Linux CUDA 'Special' App finally available, featuring Low CPU use

Message boards : Number crunching : Linux CUDA 'Special' App finally available, featuring Low CPU use
Message board moderation

To post messages, you must log in.

Previous · 1 · 2 · 3 · 4 · 5 · 6 · 7 . . . 71 · Next

AuthorMessage
rob smithProject Donor
Volunteer tester

Send message
Joined: 7 Mar 03
Posts: 15194
Credit: 251,445,793
RAC: 324,129
United Kingdom
Message 1836133 - Posted: 15 Dec 2016, 21:05:15 UTC

First few tasks running through just now.
(Following a "memo to self" moment = "Don't remove 'execute permissions' from executables"....)

I notice that there are a very large number of VLARs going through just now.

First few I've noticed:
GTX970 VLAR run-time ~17:25s
GTX1080 VLAR run-time ~12:50s
(running with whatever the base settings are)
host = https://setiathome.berkeley.edu/show_host_detail.php?hostid=8161267
Bob Smith
Member of Seti PIPPS (Pluto is a Planet Protest Society)
Somewhere in the (un)known Universe?
ID: 1836133 · Report as offensive     Reply Quote
TBar
Volunteer tester

Send message
Joined: 22 May 99
Posts: 3789
Credit: 186,212,712
RAC: 237,421
United States
Message 1836139 - Posted: 15 Dec 2016, 21:40:00 UTC - in response to Message 1836133.  

Seems to be working now. I've found that transferring the App to another computer via a USB stick will reset the Execute bit, while just downloading it on the target machine will leave it set.
The way I remember it, the problem with the VLARs is that the nature of the observation causes the code to just use 1 Compute Unit. That's why the higher end GPUs see more of a slowdown than the lower end GPUs when running the VLARs. The unroll function was added to send data to those CUs that are not being used. The general rule would be to set the unroll to the number of Compute Units. The app_info.xml is set to the number of the Low end GPUs, which is 2. For the 970 you would set it to 13, the 1080 needs 20 but that might slowdown the 970s. I would set it to 13, run it for an hour or so and then try it at 14. If that works acceptably you might try higher. It all depends on how the run times change as to the best setting.
<version_num>801</version_num>
<plan_class>cuda60</plan_class>
<cmdline>-bs -unroll 13</cmdline>
ID: 1836139 · Report as offensive     Reply Quote
rob smithProject Donor
Volunteer tester

Send message
Joined: 7 Mar 03
Posts: 15194
Credit: 251,445,793
RAC: 324,129
United Kingdom
Message 1836142 - Posted: 15 Dec 2016, 21:48:32 UTC

...giving it a go - might stay at 13 until sometime later tomorrow.

(Refresh my little grey-cell...do I have to restart BOINC for this to take effect, or will it happen automagically next time a task starts?)
Bob Smith
Member of Seti PIPPS (Pluto is a Planet Protest Society)
Somewhere in the (un)known Universe?
ID: 1836142 · Report as offensive     Reply Quote
TBar
Volunteer tester

Send message
Joined: 22 May 99
Posts: 3789
Credit: 186,212,712
RAC: 237,421
United States
Message 1836145 - Posted: 15 Dec 2016, 21:57:15 UTC - in response to Message 1836142.  

Changing the app_info requires a BOINC restart. It appears the tasks usually resume without any trouble, but, I have experienced an Overflow after a resume. So, I'd try to avoid restarting BOINC as much as possible ;-)
ID: 1836145 · Report as offensive     Reply Quote
rob smithProject Donor
Volunteer tester

Send message
Joined: 7 Mar 03
Posts: 15194
Credit: 251,445,793
RAC: 324,129
United Kingdom
Message 1836148 - Posted: 15 Dec 2016, 22:07:55 UTC

Re-start done, now let's watch this for a few minutes before I head to bed....
Bob Smith
Member of Seti PIPPS (Pluto is a Planet Protest Society)
Somewhere in the (un)known Universe?
ID: 1836148 · Report as offensive     Reply Quote
rob smithProject Donor
Volunteer tester

Send message
Joined: 7 Mar 03
Posts: 15194
Credit: 251,445,793
RAC: 324,129
United Kingdom
Message 1836151 - Posted: 15 Dec 2016, 22:14:57 UTC

Well that looks as if it is working nicely - I'll let it run for a few hours while I get some sleep then check for errors.
Times are "better than halved" - 1080 is down to about 5 minutes and 97s to about 8:40 - I've never had an Nvidia card do VLARs in that sort of time :-)
(It looks as if I escaped with no overflows this time...)
Bob Smith
Member of Seti PIPPS (Pluto is a Planet Protest Society)
Somewhere in the (un)known Universe?
ID: 1836151 · Report as offensive     Reply Quote
qbit
Volunteer tester
Avatar

Send message
Joined: 19 Sep 04
Posts: 630
Credit: 6,868,528
RAC: 0
Austria
Message 1836319 - Posted: 16 Dec 2016, 18:01:49 UTC

TBar, I think about switching to Linux on my cruncher and using your app. Currently running the latest OpenCL app from raistmer, do you think it's worth it?:
https://setiathome.berkeley.edu/show_host_detail.php?hostid=7563243
ID: 1836319 · Report as offensive     Reply Quote
TBar
Volunteer tester

Send message
Joined: 22 May 99
Posts: 3789
Credit: 186,212,712
RAC: 237,421
United States
Message 1836330 - Posted: 16 Dec 2016, 18:53:15 UTC - in response to Message 1836319.  

The only problem I see might be the vRam. Petri's App has only been used with GPUs having 2GB or more as far as I know. My two 750Ti cards are showing 1280MB used for the one running the screen and 1090 for the other one. I suppose you could try it and see. If it does work it would probably run quite a bit better on your dual core CPU. It looks as though you are using sleep to save CPU and the run times are kinda long if you are running one at a time. I'm currently running a dual core CPU with two GPU tasks and 1 CPU task and the machine is still very responsive. I've been trying to sort out a problem with this machine and have been trying different CPUs. It seems this old board doesn't support the quad core CPUs listed on the manufacturers site. I may be changing machines soon.

If you do try it, I'd suggest using a system based on Ubuntu 14.04.x, the Berkeley BOINC in your Home folder, and driver 367.xx.
ID: 1836330 · Report as offensive     Reply Quote
qbit
Volunteer tester
Avatar

Send message
Joined: 19 Sep 04
Posts: 630
Credit: 6,868,528
RAC: 0
Austria
Message 1836339 - Posted: 16 Dec 2016, 19:09:51 UTC

Thx TBar. Yes, I use use_sleep but I run 2 tasks at a time. I will decide tomorrow if I try or not. My main concern is that I have zero experience with linux. But how hard can it be? Also, it looks like new cuda apps for windows will still take some time.
ID: 1836339 · Report as offensive     Reply Quote
Rockhount
Avatar

Send message
Joined: 29 May 00
Posts: 34
Credit: 18,676,697
RAC: 23,964
Germany
Message 1836360 - Posted: 16 Dec 2016, 21:41:14 UTC

With lower RAM the unroll variable is limiting.
At bs-unroll 5 you need about 1090MiB. With bs-unroll 4 you could use it with fewer ram and less performance
Regards from nothern Germany
Roman

SETI@home classic workunits 207,059
SETI@home classic CPU time 1,251,095 hours

ID: 1836360 · Report as offensive     Reply Quote
qbit
Volunteer tester
Avatar

Send message
Joined: 19 Sep 04
Posts: 630
Credit: 6,868,528
RAC: 0
Austria
Message 1836489 - Posted: 17 Dec 2016, 13:40:43 UTC

OK, so today I decided to give it a try. I installed the latest Ubuntu (the LTS version) on my cruncher (dual boot, keeping my Windows installation). Seems to work fine besides WiFi beeing veeeeerrrrryyyy slow, but that shouldn't really matter much for testing.

So I went to the BOINC site and it tells me to install BOINC by using the following command:

sudo aptitude install boinc-client boinc-manager


But when I try it tells me "aptidude: command not found"

So I googled and found another page which tells me to install BOINC on Ubuntu by using

sudo apt-get install boinc-client


Trying this gives me the message "Unable to locate package boinc-client"


As I said before I'm a complete linux noob so I guess I need some help here. What am I doing wrong?
ID: 1836489 · Report as offensive     Reply Quote
TBar
Volunteer tester

Send message
Joined: 22 May 99
Posts: 3789
Credit: 186,212,712
RAC: 237,421
United States
Message 1836494 - Posted: 17 Dec 2016, 14:27:57 UTC - in response to Message 1836489.  
Last modified: 17 Dec 2016, 14:35:05 UTC

There are reasons I suggested 14.04. Especially with an older machine. Try this one, http://releases.ubuntu.com/trusty/ 64-bit PC (AMD64) desktop image
Then download BOINC from here, 7.2.42 Recommended version
Place the BOINC file in your Home folder and double click it. If that doesn't work, open a terminal, enter a ./ then the name of the file and hit enter.
You will need to install the libwx 2.8 packages to get BOINC to work. Open the Ubuntu software center in the Dock and install some of the recommended programs such as 7zip, restricted extras, VLC, and the Synaptic PACKAGE Manager. Open the Package Manager and search for libwx, you need the versions listed as libwxgtk2.8-0, just install the libwxgtk2.8-dev package and it will install everything you need to run BOINC.
You also need to install the driver from the Additional drivers tab of the System settings/Software & Updates, it should install 367.57.
Try that.
ID: 1836494 · Report as offensive     Reply Quote
qbit
Volunteer tester
Avatar

Send message
Joined: 19 Sep 04
Posts: 630
Credit: 6,868,528
RAC: 0
Austria
Message 1836496 - Posted: 17 Dec 2016, 14:39:57 UTC
Last modified: 17 Dec 2016, 14:45:19 UTC

Thx TBar!

More complicated then I thought. Also, the "additional driver" tab doesn't offer me anything. I just dl'd the 367.57 driver directly from nvidia and will try to install this in a minute.

*edit* Didn't really see that you suggested an older version. But why do those commands not work? I suppose commands don't change with each new version? Just trying to learn a bit about how Linux works.
ID: 1836496 · Report as offensive     Reply Quote
TBar
Volunteer tester

Send message
Joined: 22 May 99
Posts: 3789
Credit: 186,212,712
RAC: 237,421
United States
Message 1836501 - Posted: 17 Dec 2016, 14:54:15 UTC - in response to Message 1836496.  
Last modified: 17 Dec 2016, 14:55:52 UTC

Every version of Ubuntu is different. You Must be using the System I linked, 14.04.5, for my suggestions to work, BOINC 7.2.42 will Not work in newer versions of Ubuntu, Driver 367.57 will Not appear in Additional Drivers...etc.
ID: 1836501 · Report as offensive     Reply Quote
qbit
Volunteer tester
Avatar

Send message
Joined: 19 Sep 04
Posts: 630
Credit: 6,868,528
RAC: 0
Austria
Message 1836507 - Posted: 17 Dec 2016, 15:01:43 UTC - in response to Message 1836501.  

Every version of Ubuntu is different. You Must be using the System I linked, 14.04.5, for my suggestions to work, BOINC 7.2.42 will Not work in newer versions of Ubuntu, Driver 367.57 will Not appear in Additional Drivers...etc.

Damn! Ok, so if the suggested version of BOINC will not work in the OS I installed then I can stop right here. Maybe I try again another time. Thx again!
ID: 1836507 · Report as offensive     Reply Quote
TBar
Volunteer tester

Send message
Joined: 22 May 99
Posts: 3789
Credit: 186,212,712
RAC: 237,421
United States
Message 1836760 - Posted: 18 Dec 2016, 13:07:58 UTC

Here are some results from some recent Inconclusive tasks with WUs included;
http://www.arkayn.us/forum/index.php?topic=197.msg4500#msg4500
ID: 1836760 · Report as offensive     Reply Quote
rob smithProject Donor
Volunteer tester

Send message
Joined: 7 Mar 03
Posts: 15194
Credit: 251,445,793
RAC: 324,129
United Kingdom
Message 1836791 - Posted: 18 Dec 2016, 15:55:24 UTC

Well, I just increased unroll to 14.
First impression is that the 1080 likes it. However I'm not sure it has done anything (positive or negative) for the pair of 970s. I'll let it run for a few hours and see what happens.
Bob Smith
Member of Seti PIPPS (Pluto is a Planet Protest Society)
Somewhere in the (un)known Universe?
ID: 1836791 · Report as offensive     Reply Quote
TBar
Volunteer tester

Send message
Joined: 22 May 99
Posts: 3789
Credit: 186,212,712
RAC: 237,421
United States
Message 1836813 - Posted: 18 Dec 2016, 17:37:15 UTC - in response to Message 1836791.  

It looks as though it slowed down the 970s more than it sped up the 1080. Not much you can do except try to match the cards better. Changing out one 970 for the 980 would get you closer, but you'd still have one slow one. The Host seems to be doing well, it will probably be in the Top 20 in a couple weeks.
ID: 1836813 · Report as offensive     Reply Quote
Profile -= Vyper =-Project Donor
Volunteer tester
Avatar

Send message
Joined: 5 Sep 99
Posts: 1363
Credit: 505,053,019
RAC: 165,349
Sweden
Message 1836823 - Posted: 18 Dec 2016, 19:04:27 UTC - in response to Message 1836501.  

Huh!

But i'm using 16.04 on one of my hosts. Worked like a charm!
Though i started with installing only the server version with nothing more than OpenSSH server to continue from there though.
Good that more ppl are jumping the Linux bandwagon for eeking out the most of their hardware at the moment.

_________________________________________________________________________
Addicted to SETI crunching!
Founder of GPU Users Group
ID: 1836823 · Report as offensive     Reply Quote
Stephen "Heretic"Project Donor
Volunteer tester
Avatar

Send message
Joined: 20 Sep 12
Posts: 2628
Credit: 48,068,494
RAC: 130,648
Australia
Message 1836829 - Posted: 18 Dec 2016, 20:18:19 UTC - in response to Message 1836791.  

Well, I just increased unroll to 14.
First impression is that the 1080 likes it. However I'm not sure it has done anything (positive or negative) for the pair of 970s. I'll let it run for a few hours and see what happens.


. . Hi Rob,

. . Should not the unroll value be 13 or less for the 970s? I have gathered the impression it should not exceed the compute units on the GPU it is applied to.

Stephen

.
ID: 1836829 · Report as offensive     Reply Quote
Previous · 1 · 2 · 3 · 4 · 5 · 6 · 7 . . . 71 · Next

Message boards : Number crunching : Linux CUDA 'Special' App finally available, featuring Low CPU use


 
©2017 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.