Please do, it's likely to be useful to many.
Thanks,

---

FAQ's

HOW TO:
- Opt out of Beta Tests
- Ask for Help

When I first installed 16.04 I was using a 750Ti and I installed the 361 driver from the repository. When I tried to install my 1060 I couldn't get it to boot. I put the 750 back in and started searching. The 367.27 driver is needed for the 10xx cards. I also found it helped me to install synaptic package manager first.
I used the following commands to get started with the video drivers.

sudo add-apt-repository ppa:graphics-drivers/ppa
sudo apt-get update
sudo apt-get install nvidia-367 nvidia-settings

I then reinstalled the 1060 and it booted but there were still some issues. Installing the 367 driver did not install the needed OpenCl or CUDA libraries. I used synaptic package manager to install them. Now we're almost done.

Add the unprivileged boinc account to the video group

sudo usermod -a -G video boinc

Move BOINC start-up to the end of the start-up processing.

sudo update-rc.d -f boinc-client remove
sudo update-rc.d boinc-client defaults 99

Without nvidia-modprobe installed boinc will not find the GPU

sudo apt-get install nvidia-modprobe <-- when I did this I found that it was already installed YMMV

sudo modprobe nvidia_uvm

sudo service boinc-client restart

That's it but in some cases a reboot is needed. Hope this will help someone. My 1060 now happily runs POEM and Einstein. Patiently waiting for support here
You can find more driver info here.

Tried to use the 375.20 repo drivers on Ubuntu 16.04_x64 LTS but could not get Boinc to recognise the GPU. Tried lots of tricks (repeatedly) but none worked. Reverted to 370.28 and things are working again. X server said the 375.20 driver was there but maybe it didn't fully install.
For here there is no known reason to upgrade - I was just looking to see if there was anything in it (library updates that might make things faster), on the off-chance. There usually isn't BTW. If anyone tries to use 375.20 and can't get it to work, don't spend too much time on it.

---

FAQ's

HOW TO:
- Opt out of Beta Tests
- Ask for Help

I have a setup running


Boinc 7.6.31
GenuineIntel Intel(R) Core(TM) i7-4770K CPU @ 3.50GHz [Family 6 Model 60 Stepping 3]
(8 processors) [2] NVIDIA GeForce GTX 780 (3052MB) driver: 375.20 Linux
4.4.0-21-generic 25 Nov 2016 | 16:44:32 UTC


I was doing an upgrade from Linux Mint 17.3 (ubuntu 14.04) to 18 (ubuntu 16.04) ,
I did test the 375.20 drivers with both versions. however to get the latest boinc client I had to get to mint 18.

Basic approach - your milage may vary :)

sudo apt-get purge nvidia*

Note - Once you reboot here you have to do the next steps from the Console as it will not boot into X
sudo reboot

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

sudo apt-get install nvidia-375 nvidia-settings libcuda1-375

sudo apt-get install boinc 7.6


I found these sites usefull -
http://askubuntu.com/questions/757080/why-is-it-so-hard-to-install-nvidia-drivers-what-is-the-most-fool-proof-method
https://sites.google.com/site/easylinuxtipsproject/12

Here is what appears to be the relevent packages - there may be more - these are the ones I touched during the upgrades

paz@stealth-mint ~ $ sudo dpkg -l | grep -i nvidia
[sudo] password for paz:
rc libcuda1-331 331.113-0ubuntu0.0.4 amd64 NVIDIA CUDA runtime library
rc libcuda1-346 346.22-0ubuntu1~xedgers14.04.1 amd64 NVIDIA CUDA runtime library
rc libcuda1-349 349.16-0ubuntu0~xedgers14.04.1 amd64 NVIDIA CUDA runtime library
ii libcuda1-375 375.20-0ubuntu0~gpu16.04.1 amd64 NVIDIA CUDA runtime library
ii nvidia-375 375.20-0ubuntu0~gpu16.04.1 amd64 NVIDIA binary driver - version 375.20
ii nvidia-settings 375.20-0ubuntu0~gpu16.04.1 amd64 Tool for configuring the NVIDIA graphics driver

paz@stealth-mint ~ $ uname -a
Linux stealth-mint 4.4.0-21-generic #37-Ubuntu SMP Mon Apr 18 18:33:37 UTC 2016 x86_64 x86_64 x86_64 GNU/Linux

paz@stealth-mint ~ $ sudo dpkg -l | grep -i boinc
ii boinc 7.6.31+dfsg-6ubuntu1 all metapackage for the BOINC client and the manager
ii boinc-client 7.6.31+dfsg-6ubuntu1 amd64 core client for the BOINC distributed computing infrastructure
ii boinc-manager 7.6.31+dfsg-6ubuntu1 amd64 GUI to control and monitor the BOINC core client
ii libboinc7:amd64 7.6.31+dfsg-6ubuntu1 amd64 libraries of BOINC the client depends on

paz@stealth-mint ~ $

I've been converting all my Windows machines over to Linux of late. I have been running Win7 but given I don't want Microsoft deciding how to manage my machines for me I went with Linux.

I've written a blog post with the instructions HERE describing how to get Debian going.

Why Debian? Its more up to date than Ubuntu - Ubuntu is based upon Debian. Also everything needed is in their standard repositories. No fiddling with driver versions and the like.

Regarding the problems getting the manager to connect with the core client I found that simply adding the password that's in gui_rpc_auth.cfg to the properties of the BOINC Manager icon was enough for it to work. No need to have two copies of gui_rpc_auth. I use a windows laptop that runs BOINC Tasks to control the crunchers but also have the manager on the desktop of each cruncher.

---

BOINC blog

UBUNTU 18.04 LTS
BOINC 7.9.3
NVIDIA driver 390.77
GPU 980 TI

How configure driver and BOINC, to reach 100% GPU load ?

GPU load about 30% 60%
T° 79° speed fan 50%

---

I am not familiar with Linux, but to me it seems that there may be some GPU throtteling due to the high temperature (79° - which, running constantly, is definitely not good for the lifetime of the card).

UBUNTU 18.04 LTS
BOINC 7.9.3
NVIDIA driver 390.77
GPU 980 TI

How configure driver and BOINC, to reach 100% GPU load ?

GPU load about 30% 60%
T° 79° speed fan 50%

I agree with the previous statement. You need to look at increasing your fan speed as your card is getting hot. Do you have nvidia settings installed? Are you able to adjust the fan speed?

Edit.. OK, just saw your PM. I have responded with instructions for you

Z

---

UBUNTU 18.04 LTS
BOINC 7.9.3
NVIDIA driver 390.77
GPU 980 TI

How configure driver and BOINC, to reach 100% GPU load ?

You can't use the SWAN_SYNC environmental variable under Linux to dedicate a CPU thread for the GPUGrid app to increase GPU usage.
(It works only under Windows.)
BTW the GPUGrid app can't reach 100% GPU usage. The maximum is about 95% (under Windows XP with SWAN_SYNC on).

GPU load about 30% 60%

This is quite low.
There's one component you forgot to post:
Your host has an AMD Ryzen Threadripper 2990WX 32-Core Processor, which is hyper-threaded, so it can run 64 tasks at the same time. In my experience if BOINC is not limited to use only the 50% of a HT CPU, the tasks will hinder each other's performance (as these scientific applications usually do a lot of floating point operations, and there's only 1 such FP unit available per core on the CPU). Memory bandwidth could be a limiting factor too when running too many tasks simultaneously. Too many CPU tasks can lower GPU performance a lot (In my experience there can be only 1 CPU task running without much GPU performance loss).

T° 79° speed fan 50%

This is quite hot, perhaps your CPU & GPU heats each other too much. If the GPU cooler emits the heat inside the PC case, you'll need some extra fans to remove the excess heat.

UBUNTU 18.04 LTS
BOINC 7.9.3
NVIDIA driver 390.77
GPU 980 TI

How configure driver and BOINC, to reach 100% GPU load ?

GPU load about 30% 60%
T° 79° speed fan 50%

I am not quite sure how you managed to get that, except insufficient CPU support as Zoltan suggested. But I have a GTX 980 running under Ubuntu 16.04 at the moment (PABLO_2IDP), supported by a single core of an i7-3770. (The other seven cores are running Universe.) And I have the 396.54 driver, though I doubt that matters. It runs at 91 to 95% utilization, averaging about 93%. So I think you need to do something with the CPU to support your card better.

As for the temps, they are high. Is your room warm? But I see your fan is running at only 50%, so maybe it is designed to run hot. Some cards are. (Mine runs cool, since it has three fans; currently 65C with 38% fan speed.) Good luck.

PS - I have a Ryzen 1700 machine, and it has no trouble supporting Nvidia cards. I think it does at least as well as my Haswell (i7-4770), so the Threadripper should do OK. But what other projects are you running? How many cores reserved?

PPS: I recall it being discussed a while ago that Nvidia cards down-clock above a certain temperature, but I don't know what it is. You might want to check your GPU clock speed.

I use a windows laptop that runs BOINC Tasks to control the crunchers but also have the manager on the desktop of each cruncher.

I just run BoincTasks under Wine. I have to have BT to monitor and control my hosts but don't want to ever touch Windows again. BT runs fine under Wine.

My GTX 750 Ti on a SuSE Leap 15.0 Linux and an Opteron 1210 CPU runs at 57 C, GPU clock 1137 MHz, memory clock 2700 MHz. Driver is 390.87.
Tullio

Yes, all the Nvidia cards down/up clock in 12 Mhz bins for every 5 degrees decrease/increase in temps. I think the temp where the card doesn't hit the thermal binning point is somewhere around 54°C. At that point it will run at max clocks. All my EVGA Hybrid cards always run at max clocks since they are always cooler than 54°C. The only card that even gets close or rarely goes past that point is the GTX 1080Ti in the summer. So the cooler you can run the card, the higher the clocks it will maintain.

I run multiple projects on the cards but only 1 gpu task at a time. So any card at any time might have a Seti, MilkyWay, Einstein or GPUGrid task on it. Yet any time I look at nvidia-smi I always see all cards always at 95 - 100% utilization.

With Ryzen or TR, try and keep cpu utilization below 75% to allow enough cpu support for the gpu tasks.

You can't use the SWAN_SYNC environmental variable under Linux to dedicate a CPU thread for the GPUGrid app to increase GPU usage.
(It works only under Windows.)

Maybe you can explain why I have one host that is able to run all GPUGrid tasks with SPIN synchronization. All my other hosts run with BLOCKING.

All the hosts have identical hardware for mobo, cpu and memory. The only differences are in the mix of Nvidia gpus each host has.

All the hosts run Ubuntu 18.04 with Nvidia 410.66 drivers.

So how is this host any different than the others? https://www.gpugrid.net/show_host_detail.php?hostid=456812

The only obvious difference is this host has three identical 1070Ti's and the other hosts have a mix of 1070/1070Ti/1080/1080TI cards in each host.

I've been asked a couple times now how I managed this feat. I don't have a clue.

You can't use the SWAN_SYNC environmental variable under Linux to dedicate a CPU thread for the GPUGrid app to increase GPU usage.
(It works only under Windows.)

Maybe you can explain why I have one host that is able to run all GPUGrid tasks with SPIN synchronization. All my other hosts run with BLOCKING.

I wish I could explain that, but I'm glad your host proves that I am wrong.

All the hosts have identical hardware for mobo, cpu and memory. The only differences are in the mix of Nvidia gpus each host has.

All the hosts run Ubuntu 18.04 with Nvidia 410.66 drivers.

So how is this host any different than the others? https://www.gpugrid.net/show_host_detail.php?hostid=456812

The only obvious difference is this host has three identical 1070Ti's and the other hosts have a mix of 1070/1070Ti/1080/1080TI cards in each host.

I've been asked a couple times now how I managed this feat. I don't have a clue.

There's total confusion about it on the forum.
I've tried every setting I can think of.
I've restarted my host after each configuration change.
I've tried to change the user account under which boinc runs, by changing the BOINC_USER setting in /etc/default/boinc-client and in /etc/init.d/boinc-client to my user and to root, none of them took effect (boinc still runs under user "boinc").
I've changed the access permissions of /var/lib/boinc-client to full access.
I've put "export swan_sync=0", "export swan_sync=1" to ~/.bashrc, /etc/init.d/boinc-client, /etc/bash.bashrc, /etc/profile
I've put "swan_sync=0", "swan_sync=1" to /etc/default/boinc-client, ~/.profile, /etc/profile
None of the above worked. (I don't know how Linux works, so it's no wonder.)
BTW I'm using Ubuntu 18.04.1 LTS.

References:
ACEMD2 6.12 cuda and 6.13 cuda31 for windows and linux
GTX 460
New applications ACEMD2 6.07/6.08 for Win and Lin
Load balancing GPUGRID and other boinc projects on Linux.
Setting up BOINC/GPUGRID on Fedora
GT240 and Linux: niceness and overclocking
Just a few Linux/GPU questions
The punchline: (message 38355)
CUDA 6.5 app for Linux now available on acemdbeta and acemdshort

Well mystery solved after reading your history of threads regarding swan_sync. I believe that host 456812 was the first host I brought up on Linux and the project.

Seems I somewhere in the past put the SWAN_SYNC=1 entry into the /etc/environment file.

I assume that is the reason that host has been running tasks with the SPIN synchronization parameter.

Even though everyone said their attempt with the same entry never worked, I wonder if the reason it works with my hosts is that I don't use the repository versions of BOINC.

I use a specially Seti@home user compiled version 7.4.44 with settings allowing 3000 tasks on board and also the developer compiles it on much older distributions for compatibility. I believe he makes it compatible with distribution as old as 12.04.

Anyway, I am now modifying the environment file for the other hosts and see whether they too start running in SPIN synch.

So I experimented and now have changed to SWAN_SYNC=1 on all my hosts. I ran overnight on the odd BOINC 7.8.3 host to make sure that it wasn't just the BOINC 7.4.44 that was able to use the SWAN_SYNC environment variable. I have the tasks now marked with SPIN synchronization too and the cpu_times equal the run_times so I used 100% of a cpu core to support the gpu task. Looks like I shaved 30-60 minutes of the typical PABLO task.

So I experimented and now have changed to SWAN_SYNC=1 on all my hosts. I ran overnight on the odd BOINC 7.8.3 host to make sure that it wasn't just the BOINC 7.4.44 that was able to use the SWAN_SYNC environment variable. I have the tasks now marked with SPIN synchronization too and the cpu_times equal the run_times so I used 100% of a cpu core to support the gpu task. Looks like I shaved 30-60 minutes of the typical PABLO task.

Do you think you could give a complete guide of how to accomplish this for the linux novice? I think everyone would benefit from this information.

It's fairly easy. Question is whether the version of BOINC manager you are using will allow you to use it.

Keith and I both run a BM that was compiled by one of the Seti Gurus.

First thing is to unhide your files. In the right upper corner you should have a button to see files, you want to click show hidden files.

Once you do that, open files. Find your computer, click on it. Find the etc folder, click on that and then locate the environment file. Check it with texteditor. Might say something with "games" in the name. All you are doing is looking at it. Then close the file

move mouse over the open etc folder and right click and open the terminal.

Next get to root, so

"sudo su"

hit return then it will ask you to enter your password

you should be at root

"sudo nano environment"

Now you can edit the environment file

Add the

"SWAN_SYNC=1"

then control o to write, hit enter. You have now overwritten that file

control x to exit the edit

close terminal.

now using textedit, recheck to see if the edit took. If it did then you should be good.

I don't think you need to reboot. That should be it. I've tested it on one of my system and it works.

Edit..

Ok, reboot then. I always reboot anytime I do anything to a computer. Thanks Keith.

---

So I experimented and now have changed to SWAN_SYNC=1 on all my hosts. I ran overnight on the odd BOINC 7.8.3 host to make sure that it wasn't just the BOINC 7.4.44 that was able to use the SWAN_SYNC environment variable. I have the tasks now marked with SPIN synchronization too and the cpu_times equal the run_times so I used 100% of a cpu core to support the gpu task. Looks like I shaved 30-60 minutes of the typical PABLO task.

Did the tasks change w/in the past day or have you not been getting the 24hr bonus? All your tasks are worth 73k, 92k or 120k. I had mainly been getting 110k and 180k credit tasks. I was hoping to compare vs my 1070 running 2x tasks at once but want to make sure I compare proper task length.