Setting up Linux to crunch CUDA90 and above for Windows users

Message boards : Number crunching : Setting up Linux to crunch CUDA90 and above for Windows users
Message board moderation

To post messages, you must log in.

Previous · 1 . . . 114 · 115 · 116 · 117 · 118 · 119 · 120 . . . 162 · Next

AuthorMessage
Stephen "Heretic" Crowdfunding Project Donor*Special Project $75 donorSpecial Project $250 donor
Volunteer tester
Avatar

Send message
Joined: 20 Sep 12
Posts: 5557
Credit: 192,787,363
RAC: 628
Australia
Message 2001718 - Posted: 8 Jul 2019, 21:44:01 UTC - in response to Message 2001639.  
Last modified: 8 Jul 2019, 21:48:59 UTC

On My machines I see around a 10% speedup on the Arecibo tasks, plus you can now run a GTX 960 2 GB without it running out of vRAM at unroll 8. Those items alone should be enough for everyone to update. If you have a Turing GPU the BLC tasks also see a nice speed improvement whereas it's not quite so much on the older GPUs. My 750 Ti sees a BLC speedup, the 970 doesn't see a BLC speedup, EVERYTHING sees an Arecibo Speedup.

. . Thanks for that TBar, I take it that comparison is between 0.98b1-90 and 0.98b1-101. I do remember the mention of the lower rate of inconclusives but with very little Arecibo work and an otherwise low rate already it was not compelling for me. I guess it is doable then on both my machines with Ubuntu 14.04. Again thanks.

. . One thing that I find amusing is that there was an improvement on GBT WUs on the 750ti but not the 970s, those little cards are amazing. Such low power usage but very impressive performance.

Stephen

:)
ID: 2001718 · Report as offensive     Reply Quote
elec999 Project Donor

Send message
Joined: 24 Nov 02
Posts: 375
Credit: 416,969,548
RAC: 141
Canada
Message 2001720 - Posted: 8 Jul 2019, 21:51:40 UTC - in response to Message 2001718.  

Is this P2000 card worth anything? Anyone using it?
ID: 2001720 · Report as offensive     Reply Quote
Stephen "Heretic" Crowdfunding Project Donor*Special Project $75 donorSpecial Project $250 donor
Volunteer tester
Avatar

Send message
Joined: 20 Sep 12
Posts: 5557
Credit: 192,787,363
RAC: 628
Australia
Message 2001721 - Posted: 8 Jul 2019, 21:52:17 UTC - in response to Message 2001641.  

it's absolutely worth updating. just because there is still a CUDA 9.0 version of the app, doesn't mean it's the same app. v0.98 is quite a bit faster than v0.97.
if you have a Turing card, then the CUDA 10+ apps are very slightly faster than 9.0. but v0.98 is faster than 0.97 all else being equal.


. . Thanks, that is the sort of answer I was hoping for ...

Stephen

:)
ID: 2001721 · Report as offensive     Reply Quote
Stephen "Heretic" Crowdfunding Project Donor*Special Project $75 donorSpecial Project $250 donor
Volunteer tester
Avatar

Send message
Joined: 20 Sep 12
Posts: 5557
Credit: 192,787,363
RAC: 628
Australia
Message 2001723 - Posted: 8 Jul 2019, 22:09:55 UTC - in response to Message 2001720.  

Is this P2000 card worth anything? Anyone using it?


. . If you read Keith's advice and make sure you reboot the machine after updating the Video drivers to release .418 then that card should work quite well for you. Whenever you see the word "nouveau" in the driver info section you have the WRONG video drivers.

. . Good luck ...

Stephen

:)
ID: 2001723 · Report as offensive     Reply Quote
Profile Keith Myers Special Project $250 donor
Volunteer tester
Avatar

Send message
Joined: 29 Apr 01
Posts: 13161
Credit: 1,160,866,277
RAC: 1,873
United States
Message 2001726 - Posted: 8 Jul 2019, 22:24:47 UTC - in response to Message 2001720.  

Is this P2000 card worth anything? Anyone using it?

Probably equivalent to a GTX 750 Ti or a GTX 1050 Ti with the special app.
Seti@Home classic workunits:20,676 CPU time:74,226 hours

A proud member of the OFA (Old Farts Association)
ID: 2001726 · Report as offensive     Reply Quote
Stephen "Heretic" Crowdfunding Project Donor*Special Project $75 donorSpecial Project $250 donor
Volunteer tester
Avatar

Send message
Joined: 20 Sep 12
Posts: 5557
Credit: 192,787,363
RAC: 628
Australia
Message 2001733 - Posted: 8 Jul 2019, 22:47:33 UTC - in response to Message 2001726.  
Last modified: 8 Jul 2019, 23:05:44 UTC

Is this P2000 card worth anything? Anyone using it?

Probably equivalent to a GTX 750 Ti or a GTX 1050 Ti with the special app.


. . With a small dose of steroids ... :) It is a pascal card but it has 5GB GDDR5 ram, more than the 1050ti for some reason.

. . On looking further it has 1024 Cuda cores and 8 CUs. So I am thinking a restricted 1060 maybe?

Stephen

:)
ID: 2001733 · Report as offensive     Reply Quote
elec999 Project Donor

Send message
Joined: 24 Nov 02
Posts: 375
Credit: 416,969,548
RAC: 141
Canada
Message 2001735 - Posted: 8 Jul 2019, 22:51:06 UTC - in response to Message 2001726.  

Is this P2000 card worth anything? Anyone using it?

Probably equivalent to a GTX 750 Ti or a GTX 1050 Ti with the special app.

should I do the -nobs switch and use the cuda 101?
ID: 2001735 · Report as offensive     Reply Quote
Profile Keith Myers Special Project $250 donor
Volunteer tester
Avatar

Send message
Joined: 29 Apr 01
Posts: 13161
Credit: 1,160,866,277
RAC: 1,873
United States
Message 2001737 - Posted: 8 Jul 2019, 22:53:35 UTC - in response to Message 2001735.  
Last modified: 8 Jul 2019, 23:22:30 UTC

I think the consensus we've arrived at today in this and other threads is that a Pascal card should probably stick to the CUDA90 application. Which means a minimum of 384 series drivers but not to exceed the 410 series drivers for best performance. The -nobs will help but no matter which driver as long as you have enough cpu cores to support the gpu task.
Seti@Home classic workunits:20,676 CPU time:74,226 hours

A proud member of the OFA (Old Farts Association)
ID: 2001737 · Report as offensive     Reply Quote
Profile ThePHX264

Send message
Joined: 29 May 19
Posts: 86
Credit: 6,688,090
RAC: 32
United States
Message 2001742 - Posted: 8 Jul 2019, 23:10:04 UTC - in response to Message 2001737.  

I think the consensus we've arrived at today in this and other threads is that a Pascal card should probably stick to the CUDA90 application. Which means a minimum of 384 series drivers but not to exceed the 410 series drivers for best performance. The -nobs will help put no matter which driver as long as you have enough cpu cores to support the gpu task.



I think that's what the new computer is missing, some nobs. I'm going to have to add that to the file
ID: 2001742 · Report as offensive     Reply Quote
Stephen "Heretic" Crowdfunding Project Donor*Special Project $75 donorSpecial Project $250 donor
Volunteer tester
Avatar

Send message
Joined: 20 Sep 12
Posts: 5557
Credit: 192,787,363
RAC: 628
Australia
Message 2001743 - Posted: 8 Jul 2019, 23:10:07 UTC - in response to Message 2001737.  

I think the consensus we've arrived at today in this and other threads is that a Pascal card should probably stick to the CUDA90 application. Which means a minimum of 384 series drivers but not to exceed the 410 series drivers for best performance. The -nobs will help put no matter which driver as long as you have enough cpu cores to support the gpu task.


. . He has a i7-8700 so he has 12 threads to play with. As long as he sets the Use CPUs to no more than about 85% he should be fine with -nobs.

Stephen

:)
ID: 2001743 · Report as offensive     Reply Quote
elec999 Project Donor

Send message
Joined: 24 Nov 02
Posts: 375
Credit: 416,969,548
RAC: 141
Canada
Message 2001746 - Posted: 9 Jul 2019, 0:06:56 UTC - in response to Message 2001742.  

I think the consensus we've arrived at today in this and other threads is that a Pascal card should probably stick to the CUDA90 application. Which means a minimum of 384 series drivers but not to exceed the 410 series drivers for best performance. The -nobs will help put no matter which driver as long as you have enough cpu cores to support the gpu task.



I think that's what the new computer is missing, some nobs. I'm going to have to add that to the file


i TRIED to remove the -418 drivers and install the 410, but it still gave me the 418.
ID: 2001746 · Report as offensive     Reply Quote
Profile Keith Myers Special Project $250 donor
Volunteer tester
Avatar

Send message
Joined: 29 Apr 01
Posts: 13161
Credit: 1,160,866,277
RAC: 1,873
United States
Message 2001748 - Posted: 9 Jul 2019, 0:17:50 UTC - in response to Message 2001746.  

I think the consensus we've arrived at today in this and other threads is that a Pascal card should probably stick to the CUDA90 application. Which means a minimum of 384 series drivers but not to exceed the 410 series drivers for best performance. The -nobs will help put no matter which driver as long as you have enough cpu cores to support the gpu task.



I think that's what the new computer is missing, some nobs. I'm going to have to add that to the file


i TRIED to remove the -418 drivers and install the 410, but it still gave me the 418.

HOW did you try to shift to the 410 drivers? Are you using the ppa? I think the latest Ubuntu 19.04 distro repo only has the 418 drivers. You should install the ppa so you have access to all the Nvidia drivers from 390 to 430.

To install the ppa drivers from the command line

sudo add-apt-repository ppa:graphics-drivers/ppa

sudo apt install nvidia-driver-410 && reboot

Seti@Home classic workunits:20,676 CPU time:74,226 hours

A proud member of the OFA (Old Farts Association)
ID: 2001748 · Report as offensive     Reply Quote
TBar
Volunteer tester

Send message
Joined: 22 May 99
Posts: 5204
Credit: 840,779,836
RAC: 2,768
United States
Message 2001769 - Posted: 9 Jul 2019, 2:48:21 UTC - in response to Message 2001678.  
Last modified: 9 Jul 2019, 3:20:07 UTC

i don't know what you mean by "CUDA10 archive"
but the app you want to use with the 410 (or lower) drivers will be the CUDA90 app. This is CUDA 9.0 enabled. not CUDA 10. I have a beta version that is CUDA 10 (not 10.1) enabled.
the apps available to you in the latest standard All-In-One package are the CUDA 9.0 and CUDA 10.1 app. choose the 9.0 one. You wont see any benefit using a CUDA 10+ app on your GTX 10-series cards
I downloaded a new copy. And it appears I was running an unknown previous version of the 10.1 archive.
It was throwing computation errors.
Installed the "new" archive over the top of the "old" stuff. Put in the "-nobs" in the app_info.xml file.
Running it now with NNT while I confirm it has stopped throwing "computation errors".
Tom
I really don't see why you are still running V0.97b2, Cuda 9.10. On My machines 10.1/10.0 runs about the same and are a few seconds faster than any version of 9.x. As for v0.98, look at the differences between my 1070s and your 1070 on Arecibo tasks. The first is on the mining rig running driver 318, the Second is on the Hack running the latest 10.1 Update 1 (newer than 10.1);
My 1070
Name 29jn19ab.22239.10905.9.36.205.vlar_1
Run time: 2 min 19 sec
CPU time: 1 min 8 sec

My Hack
Name 29jn19ab.23659.18369.10.37.15.vlar_0
Run time: 2 min 11 sec
CPU time: 2 min 9 sec

Your 1070
Name 29jn19ab.27599.11519.14.41.68.vlar_0
Run time: 2 min 39 sec
CPU time: 2 min 36 sec


My 1070s use to run the Arecibo VLARs about the same with v0.97, they are Much better with 0.98. Even the 1060s are much better with 0.98, they use to run over 4 mins with 0.97;
Name 29jn19ab.23659.11007.10.37.176.vlar_0
Run time: 3 min 28 sec
CPU time: 1 min 42 sec


And Yes, the Mining rig has nobs set. It doesn't matter there are twice as many GPUs as CPU Cores, the available CPU is divided between the GPUs and all that matters is the CPU remains assigned to the GPU for faster response than without nobs set. Obviously the GPU doesn't need a Full CPU, just having nobs set is enough for those few less seconds. It just means your CPU stays around 99%, so, have good CPU cooling. At some point over twice as many GPUs per CPU cores may make a difference, but My GPUs are happy at around 50% CPU with nobs set.
ID: 2001769 · Report as offensive     Reply Quote
elec999 Project Donor

Send message
Joined: 24 Nov 02
Posts: 375
Credit: 416,969,548
RAC: 141
Canada
Message 2001864 - Posted: 10 Jul 2019, 1:02:02 UTC - in response to Message 2001769.  

i don't know what you mean by "CUDA10 archive"
but the app you want to use with the 410 (or lower) drivers will be the CUDA90 app. This is CUDA 9.0 enabled. not CUDA 10. I have a beta version that is CUDA 10 (not 10.1) enabled.
the apps available to you in the latest standard All-In-One package are the CUDA 9.0 and CUDA 10.1 app. choose the 9.0 one. You wont see any benefit using a CUDA 10+ app on your GTX 10-series cards
I downloaded a new copy. And it appears I was running an unknown previous version of the 10.1 archive.
It was throwing computation errors.
Installed the "new" archive over the top of the "old" stuff. Put in the "-nobs" in the app_info.xml file.
Running it now with NNT while I confirm it has stopped throwing "computation errors".
Tom
I really don't see why you are still running V0.97b2, Cuda 9.10. On My machines 10.1/10.0 runs about the same and are a few seconds faster than any version of 9.x. As for v0.98, look at the differences between my 1070s and your 1070 on Arecibo tasks. The first is on the mining rig running driver 318, the Second is on the Hack running the latest 10.1 Update 1 (newer than 10.1);
My 1070
Name 29jn19ab.22239.10905.9.36.205.vlar_1
Run time: 2 min 19 sec
CPU time: 1 min 8 sec

My Hack
Name 29jn19ab.23659.18369.10.37.15.vlar_0
Run time: 2 min 11 sec
CPU time: 2 min 9 sec

Your 1070
Name 29jn19ab.27599.11519.14.41.68.vlar_0
Run time: 2 min 39 sec
CPU time: 2 min 36 sec


My 1070s use to run the Arecibo VLARs about the same with v0.97, they are Much better with 0.98. Even the 1060s are much better with 0.98, they use to run over 4 mins with 0.97;
Name 29jn19ab.23659.11007.10.37.176.vlar_0
Run time: 3 min 28 sec
CPU time: 1 min 42 sec


And Yes, the Mining rig has nobs set. It doesn't matter there are twice as many GPUs as CPU Cores, the available CPU is divided between the GPUs and all that matters is the CPU remains assigned to the GPU for faster response than without nobs set. Obviously the GPU doesn't need a Full CPU, just having nobs set is enough for those few less seconds. It just means your CPU stays around 99%, so, have good CPU cooling. At some point over twice as many GPUs per CPU cores may make a difference, but My GPUs are happy at around 50% CPU with nobs set.


CUDA install keeps on falling for me.
ID: 2001864 · Report as offensive     Reply Quote
Profile Keith Myers Special Project $250 donor
Volunteer tester
Avatar

Send message
Joined: 29 Apr 01
Posts: 13161
Credit: 1,160,866,277
RAC: 1,873
United States
Message 2001866 - Posted: 10 Jul 2019, 1:17:18 UTC - in response to Message 2001864.  

CUDA install keeps on falling for me.

We're going to need more information than this. What host? What is failing? What error messages?
Seti@Home classic workunits:20,676 CPU time:74,226 hours

A proud member of the OFA (Old Farts Association)
ID: 2001866 · Report as offensive     Reply Quote
Profile Bernie Vine
Volunteer moderator
Volunteer tester
Avatar

Send message
Joined: 26 May 99
Posts: 9954
Credit: 103,452,613
RAC: 328
United Kingdom
Message 2001907 - Posted: 10 Jul 2019, 10:03:07 UTC

Quick question

If I have a standard Linux install of Boinc and I wanted to change to the AIO, what is the easiest way?
ID: 2001907 · Report as offensive     Reply Quote
Profile Tom M
Volunteer tester

Send message
Joined: 28 Nov 02
Posts: 5124
Credit: 276,046,078
RAC: 462
Message 2001937 - Posted: 10 Jul 2019, 14:23:36 UTC - in response to Message 2001907.  

Quick question

If I have a standard Linux install of Boinc and I wanted to change to the AIO, what is the easiest way?


EASIEST?

In my opinion. Run your current Linux install dry (NNT).
Download Tbar's AIO.
Do a clean reinstall of Linux.
Then follow the directions for installing AIO.

Plan B is run your Linux dry (NNT).
Download/unarchive the AIO in downloads.
Locate the current places where Linux is installed. A non-trivial task.
Copy files from the unpacked archive to the equivalent places in your current install.
Confirm via properties that the exe files can be executed by "anyone".
Re-start Boinic.

Both of these procedures have been previously described here.
Running Seti dry is not officially required but I believe it is safer.

What is NOT recommended is trying to move the original Seti install to another location. There has been discussion trying that and there turns out to be permissions and other issues that I don't understand that made it less trouble to follow the reinstall the OS process.

HTH,
Tom
A proud member of the OFA (Old Farts Association).
ID: 2001937 · Report as offensive     Reply Quote
Ian&Steve C.
Avatar

Send message
Joined: 28 Sep 99
Posts: 4267
Credit: 1,282,604,591
RAC: 6,640
United States
Message 2001940 - Posted: 10 Jul 2019, 14:37:41 UTC - in response to Message 2001937.  

i've done it with no ill effects, one of my systems is still running like this. but you really do want to completely purge all remnants of the old install.

sudo apt purge *boinc*


and then let it uninstall everything and all related dependencies should probably get rid of the bulk of it. This worked on my system, and i think i had to manually clean up some remnants, but as always, YMMV, and I take no responsibility if Ubuntu decides to remove important stuff, as it sometimes likes to do. Always check the list of packages to be uninstalled before committing to it.

A full OS re-install will definitely remove everything and give you a clean slate. It could be faster depending on your linux proficiency level or if you don't care about losing other stuff related to your existing install.
Seti@Home classic workunits: 29,492 CPU time: 134,419 hours

ID: 2001940 · Report as offensive     Reply Quote
Profile Keith Myers Special Project $250 donor
Volunteer tester
Avatar

Send message
Joined: 29 Apr 01
Posts: 13161
Credit: 1,160,866,277
RAC: 1,873
United States
Message 2001967 - Posted: 10 Jul 2019, 16:55:49 UTC

I got rid of the standard BOINC installation on my Nano by digging out all the old references, permissions and symlinks so I could move the BOINC data folder to /home. A lot more work than just doing a purge though. At least I own BOINC now and can do with it what I want. Couldn't use the AIO obviously because of the platform differences.
Seti@Home classic workunits:20,676 CPU time:74,226 hours

A proud member of the OFA (Old Farts Association)
ID: 2001967 · Report as offensive     Reply Quote
Richard Haselgrove Project Donor
Volunteer tester

Send message
Joined: 4 Jul 99
Posts: 14649
Credit: 200,643,578
RAC: 874
United Kingdom
Message 2001975 - Posted: 10 Jul 2019, 18:33:32 UTC - in response to Message 2001907.  

Quick question

If I have a standard Linux install of Boinc and I wanted to change to the AIO, what is the easiest way?
When I did mine, I did it by:

1) installing repository BOINC
2) creating custom SaH project folder in home workspace - updating app_info from 9.0 to 10.1, removing unwanted apps etc.
3) copy new SaH folder to BOINC data folder/projects as SU
4) set execute bit on application: set write permissions on new folder
5) start BOINC Manager and attach to SETI. It finds the app_info and runs with specified app

It took much longer to get a GPU driver installed! When I get home, I still need to update v7.9.3 to v7.14.2 (or preferably v7.16.x - when David pulls his finger out)

But it's running as host 8747061 while I'm away.
ID: 2001975 · Report as offensive     Reply Quote
Previous · 1 . . . 114 · 115 · 116 · 117 · 118 · 119 · 120 . . . 162 · Next

Message boards : Number crunching : Setting up Linux to crunch CUDA90 and above for Windows users


 
©2024 University of California
 
SETI@home and Astropulse are funded by grants from the National Science Foundation, NASA, and donations from SETI@home volunteers. AstroPulse is funded in part by the NSF through grant AST-0307956.