Advanced search

Message boards : Number crunching : Linux for GPUGrid - general discussion

Author Message
Profile skgiven
Volunteer moderator
Volunteer tester
Avatar
Send message
Joined: 23 Apr 09
Posts: 3968
Credit: 1,995,359,260
RAC: 0
Level
His
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 22418 - Posted: 29 Oct 2011 | 23:57:10 UTC

Thread for discussing Linux use with GPUGrid, asking questions, giving tips and helping others.

bigtuna
Volunteer moderator
Send message
Joined: 6 May 10
Posts: 80
Credit: 98,784,188
RAC: 0
Level
Thr
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwat
Message 22460 - Posted: 3 Nov 2011 | 2:50:13 UTC - in response to Message 22418.

AFAIK Linux is ideal for running GPUGrid. Work units run faster than Windows 7 and roughly on par with Windows XP (correct me if I'm wrong here). Linux is free and much easier to use than one would think. I'm especially fond of the FatDog 64 bit distro. Small, fast, free and easy, what could be better?

One thing about Linux (or at least the version I run) is that it is picky about capitalization. For example SWAN_SYNC=0 must be in all caps.

Another great thing is that you can run Linux without installing it. I know this sounds crazy if you have never heard of such a thing but many Linux distros will run fine from CD/DVD media or can easily be installed to run from a USB stick so no hard drive is required at all. This is perfect if you want to keep a Windows install on your hard drive or if you want to forgo a hard drive all together.

Profile skgiven
Volunteer moderator
Volunteer tester
Avatar
Send message
Joined: 23 Apr 09
Posts: 3968
Credit: 1,995,359,260
RAC: 0
Level
His
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 22657 - Posted: 8 Dec 2011 | 17:55:07 UTC - in response to Message 22460.

SWAN_SYNC is now only recommended for Linux (6.14app), and is no longer needed on Windows operating systems (6.15app), though I think it still helps a bit.

On Ubuntu 11.10, from a terminal prompt use,

    sudo chmod a+w /etc/environment
    sudo echo SWAN_SYNC=0 >> /etc/environment
    sudo chmod go-w /etc/environment

Then restart.
If you type echo $SWAN_SYNC it should return the desired 0. (Thanks Dagorath)
____________
FAQ's

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

Profile skgiven
Volunteer moderator
Volunteer tester
Avatar
Send message
Joined: 23 Apr 09
Posts: 3968
Credit: 1,995,359,260
RAC: 0
Level
His
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 24197 - Posted: 31 Mar 2012 | 13:20:20 UTC - in response to Message 22657.

Damaraland has made another couple of significant Linux posts recently,


Mowskwoz spurred me on to look into coolbits,


Thanks guys,

Please post any comments or discuss these HOW TO's here, rather than following up in the FAQ's.
____________
FAQ's

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

Profile vitalidze
Send message
Joined: 15 May 09
Posts: 20
Credit: 239,712,351
RAC: 0
Level
Leu
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 24214 - Posted: 2 Apr 2012 | 17:51:37 UTC - in response to Message 22460.

I do not know why do you say that linux is as fast as XP and faster than 7. I have had just everything up side down. my linux system was at least 10% slower especially on 12-20 hours tasks. 1 core was free and it was almost always 70-100% utilized. i5 2500K@4600. The only advantage of linux is that it is free i see no advantage in performance at all only HCC in WCG uses the advantage.

Profile skgiven
Volunteer moderator
Volunteer tester
Avatar
Send message
Joined: 23 Apr 09
Posts: 3968
Credit: 1,995,359,260
RAC: 0
Level
His
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 24221 - Posted: 3 Apr 2012 | 12:23:38 UTC - in response to Message 24214.
Last modified: 3 Apr 2012 | 12:24:08 UTC

I have 2 systems that are almost identical:
i7-2600 stock (for now), 8GB, GTX470. One has Ubuntu 11.10, the other 2003R2 server (both x64).
My Win system has an OC'ed GTX 470 @ 680MHz, so it's 12% faster. Because performance is linear wrt clock speed (up to the point when the GPU becomes unstable), the OC can be accounted for.

Just looking at a few of Nate's tasks:

Ubuntu system,
I4R85-NATHAN_FAX4-16-100-RND0016_0 3316448 2 Apr 2012 | 1:43:14 UTC 2 Apr 2012 | 14:32:38 UTC Completed and validated 44,294.76 4,098.40 71,400.00
I3R65-NATHAN_FAX4-9-100-RND0493_0 3315075 1 Apr 2012 | 13:46:50 UTC 2 Apr 2012 | 2:13:52 UTC Completed and validated 44,482.68 4,147.50 71,400.00

Average run time, 44.4Ks

Win system,
I2R48-NATHAN_FAX4-8-100-RND8586_0 3316276 1 Apr 2012 | 23:52:16 UTC 2 Apr 2012 | 12:34:25 UTC Completed and validated 41,657.19 5,764.06 71,400.00
I5R59-NATHAN_FAX4-8-100-RND0267_1 3310663 31 Mar 2012 | 19:45:16 UTC 1 Apr 2012 | 7:53:44 UTC Completed and validated 41,087.66 5,785.64 71,400.00
I3R33-NATHAN_FAX4-9-100-RND1015_0 3308808 30 Mar 2012 | 9:15:15 UTC 31 Mar 2012 | 0:41:15 UTC Completed and validated 42,084.28 5,778.84 71,400.00

Average run time, 41.6Ks

So Nate's tasks are 44.4/41.6= 5.8% faster on the 12% OC'ed GTX470 on Windows, before taking account of the GPU speed. Obviously the 12% OC is more than 5.8%, so Linux is faster than W2003Server x64 (the fastest Windows system).
44.4/(41.6*1.12)=0.95
After taking account for the GPU OC its clear that Linux (for Nate's tasks) is ~5% faster. Historically, this is about right, but there might be some variation from app to app and between task performances.

These results were without using SWAN_SYNC, and each system was using 7 from 8 threads to crunch mixed CPU tasks at the time.

WRT the performance of CPU tasks, they are going to vary from project to project. Some tasks such as HCC are known to be faster, some are the same (C4CW), and some are faster on Windows (most Climate models).

Unfortunately there isn't a relative performance list within the Boinc community comparing Windows and Linux. That would need identical (hardware and Boinc config) test systems, and testing of each operating system (Win XP, 2003, Vista, 2008, W7, W8, all the various Linux flavors) and for both x64 and x86...

What way did you determine your GPU utilization on Linux, nvidia-smi?
____________
FAQ's

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

5pot
Send message
Joined: 8 Mar 12
Posts: 411
Credit: 2,083,882,218
RAC: 0
Level
Phe
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 24236 - Posted: 4 Apr 2012 | 14:07:12 UTC
Last modified: 4 Apr 2012 | 14:15:34 UTC

Thank you all for your advice about getting rid of nouveau , was kind of a pain. If anyone has trouble installing dual-boot, and cannot seem to get acpi=off during install, as it turns out my MSI mobo had a setting for this under Power Management in CMOS, which had to be disabled for install to work. IT WOULD NOT INSTALL WITHOUT THIS BEING DISABLED IN CMOS. I tried install acpi=off with no luck.

Found this interesting though (same rig):

(Windows)
Measured floating point speed

3550.41 million ops/sec

Measured integer speed

8298.33 million ops/sec

Ubuntu 11.10

Measured floating point speed

3306.03 million ops/sec

Measured integer speed

16926.02 million ops/sec

Coolbits and SWAN_SYNC functions were a breeze to set up BTW. And skgiven first let me thank you for your sticked posts, and I did notice that sudo apt-get --purge remove xserver-xorg-video-nouveau worked while
sudo apt-get --purge remove xserver-xorg-video did not (the latter is on your color coded FAQ) Dont know why, I'm terrible with Ubuntu (more of a copy and paste guy)

EDIT:

This was also needed (i did things somewhat choppy, but everything worked)I had to do this to get X Server since jockey would not locate NVIDIA drivers while nouveau was installed. (May be talking out my a** here, but it worked for me)
sudo apt-add-repository ppa:ubuntu-x-swat/x-updates
sudo apt-get update
sudo apt-get install nvidia-current

I did this before uninstalling nouveau btw.

Again thank you everyone

Profile Damaraland
Send message
Joined: 7 Nov 09
Posts: 152
Credit: 16,181,924
RAC: 0
Level
Pro
Scientific publications
watwatwatwatwatwatwatwatwat
Message 24241 - Posted: 4 Apr 2012 | 16:42:16 UTC - in response to Message 24236.

Thank you all for your advice about getting rid of nouveau, was kind of a pain.

I'm really pleased to hear that! It took me a while put together everything found on Internet and put it together to make it work and to make the tutorial.

Glad someone took advantage of it.

____________
HOW TO - Full installation Ubuntu 11.10

5pot
Send message
Joined: 8 Mar 12
Posts: 411
Credit: 2,083,882,218
RAC: 0
Level
Phe
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 24247 - Posted: 4 Apr 2012 | 19:30:27 UTC

I do have one question; the nvidia-smi command works, but it my GPU usage is displayed as N/A, the memory load, fan %, and temps are all there (recognizes it's a 570), but the GPU usage is not present, and this I would like to see if there's a difference.

What command, and/or program would help me out with this?

Again, thank you for your hard work.

5pot
Send message
Joined: 8 Mar 12
Posts: 411
Credit: 2,083,882,218
RAC: 0
Level
Phe
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 24248 - Posted: 4 Apr 2012 | 23:40:08 UTC

How do you disable swan_sync? I think since I can't change my LLC on this board, its causing my system monitor to look kinda crazy. When gpu's CPU is at let's say 50 and goes.to 100 a different core will decrease by 50. Thinking its a voltage issue due to OC. (vdroop or vdrop?). Cores fighting for power to stay at 100. I THINK.

5pot
Send message
Joined: 8 Mar 12
Posts: 411
Credit: 2,083,882,218
RAC: 0
Level
Phe
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 24259 - Posted: 5 Apr 2012 | 14:08:48 UTC
Last modified: 5 Apr 2012 | 14:14:22 UTC

Ok. So I think I finally get the swan variable. By freeing up one core and allowing gpu to utilize more of that core, it "steals" that power from other cores on my I5 2500k. Meaning since boinc is told use 75% with four tasks running (3 WCG and 1 GPUgrid) at any time the gpu task uses more than the .49 it states on task, it uses the WCG cores to total 3 cores running? So is there no way to utilize that remaining core? I mean if I could tell boinc use 100% of cores, w/o running another WCG CPU task? That way all cores would show 100% usage instead of 75% spread across 4 tasks?

Edit
at times when I say.core I mean processing power of total CPU. Hope I'm making sense.

Profile skgiven
Volunteer moderator
Volunteer tester
Avatar
Send message
Joined: 23 Apr 09
Posts: 3968
Credit: 1,995,359,260
RAC: 0
Level
His
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 24263 - Posted: 5 Apr 2012 | 16:26:41 UTC - in response to Message 24259.
Last modified: 5 Apr 2012 | 16:49:28 UTC

Boinc Manager does not know SWAN_SYNC exists.
So to effectively use SWAN_SYNC you must instruct Boinc to use 1 less CPU core/thread when SWAN_SYNC is enabled. If you do not then one core will be used by SWAN_SYNC and Boinc will also try to use all the cores. In this situation there is more chance of failures. So if you use SWAN_SYNC you must free up a CPU core/thread.

On an i5 quad (no HT) there would be little point freeing up more than one core.
On an i7 (HT for 8 or more threads) freeing up an additional thread should improve GPU performance a bit further.
The bigger the GPU the more important this becomes.

Depending on the GPU app and tasks, freeing up CPU threads to expedite GPU tasks often works for up to 2 or 3 threads, but after that it usually makes little or no difference (dual cards and high end AMD cards on OpenCL excepted).
It's up to the individual to decide how much resources to commit to each project. I'm presently running C4CW tasks along with GPUGrid tasks. I found that on each occasion after freeing up 1 then 2 and then 3 CPU threads the C4CW tasks progressively took less time to run. >10% faster with 2 threads free still means doing less CPU work, but probably under 17.5% less rather than 25% less. With only 1 thread free (not including SWAN_SYNC if that's used) the loss is often ~1/2 of a CPU thread. That's only losing 6% of your CPU. The benefits are improved system responsiveness, reduction in task failures, improvement in GPU performance. With additional threads free the CPU can better manage it's temperature (by moving threads to cooler cores) preventing overheating. Obviously there is huge variation between different CPU apps and the GPU performance impact can vary considerably.
____________
FAQ's

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

5pot
Send message
Joined: 8 Mar 12
Posts: 411
Credit: 2,083,882,218
RAC: 0
Level
Phe
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 24264 - Posted: 5 Apr 2012 | 17:16:05 UTC

I too have noticed reduction in run times on the CPU tasks, which is why I've decided to leave it alone, but also because such as you stated about failures occuring, I noticed as soon as I allowed 100% usage late last night to see what would happen, my gpu temp. Plummeted from 52 down to 41, and even though I can't see gpu usage, i figured that had decreased drastically, so i reverted back to 75%. Still wish they weren't "fighting" for the 75% , but all in all I'm glad to be back to Linux.

5pot
Send message
Joined: 8 Mar 12
Posts: 411
Credit: 2,083,882,218
RAC: 0
Level
Phe
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 24276 - Posted: 6 Apr 2012 | 13:49:16 UTC

What.command disabled swan_sync ?

5pot
Send message
Joined: 8 Mar 12
Posts: 411
Credit: 2,083,882,218
RAC: 0
Level
Phe
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 24303 - Posted: 7 Apr 2012 | 14:45:37 UTC

So let's see if I can get this 680 up and running linux with some help. After trying many times last night to install ubuntu, I kept getting grub-efi failed to install /target, so after unplugging Windows SATA drive out of frustration, thinking that installing on my fresh SSD would fix the issue (since windows would not be plugged in). This didn't happen. No matter what I do, I keep getting that exact same message. After that message appears, ubuntu loads off usb stick. Even after uninstalling Windows completely (using cd to delete partitions), I got the SAME error message again. Beginning to lose a little faith in Ubuntu, I've installed it quite a few times and have never had this kind of problem. The "Windows Boot Manager" option is still available on my BIOS screen, and I am asking for a detailed, hopefully easy, solution for my problems. Both drivers have been cleared entirely (as far as I can tell), and yet 11.10 just does not seem to what to install.

Profile skgiven
Volunteer moderator
Volunteer tester
Avatar
Send message
Joined: 23 Apr 09
Posts: 3968
Credit: 1,995,359,260
RAC: 0
Level
His
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 24345 - Posted: 10 Apr 2012 | 10:08:03 UTC - in response to Message 24303.
Last modified: 10 Apr 2012 | 10:24:03 UTC

The Ubuntu 11.10 repository drivers (280.13) don't support your GTX680.

I suggest you stick to Windows; the researchers are testing on Linux and will release Beta's for Windows when they are ready to test.

If you really must use that card with Linux, you will have to figure out how to download and integrate the latest drivers into Linux. You might need to uninstall the present drivers, download the latest drivers, remove the card, boot to grub and configure Ubuntu to work with the new drivers...
See these threads,
HOW TO - Install NVIDIA drivers on Linux 64 bit (generic)
HOW TO - Full installation Ubuntu 11.10

Ubuntu 12.04 will have Long Term Support (LTS) but at present it's Beta only.
I have not tried it but the 12.04 repositories are more likely to have the latest Boinc client and drivers. IF you are going to try it remember to get the x64bit OS.
____________
FAQ's

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

5pot
Send message
Joined: 8 Mar 12
Posts: 411
Credit: 2,083,882,218
RAC: 0
Level
Phe
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 24346 - Posted: 10 Apr 2012 | 12:34:49 UTC

Yea i gave up skgiven on trying to set that one.up, I'm not sure why the drivers wouldnt allow.grub to install, and NVIDIA has drivers for the 680 out, or are they the sleep bug one and u cant disable it in Linux like Windows? Screen worked and if anything since my mobo is a z series I could in theory use my vga on it.until drivers were installed. As of.right now I am unfortunately using it for.primegrids sieve apps, till WCG betas are grabbed or I get the word from gianni to attach. Thx for the reply though.

Post to thread

Message boards : Number crunching : Linux for GPUGrid - general discussion

//