V8 CUDA for Linux?

Message boards : Number crunching : V8 CUDA for Linux?
Message board moderation

To post messages, you must log in.

1 · 2 · 3 · 4 . . . 10 · Next

AuthorMessage
Profile Francesco Forti
Avatar

Send message
Joined: 24 May 00
Posts: 334
Credit: 204,421,005
RAC: 15
Switzerland
Message 1758156 - Posted: 22 Jan 2016, 11:21:22 UTC

Hi!
I have a mix of windows and linux hosts (7.2.41 x64).
Now I see that my 4 win hosts are cruncing V8(cuda42, ...23, ..50)
No V8 cuda job for my 4 Linux host with cuda.
Have I something to do?

Thanks
FF
ID: 1758156 · Report as offensive
Profile William
Volunteer tester
Avatar

Send message
Joined: 14 Feb 13
Posts: 2037
Credit: 17,689,662
RAC: 0
Message 1758165 - Posted: 22 Jan 2016, 11:55:14 UTC

Short answer: Wait for somebody (e.g. Jason) to do a Linux version of x41zi, hand it to Eric for beta testing and after successful beta testing have Eric transfer the app to Main.
A person who won't read has no advantage over one who can't read. (Mark Twain)
ID: 1758165 · Report as offensive
OTS
Volunteer tester

Send message
Joined: 6 Jan 08
Posts: 371
Credit: 20,533,537
RAC: 0
United States
Message 1758176 - Posted: 22 Jan 2016, 12:25:35 UTC
Last modified: 22 Jan 2016, 12:28:50 UTC

Based on the posting in the "News" area from Eric, it appears there is a problem with the Linux64/Nvidia app. Does anyone know if it is a major or minor setback and if there is an estimate on when it will be available?
ID: 1758176 · Report as offensive
Profile Francesco Forti
Avatar

Send message
Joined: 24 May 00
Posts: 334
Credit: 204,421,005
RAC: 15
Switzerland
Message 1772446 - Posted: 18 Mar 2016, 16:14:36 UTC

Any update?
I have 3500 Gflops sleeping ....
ID: 1772446 · Report as offensive
Profile jason_gee
Volunteer developer
Volunteer tester
Avatar

Send message
Joined: 24 Nov 06
Posts: 7489
Credit: 91,093,184
RAC: 0
Australia
Message 1772450 - Posted: 18 Mar 2016, 16:39:10 UTC - in response to Message 1772446.  

Some Linux library issues preventing stock distribution'for older kernels (namely GLIBC), and also partly because Linux Boinc doesn't seem to display/detect Cuda driver version (?, will happily be corrected if that's been fixed).

Temporary solution (if you have a newish Linux Kernel and driver version) is to use the build on my site ( http://jgopt.org/download.html under anonymous platform.

Will probably end up having to build an old Cuda version on an actual old distro, in order to make it use old GLibC, and hope the scheduler can be taught enough to figure out what to send.
"Living by the wisdom of computer science doesn't sound so bad after all. And unlike most advice, it's backed up by proofs." -- Algorithms to live by: The computer science of human decisions.
ID: 1772450 · Report as offensive
TBar
Volunteer tester

Send message
Joined: 22 May 99
Posts: 5204
Credit: 840,779,836
RAC: 2,768
United States
Message 1772455 - Posted: 18 Mar 2016, 16:53:35 UTC - in response to Message 1772446.  

There are currently Two Linux CUDA Apps available. The CUDA 42 version is best for your pre-fermi cards with less than driver version 331. I think the CUDA 6.0 version needs driver 331 or maybe 334, but even then doesn't seem to work well on my pre-fermi cards. The CUDA 42 version needs driver 295 to 337. I would suggest using the Berkeley version of BOINC where the BOINC folder is in your Home folder and contains ALL the BOINC components, I prefer version 7.2.33. This makes changing Apps and others things very simple as opposed to the repository version which scatters the components among different folders. It also makes moving the host from one Linux version to another very simple. You won't get the same times you do in XP, but they will be comparable to times in other OSes.

CUDA 42: http://www.arkayn.us/forum/index.php?topic=197.msg4489
CUDA 60: http://jgopt.org/download.html
ID: 1772455 · Report as offensive
Profile Francesco Forti
Avatar

Send message
Joined: 24 May 00
Posts: 334
Credit: 204,421,005
RAC: 15
Switzerland
Message 1772684 - Posted: 19 Mar 2016, 20:45:11 UTC - in response to Message 1772455.  

There are currently Two Linux CUDA Apps available. The CUDA 42 version is best for your pre-fermi cards with less than driver version 331. I think the CUDA 6.0 version needs driver 331 or maybe 334, but even then doesn't seem to work well on my pre-fermi cards. The CUDA 42 version needs driver 295 to 337. I would suggest using the Berkeley version of BOINC where the BOINC folder is in your Home folder and contains ALL the BOINC components, I prefer version 7.2.33. This makes changing Apps and others things very simple as opposed to the repository version which scatters the components among different folders. It also makes moving the host from one Linux version to another very simple. You won't get the same times you do in XP, but they will be comparable to times in other OSes.

CUDA 42: http://www.arkayn.us/forum/index.php?topic=197.msg4489
CUDA 60: http://jgopt.org/download.html

Ok, I (perhaps) understand. Now I'm running standard 7.2.42 (x64) and Nvidia-304.
304 is the only way to run Astropulse V7 7.08 opencl_nvidia_100 on this GTX 560.
In the driver management for my Linux Mint 17.3 I can get also 340 and 352 (and connected updates) but not 331 or 334 (and not 295 to 337).

What can I do?
ID: 1772684 · Report as offensive
TBar
Volunteer tester

Send message
Joined: 22 May 99
Posts: 5204
Credit: 840,779,836
RAC: 2,768
United States
Message 1772689 - Posted: 19 Mar 2016, 21:05:44 UTC - in response to Message 1772684.  
Last modified: 19 Mar 2016, 21:55:09 UTC

Driver 304 will work perfectly fine with the CUDA 42 App. All you have to do is install it and the included libraries.
I've tested the CUDA 42 App with Ubuntu 11.04 and driver 295 without any problems, others have used it with other Linux distros with kernel 2.6.x and up with 100% success. For pre-fermi cards you would use a driver from 295 to 337, 304 falls inside this range. For Fermi and up there isn't a limit, and the App works very well on my 750Ti with current drivers.

Here are the nVidia Linux drivers if you would like to use something different; http://www.nvidia.com/object/linux-amd64-display-archive.html
In my tests the Driver 343.x versions have a broken OpenCL build program, so, I wouldn't bother with those versions.
ID: 1772689 · Report as offensive
TBar
Volunteer tester

Send message
Joined: 22 May 99
Posts: 5204
Credit: 840,779,836
RAC: 2,768
United States
Message 1772698 - Posted: 19 Mar 2016, 22:55:45 UTC - in response to Message 1772684.  

304 is the only way to run Astropulse V7 7.08 opencl_nvidia_100 on this GTX 560.

After thinking this over the only explanation I can think of is the Problem with the Repository versions of 340 and 352. Many people, including myself, have problems with BOINC not seeing OpenCL with these drivers from the Repository. If you were to use one of the recent drivers from Nvidia you shouldn't have this problem. Your Series 2xx cards should say at driver 337 or below, the Series 4, 5, and 6 cards should easily run the recent drivers from nVidia without the problem of the Missing OpenCL in BOINC.
You'll just have to remove the Repository drivers and install the newer nVidia drivers manually if you want to run a newer driver on your newer cards.
ID: 1772698 · Report as offensive
Profile jason_gee
Volunteer developer
Volunteer tester
Avatar

Send message
Joined: 24 Nov 06
Posts: 7489
Credit: 91,093,184
RAC: 0
Australia
Message 1772740 - Posted: 20 Mar 2016, 3:26:02 UTC - in response to Message 1772684.  

What can I do?


304 driver should run the Cuda 4.2 version properly, if that's what you need for AP/OpenCL.
"Living by the wisdom of computer science doesn't sound so bad after all. And unlike most advice, it's backed up by proofs." -- Algorithms to live by: The computer science of human decisions.
ID: 1772740 · Report as offensive
W3Perl Project Donor
Volunteer tester

Send message
Joined: 29 Apr 99
Posts: 251
Credit: 3,696,783,867
RAC: 12,606
France
Message 1772770 - Posted: 20 Mar 2016, 8:31:28 UTC - in response to Message 1772698.  


After thinking this over the only explanation I can think of is the Problem with the Repository versions of 340 and 352. Many people, including myself, have problems with BOINC not seeing OpenCL with these drivers from the Repository. If you were to use one of the recent drivers from Nvidia you shouldn't have this problem. Your Series 2xx cards should say at driver 337 or below, the Series 4, 5, and 6 cards should easily run the recent drivers from nVidia without the problem of the Missing OpenCL in BOINC.
You'll just have to remove the Repository drivers and install the newer nVidia drivers manually if you want to run a newer driver on your newer cards.



I had to install nvidia-modprobe in order GPU to be seen by boinc.

I have GTX 950, do I need to use opencl_nvidia_sah, opencl_nvidia_SoG or cuda60 ?
Both of them are working...
Both of them display '(0.173 CPUs + 1 NVIDIA GPU)' but require nearly 100% CPU to get full speed.

What's the main difference between opencl_nvidia_sah and cuda60 ?
ID: 1772770 · Report as offensive
Profile jason_gee
Volunteer developer
Volunteer tester
Avatar

Send message
Joined: 24 Nov 06
Posts: 7489
Credit: 91,093,184
RAC: 0
Australia
Message 1772771 - Posted: 20 Mar 2016, 8:35:36 UTC - in response to Message 1772770.  

What's the main difference between opencl_nvidia_sah and cuda60 ?


The Cuda60 build is basically generic standard code updated to work with v8, and would 'feel' familiar, in terms of CPU usage and throughput etc. The OpenCL build, If I understand correctly, can be faster for some situations, at some CPU usage cost and a different 'feel'. So best for your system will depend on your system and expectations (heavy doses of your mileage may vary etc)
"Living by the wisdom of computer science doesn't sound so bad after all. And unlike most advice, it's backed up by proofs." -- Algorithms to live by: The computer science of human decisions.
ID: 1772771 · Report as offensive
TBar
Volunteer tester

Send message
Joined: 22 May 99
Posts: 5204
Credit: 840,779,836
RAC: 2,768
United States
Message 1772838 - Posted: 20 Mar 2016, 17:14:59 UTC - in response to Message 1772770.  
Last modified: 20 Mar 2016, 17:24:30 UTC


After thinking this over the only explanation I can think of is the Problem with the Repository versions of 340 and 352. Many people, including myself, have problems with BOINC not seeing OpenCL with these drivers from the Repository. If you were to use one of the recent drivers from Nvidia you shouldn't have this problem. Your Series 2xx cards should say at driver 337 or below, the Series 4, 5, and 6 cards should easily run the recent drivers from nVidia without the problem of the Missing OpenCL in BOINC.
You'll just have to remove the Repository drivers and install the newer nVidia drivers manually if you want to run a newer driver on your newer cards.


I had to install nvidia-modprobe in order GPU to be seen by boinc.

I have GTX 950, do I need to use opencl_nvidia_sah, opencl_nvidia_SoG or cuda60 ?
Both of them are working...
Both of them display '(0.173 CPUs + 1 NVIDIA GPU)' but require nearly 100% CPU to get full speed.

What's the main difference between opencl_nvidia_sah and cuda60 ?

Generally the CUDA Apps preform the Lower Angle Range tasks better than OpenCL. On the Higher Angle Range tasks both Apps are about the same. Usually the OpenCL Apps Slow down and trend to create more Screen Lag on the Longer tasks. It's especially noticeable on the Mid and Lower End cards. Recent tasks seem to be mostly Lower Angle Range.

The OpenCL Mac Apps don't appear to be working very well on the LapTops, just look in your Inconclusive List and you'll be sure to find some examples. Some of the LapTops aren't preforming that badly but still appear to be running much slower than the CUDA Apps on the same Mac Laptops. Look at Your Mac LapTop, http://setiathome.berkeley.edu/results.php?hostid=7523831&offset=40. I'm finding it hard to believe your 320M could be that much slower than my 8800 GT & GTS 250, or even the 650M, http://setiathome.berkeley.edu/results.php?hostid=7366840&offset=260 Right now my old cards are running in Linux, but they have similar times in OSX running Mavericks or lower using the CUDA 42 App here, http://www.arkayn.us/forum/index.php?topic=191.msg4411#msg4411 Unfortunately nVidia dropped support for the Old Pre-Fermi cards in Yosemite and above, they do work well in Mavericks and lower with CUDA.
ID: 1772838 · Report as offensive
Profile Francesco Forti
Avatar

Send message
Joined: 24 May 00
Posts: 334
Credit: 204,421,005
RAC: 15
Switzerland
Message 1773229 - Posted: 22 Mar 2016, 11:21:27 UTC
Last modified: 22 Mar 2016, 11:50:26 UTC

I have changed driver form 304 to 352 using Driver Manager in Linux Mint.

Now I get

mar 22 mar 2016 12:17:26 CET |  | Starting BOINC client version 7.2.42 for x86_64-pc-linux-gnu
mar 22 mar 2016 12:17:26 CET |  | log flags: file_xfer, sched_ops, task
mar 22 mar 2016 12:17:26 CET |  | Libraries: libcurl/7.35.0 OpenSSL/1.0.1f zlib/1.2.8 libidn/1.28 librtmp/2.3
mar 22 mar 2016 12:17:26 CET |  | Data directory: /var/lib/boinc-client
mar 22 mar 2016 12:17:26 CET |  | No usable GPUs found
mar 22 mar 2016 12:17:26 CET |  | app version refers to missing GPU type NVIDIA
mar 22 mar 2016 12:17:26 CET | SETI@home | Application uses missing NVIDIA GPU
mar 22 mar 2016 12:17:26 CET |  | app version refers to missing GPU type NVIDIA
mar 22 mar 2016 12:17:26 CET | SETI@home | Application uses missing NVIDIA GPU
mar 22 mar 2016 12:17:26 CET |  | Host name: cil-System-Product-Name
mar 22 mar 2016 12:17:26 CET |  | Processor: 8 GenuineIntel Intel(R) Core(TM) i7-2600 CPU @ 3.40GHz [Family 6 Model 42 Stepping 7]
mar 22 mar 2016 12:17:26 CET |  | Processor features: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush dts acpi mmx fxsr sse sse2 ss ht tm pbe syscall nx rdtscp lm constant_tsc arch_perfmon pebs bts rep_good nopl xtopology nonstop_tsc aperfmperf eagerfpu pni pclmulqdq dtes64 monitor ds_cpl vmx smx est tm2 ssse3 cx16 xtpr pdcm pcid sse4_1 sse4_2 x2apic popcnt tsc_deadline_timer aes xsave avx lahf_lm ida arat epb pln pts dtherm tpr_shadow vnmi flexpriority ept vpid xsaveopt
mar 22 mar 2016 12:17:26 CET |  | OS: Linux: 3.19.0-32-generic
mar 22 mar 2016 12:17:26 CET |  | Memory: 7.77 GB physical, 7.97 GB virtual
mar 22 mar 2016 12:17:26 CET |  | Disk: 212.09 GB total, 126.47 GB free
mar 22 mar 2016 12:17:26 CET |  | Local time is UTC +1 hours

ID: 1773229 · Report as offensive
Profile Francesco Forti
Avatar

Send message
Joined: 24 May 00
Posts: 334
Credit: 204,421,005
RAC: 15
Switzerland
Message 1773235 - Posted: 22 Mar 2016, 11:48:29 UTC
Last modified: 22 Mar 2016, 11:54:35 UTC

same for 340.

I see that the driver update via driver manager modify nvidia and opencl packages too.
No usable GPUs found.
If I use 304 drivers (updated or not) I find usable GPU but no V8-Invidia job.
ID: 1773235 · Report as offensive
Profile Francesco Forti
Avatar

Send message
Joined: 24 May 00
Posts: 334
Credit: 204,421,005
RAC: 15
Switzerland
Message 1773237 - Posted: 22 Mar 2016, 12:03:12 UTC - in response to Message 1772770.  
Last modified: 22 Mar 2016, 12:03:52 UTC

I had to install nvidia-modprobe in order GPU to be seen by boinc.

Great suggestion!

Now with 340 drivers I get:

mar 22 mar 2016 12:59:26 CET | | CUDA: NVIDIA GPU 0: GeForce GTX 560 (driver version unknown, CUDA version 6.5, compute capability 2.1, 1023MB, 761MB available, 1243 GFLOPS peak)
mar 22 mar 2016 12:59:26 CET | | OpenCL: NVIDIA GPU 0: GeForce GTX 560 (driver version 340.96, device version OpenCL 1.1 CUDA, 1023MB, 761MB available, 1243 GFLOPS peak)

Prevoiusly no GPU available.

Now I wait for job and if I get it, I'll inform you.
ID: 1773237 · Report as offensive
W3Perl Project Donor
Volunteer tester

Send message
Joined: 29 Apr 99
Posts: 251
Credit: 3,696,783,867
RAC: 12,606
France
Message 1773290 - Posted: 22 Mar 2016, 21:21:49 UTC - in response to Message 1773237.  


Now I wait for job and if I get it, I'll inform you.


There is no nvidia wu available for v8 yet.
https://setiathome.berkeley.edu/apps.php

Seems there are many issues to fix before a release will be available. Hope it will be release soon

If you really want to process nvidia wu, you can join beta where
opencl_nvidia_SoG, opencl_nvidia_sah and cuda60 are the binaries.

https://setiathome.berkeley.edu/beta/apps.php


But I don't know if joining beta is only helful to debug or if processed wu
can be used for the seti project.
ID: 1773290 · Report as offensive
TBar
Volunteer tester

Send message
Joined: 22 May 99
Posts: 5204
Credit: 840,779,836
RAC: 2,768
United States
Message 1773303 - Posted: 22 Mar 2016, 21:45:58 UTC - in response to Message 1773237.  

You can see the available Linux Apps here, http://setiathome.berkeley.edu/apps.php
As you see there aren't any Linux nVidia Apps on Main. To run Linux nVidia Apps on Main you will have to manually install the previously mentioned Apps.
The CUDA 60 App requires Driver 331 or higher and gives longer runtimes on Pre-Fermi cards, as seen here; http://setiweb.ssl.berkeley.edu/beta/results.php?hostid=76565
The CUDA 42 App works better on the Pre-Fermi cards giving runtimes on a GTS 250 similar to these, http://setiweb.ssl.berkeley.edu/beta/results.php?hostid=72013

It is unknown as to when a Linux nVidia App will be released on Main.
ID: 1773303 · Report as offensive
Profile Francesco Forti
Avatar

Send message
Joined: 24 May 00
Posts: 334
Credit: 204,421,005
RAC: 15
Switzerland
Message 1773307 - Posted: 22 Mar 2016, 22:04:04 UTC
Last modified: 22 Mar 2016, 22:04:53 UTC

Infact No job:

mar 22 mar 2016 22:57:25 CET | SETI@home | Sending scheduler request: To fetch work.
mar 22 mar 2016 22:57:25 CET | SETI@home | Requesting new tasks for CPU and NVIDIA
mar 22 mar 2016 22:57:27 CET | SETI@home | Scheduler request completed: got 0 new tasks
mar 22 mar 2016 22:57:27 CET | SETI@home | No tasks sent
mar 22 mar 2016 22:57:27 CET | SETI@home | No tasks are available for SETI@home v7
mar 22 mar 2016 22:57:27 CET | SETI@home | No tasks are available for AstroPulse v7
mar 22 mar 2016 22:57:27 CET | SETI@home | No tasks are available for SETI@home v8
mar 22 mar 2016 22:57:27 CET | SETI@home | This computer has reached a limit on tasks in progress
ID: 1773307 · Report as offensive
Profile jason_gee
Volunteer developer
Volunteer tester
Avatar

Send message
Joined: 24 Nov 06
Posts: 7489
Credit: 91,093,184
RAC: 0
Australia
Message 1773311 - Posted: 22 Mar 2016, 22:24:42 UTC - in response to Message 1773303.  

It is unknown as to when a Linux nVidia App will be released on Main.


Yes, It's a mystery to me why the driver version is not detected/reported by the client on Linux (might do some homework there at some point). That's going to make things tricky for the scheduler to send the right application. For stock probably will have to guess from the Max compute capability GPU installed, put up with conservative issue, and encourage anonymous platform if not getting the one you wanted.
"Living by the wisdom of computer science doesn't sound so bad after all. And unlike most advice, it's backed up by proofs." -- Algorithms to live by: The computer science of human decisions.
ID: 1773311 · Report as offensive
1 · 2 · 3 · 4 . . . 10 · Next

Message boards : Number crunching : V8 CUDA for Linux?


 
©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.