Message boards : BOINC client : BOINC 7 development discussion thread.
Message board moderation
Previous · 1 · 2 · 3 · 4 · 5 · 6 . . . 11 · Next
Author | Message |
---|---|
Send message Joined: 29 Aug 05 Posts: 15566 |
BOINC 7.4.21 released for testing for all platforms. Rom Walton, BOINC developer wrote: Howdy Folks, |
Send message Joined: 14 Jun 11 Posts: 48 |
Although not an alpha tester, I downloaded, installed and am running the 7.4.21 version of BOINC. As far as I am concerned, no problems whatsoever occurred. Runs fine and smooth. I am in five projects involved, and my system specs are: AMD Phenom II X4 945 2 GB RAM Win XP SP3 Keep up the good work guys! Thanks a lot! :) |
Send message Joined: 29 Aug 05 Posts: 15566 |
BOINC 7.4.22 available for testing for all platforms. Rom Walton, BOINC Developer Windows and Linux wrote: Howdy Folks, |
Send message Joined: 14 Jun 11 Posts: 48 |
Same as my post from 15 Sep 2014, now I am using version 7.4.22: Works fine, no problems at all. :) |
Send message Joined: 29 Aug 05 Posts: 15566 |
Charlie Fenton, BOINC Developer for Macintosh wrote: Hello Mac testers, |
Send message Joined: 29 Aug 05 Posts: 15566 |
When BOINC 7.4 is released to the public, you who run older Nvidia GPUs may notice that your GPU is no longer being used by BOINC for Macintosh. This happens only when you updated your drivers to 340.xx or later, or otherwise added CUDA 6.5 support on your computer. Nvidia released CUDA 6.5 earlier this year and in its documentation it states that it will drop support for all compute capability GPUs 1.0 to 1.3. Despite NVIDIA's claims that CUDA 6.5 still supports GPUs with compute capabilities 1.0 through 1.3, we have found that it does not. Please see: http://setiathome.berkeley.edu/forum_thread.php?id=75633&postid=1571811#1571811 and https://devtalk.nvidia.com/default/topic/774899/cuda-programming-and-performance/340-52-driver-breaks-opencl-support-for-pre-fermi-cc-1-x-cards/ for more information. We have found that once CUDA 6.5 is installed on Macintosh computers, OpenCL applications (e.g. Seti Astropulse v7 and Collatz Conjecture) cause frequent kernel panics (i.e, they crash the OS!) on systems with older NVIDIA GPUs. In addition, the people working on SETI@home have found that OpenCL applications return incorrect results on Windows computers with older GPUs which have the NVIDIA 340.xx range of drivers (those which have CUDA 6.5), even though it appears to the user that it works properly. For the moment Seti will rely on the use of the validators to weed out these erroneous tasks from the correct ones. Notice we are talking about OpenCL applications, not just CUDA apps. The Macintosh installer for BOINC 7.4.26 includes a ReadMe, which now says: CUDA UPGRADE WARNING: Do not upgrade to CUDA 6.5 or later if you have an older NVIDIA GPU with Compute Capability 1.3 or less. You can check your GPU’s Compute Capability at https://developer.nvidia.com/cuda-gpus To use BOINC 7.4.26 and beyond with older NVIDIA GPUs on Macintosh computers, you must install an older version of the drivers. You can find older CUDA drivers at http://www.nvidia.com/object/mac-driver-archive.html. Note: after mounting the downloaded disk image, you may need to control-click on the CUDA installer package to open it. Below is an example of how BOINC will report your GPU: NVIDIA GPU 0: Quadro FX 3800M cannot be used for CUDA or OpenCL computation with CUDA driver 6.5 or later CUDA: NVIDIA GPU 0 (not used): Quadro FX 3800M (driver version 340.84, CUDA version 6.5, compute capability 1.1, 1024MB, 862MB available, 634 GFLOPS peak) OpenCL: NVIDIA GPU 0 (not used): Quadro FX 3800M (driver version 340.84, device version OpenCL 1.0 CUDA, 1024MB, 862MB available, 634 GFLOPS peak) Note that there is no cc_config.xml option to force the GPU to be used anyway, as it doesn't make sense to allow a GPU that returns only erroneous work to be used. |
Send message Joined: 29 Aug 05 Posts: 15566 |
Charlie Fenton, BOINC developer for Macintosh wrote: While one can debate the issue for Windows where the "only" issue is silently returning incorrect results, we have demonstrated that running OpenCL apps on Macs with older NVIDIA GPUs and CUDA 6.5 installed causes _frequent_ kernel panics, which are the Mac equivalent of Windows blue screens of death. |
Send message Joined: 12 Feb 11 Posts: 419 |
BOINC 7.4 removes support for CUDA GPUs that run CUDA 6.5 drivers A curiosity, which version of OpenCL is supported by Boinc? 1.1? 1.2? 2.0?? |
Send message Joined: 29 Aug 05 Posts: 15566 |
BOINC itself doesn't need to support CUDA or OpenCL, that's something the science applications need to have. BOINC will be able to detect a GPU that can do any version of OpenCL. |
Send message Joined: 29 Aug 05 Posts: 15566 |
Rom Walton, BOINC developer Linux and Windows wrote: We will be backing out the change that blocked the using older Nvidia cards with newer drivers on Windows. |
Send message Joined: 29 Aug 05 Posts: 15566 |
Rom Walton, BOINC developer for Windows and Linux wrote: Howdy Folks, |
Send message Joined: 12 Jul 14 Posts: 6 |
Probably a bug in 7.4.27: At present I'm running skynetPOGS and WU-Prop with resource share 100 and WCG as backup project with resource share 0 As POGS doesn't send new tasks at present, WCG should provide my CPU's with work. No problems with 7.2.42, BOINC client asked WCG properly for work each time a task was finished. When updating to 7.4.27, BOINC client messaged "no tasks needed" (WCG) and only WUprop was running on one core. Reverting back to 7.2.42 all cores were immediately provided with sufficient tasks from the backup project. |
Send message Joined: 29 Aug 05 Posts: 15566 |
Can you run a couple of rounds with these debug flags and post their output? <work_fetch_debug>1</work_fetch_debug> <cpu_sched_debug>1</cpu_sched_debug> You can use the new Advanced->Event Log Diagnostic Flags window to easily set these without having to mess with Notepad or another editor. Edit: I am testing it and have sent a preliminary report to the developers. Could still use a couple of rounds of logs from you though. Edit2: Elektra did post some logs, but due to their length I removed them from this thread. |
Send message Joined: 29 Aug 05 Posts: 15566 |
Did WCG in your case already have the application(s) downloaded for the sub-project you want it to run at zero RS? |
Send message Joined: 29 Aug 05 Posts: 15566 |
At least a couple of problems I see through your logs: Sorry, a very long event log: These continue like this all the way through. The one thing I do see why your WCG won't request work is because you set it to No New Tasks through BOINC Manager. So that accounts for WCG. As for 7.0.64 you used the wrong URL for iGEM@Home: 12/11/2014 04:01:04 | iGEM@home | URL http://igemathome.org/; Computer ID 567; resource share 0 If this was added this way through BAM, you should inform them of the new URL this project has. And WCG is still: 12/11/2014 04:02:17 | World Community Grid | [work_fetch] REC 382.432 prio -1000.000000 can't req work: "no new tasks" requested via Manager Thanks for both logs, I will forward these to the developers. |
Send message Joined: 12 Jul 14 Posts: 6 |
The issue regarding a wrong URL for iGEM@home is known at BoincStats but without harm for the BOINC client resp. the scheduler. I've removed the <no new tasks> setting for WCG and I am currently running another test. I'll post the results and log files per eMail to the developer group to avoid an overrunning of this thread with endless debug logs. But the error remains the same: BOINC client 7.4.27 isn't able to replace finished tasks with tasks from running backup projects (resource share = 0) when the only main project (resource share = 100) is out of work, and in the end all CPU cores are without work from BOINC. |
Send message Joined: 29 Aug 05 Posts: 15566 |
David Anderson wrote: There was a bug involving zero-share projects. |
Send message Joined: 27 Nov 14 Posts: 2 |
David Anderson wrote:There was a bug involving zero-share projects. I moved to previous version due to the zero share bug. As I did that I noticed a huge difference 33%(But only for the NVIDIA card) in floating point measurement between these versions. Now the measurement is: 27-Nov-2014 18:57:44 [---] Starting BOINC client version 7.2.42 for windows_x86_64 27-Nov-2014 18:57:44 [---] log flags: file_xfer, sched_ops, task 27-Nov-2014 18:57:44 [---] Libraries: libcurl/7.25.0 OpenSSL/1.0.1 zlib/1.2.6 27-Nov-2014 18:57:44 [---] Data directory: C:\ProgramData\BOINC 27-Nov-2014 18:57:44 [---] Running under account Jóhannes 27-Nov-2014 18:57:44 [---] CUDA: NVIDIA GPU 0: GeForce GTX 980 (driver version 344.75, CUDA version 6.5, compute capability 5.2, 4096MB, 3525MB available, 8165 GFLOPS peak) 27-Nov-2014 18:57:44 [---] OpenCL: NVIDIA GPU 0: GeForce GTX 980 (driver version 344.75, device version OpenCL 1.1 CUDA, 4096MB, 3525MB available, 8165 GFLOPS peak) 27-Nov-2014 18:57:44 [---] OpenCL: Intel GPU 0: Intel(R) HD Graphics 4000 (driver version 10.18.10.3621, device version OpenCL 1.2, 632MB, 632MB available, 154 GFLOPS peak) 27-Nov-2014 18:57:44 [---] OpenCL CPU: Intel(R) Core(TM) i7-3770K CPU @ 3.50GHz (OpenCL driver vendor: Intel(R) Corporation, driver version 3.0.1.10878, device version OpenCL 1.2 (Build 76413)) Was 25-Nov-2014 17:53:31 [---] Starting BOINC client version 7.4.27 for windows_x86_64 25-Nov-2014 17:53:31 [---] log flags: file_xfer, sched_ops, task 25-Nov-2014 17:53:31 [---] Libraries: libcurl/7.33.0 OpenSSL/1.0.1h zlib/1.2.8 25-Nov-2014 17:53:31 [---] Data directory: C:\ProgramData\BOINC 25-Nov-2014 17:53:31 [---] Running under account Jóhannes 25-Nov-2014 17:53:31 [---] CUDA: NVIDIA GPU 0: GeForce GTX 980 (driver version 344.75, CUDA version 6.5, compute capability 5.2, 4096MB, 3816MB available, 5444 GFLOPS peak) 25-Nov-2014 17:53:31 [---] OpenCL: NVIDIA GPU 0: GeForce GTX 980 (driver version 344.75, device version OpenCL 1.1 CUDA, 4096MB, 3816MB available, 5444 GFLOPS peak) 25-Nov-2014 17:53:31 [---] OpenCL: Intel GPU 0: Intel(R) HD Graphics 4000 (driver version 10.18.10.3621, device version OpenCL 1.2, 632MB, 632MB available, 154 GFLOPS peak) 25-Nov-2014 17:53:31 [---] OpenCL CPU: Intel(R) Core(TM) i7-3770K CPU @ 3.50GHz (OpenCL driver vendor: Intel(R) Corporation, driver version 3.0.1.10878, device version OpenCL 1.2 (Build 76413)) There is also a difference in available memory Accessible memory is considered more in the newer version but is the card considered a lesser performer by approx. 1/3rd. Should these figures not be constant between versions? Johannes |
Send message Joined: 20 Nov 12 Posts: 801 |
There is also a difference in available memory Nvidia changed the GPU configuration in 900 series/compute capability 5(???) cards. The formula used to calculate the peak flops is incorrect in old BOINC clients for the new cards. No idea why the available memory is different. |
Send message Joined: 5 Oct 06 Posts: 5129 |
There is also a difference in available memory Specifically, when v7.2.42 was developed, the newest NVidia architecture for which specifications were known was the Kepler generation, with 48 CUDA cores per multiplexer. No details for the next generation were released in advance, but when the Maxwell generation (9xx series, compute capability 5) arrived, it had 32 cores per multiplexer. That is properly accounted for in the v7.4.27 client - hence the change in calculated 'Peak GFLOPS'. I suspect that the v7.4 client range has better handling of the 4GB (32-bit) VRAM limit, but I don't know the details. |
Copyright © 2024 University of California.
Permission is granted to copy, distribute and/or modify this document
under the terms of the GNU Free Documentation License,
Version 1.2 or any later version published by the Free Software Foundation.