfaktc averaging 1400 GHz-d/day, 35% done 22 minutes left for a result.