Welcome to MilkyWay@home

Posts by Amauri

1) Message boards : Number crunching : Is there a change to no longer use Nvidia GT520 (Message 57019)
Posted 25 Jan 2013 by Amauri
Post:
Lets try updating your BOINC version as that can also cause the problem.
http://boinc.berkeley.edu/dl/boinc_7.0.45_windows_intelx86.exe

Yes, BOINC 6.x doesn't detect OpenCL capabilities, you need BOINC 7.x to run OpenCL apps.
2) Message boards : Number crunching : Linux GPU: a huge memory leak?? (Message 57004)
Posted 23 Jan 2013 by Amauri
Post:
One more info: if I change the "Frequency (in Hz) that should try to complete individual work chunks" value in MW preferences from 60 (default) to 13 (best value for my config), this "memory leak" reduces from 10MB/1% progress to about 2.5MB/1% progress, and starts later. My guess is that it depends on GPU speed / utilization / optimization.

I'm running MW, Einstein, Seti, Poem and WCG in this GPU, and this problem only occurs here in MW.
3) Message boards : Number crunching : Linux GPU: a huge memory leak?? (Message 56997)
Posted 23 Jan 2013 by Amauri
Post:
I'm running the separation GPU app in my Linux box (Ubuntu 12.04 64bit, 4 GB RAM, NVidia GT 640, BOINC 7.0.44), and the used memory is increasing at a rate of 10 MB each 1% of progress. At the end of the task, the used memory is about 1 GB. Is this a normal behaviour? Can anyone confirm this?
4) Message boards : News : Nobdy Release 1.02 (Message 56244)
Posted 20 Nov 2012 by Amauri
Post:
Four tasks successfully completed (Linux & NVidia), but can't validate...
5) Message boards : Number crunching : Problem with GT 640 (Message 55460)
Posted 3 Sep 2012 by Amauri
Post:
Yes, the GT 640 can run OpenCL tasks (mine runs a WU in about 22 minutes), but you need Boinc 7.0+ to do this. AFAIK Boinc 6.12.34 doesn't work with OpenCL.
6) Message boards : Number crunching : GPU issue (Message 55271)
Posted 31 Jul 2012 by Amauri
Post:
Jon, you can change the "Frequency" under "MW@H Preferences". In my computer (GPU isn't used for graphics, it's exclusive to Boinc), the best value is "5" instead of the default "60", resulting in lower time to crunch and less than 1% of CPU usage.

Frequency (in Hz) that should try to complete individual work chunks. Higher numbers may run slower but will provide a more responsive system. Lower may be faster but more laggy.
default 60 (corresponds to 60 fps) --> 5

7) Message boards : Number crunching : App version needs OpenCL but GPU doesn't support it (Message 55263)
Posted 30 Jul 2012 by Amauri
Post:
The recent NVidia drivers don't report the OpenCL capabilities to Boinc, like these lines in Boinc's "Event Log":

Dom 29 Jul 2012 17:24:08 BRT |  | NVIDIA GPU 0: GeForce GT 640 (driver version unknown, CUDA version 4.20, compute capability 3.0, 2048MB, 134215651MB available, 692 GFLOPS peak)
Dom 29 Jul 2012 17:24:08 BRT |  | OpenCL: NVIDIA GPU 0: GeForce GT 640 (driver version 295.49, device version OpenCL 1.1 CUDA, 2048MB, 134215651MB available)

Installed 302.17 (Linux), and that 2nd. line (OpenCL) isn't there. Without OpenCL support, the GPU app can't run. Reverted to 295.49, and everything is ok.

Can anyone confirm this?
8) Message boards : News : Separation updated to 1.00 (Message 52889)
Posted 9 Feb 2012 by Amauri
Post:
I'm using 6.12.34 and I'm getting work.

That is because you are using an app_info (like I am) and that does not bother with the OpenCL check.

I'm using 6.12.33 (Linux), no app_info file, and I'm getting work. The CPU usage was reduced, but I think I will need an app_info file to play with --gpu-polling-mode to reduce it even more.

In my "ultra-low-end" GT 520 I'm getting this message in results:
Estimated Nvidia GPU GFLOP/s: 104 SP GFLOP/s, 13 DP FLOP/s
Warning: Bizarrely low flops (13). Defaulting to 50

Bizarre ;-)

Maybe this 13 GFLOP/s info is correct and the default of 50 is too high, affecting the estimated time of polling period...

Thanks for this new version, Matt!
9) Message boards : Number crunching : MilkyWay CUDA takes up a whole core (Message 52706)
Posted 28 Jan 2012 by Amauri
Post:
Thank you!
10) Message boards : Number crunching : nbody-plum files don't run - MW stops running (Message 52682)
Posted 27 Jan 2012 by Amauri
Post:
it's the BOINC scheduler. As n-body uses all cores of your CPU, the scheduler needs to stop all other tasks to run them, so if the internal scheduler priorities of the other projects are higher, it will wait until the other tasks are completed to run n-body.

To make them run immediately, select all other projects in "Projects" tab and click in "Suspend" button. When all n-body are completed, resume the other projects.
11) Message boards : Number crunching : MilkyWay CUDA takes up a whole core (Message 52637)
Posted 25 Jan 2012 by Amauri
Post:
Any news yet on the fixed GPU (CUDA-Windows for me) applications?
I'm planning on updating everything in about a week (maybe 2).

Any news?
12) Message boards : Number crunching : No GPU tasks? (Message 52265)
Posted 6 Jan 2012 by Amauri
Post:
With the new server I updated the scheduler to correctly check for the OpenCL double extension rather than just checking the hardware; however only the most recent (and possibly unreleased) clients send back the supported OpenCL extension, so almost no Nvidia/OpenCL tasks are being sent out. I'll fix it soon.


I'm running BOINC 7.0.2 (latest version), and my GPU is double-precision capable, but the server reports: "GPU lacks necessary double precision extension"...

Sex 06 Jan 2012 15:40:47 BRST |  | Starting BOINC client version 7.0.2 for x86_64-pc-linux-gnu
Sex 06 Jan 2012 15:40:47 BRST |  | NVIDIA GPU 0: GeForce GT 520 (driver version unknown, CUDA version 4.0, compute capability 2.1, 1024MB, 1005MB available, 156 GFLOPS peak)
Sex 06 Jan 2012 15:40:47 BRST |  | OpenCL: NVIDIA GPU 0: GeForce GT 520 (driver version 280.13, device version OpenCL 1.1 CUDA, 1024MB)
Sex 06 Jan 2012 15:40:47 BRST |  | NVIDIA GPU is OpenCL-capable
...
Sex 06 Jan 2012 17:53:11 BRST | Milkyway@Home | Sending scheduler request: To fetch work.
Sex 06 Jan 2012 17:53:11 BRST | Milkyway@Home | Requesting new tasks for NVIDIA
Sex 06 Jan 2012 17:53:12 BRST | Milkyway@Home | No tasks sent
Sex 06 Jan 2012 17:53:12 BRST | Milkyway@Home | Message from server: GPU lacks necessary double precision extension


13) Message boards : Number crunching : MilkyWay CUDA takes up a whole core (Message 52016)
Posted 20 Dec 2011 by Amauri
Post:
I'm running the new version of cuda_opencl (0.90), but it's still using 100%+ of CPU. Ubuntu 64bit, NV 280.13. This bug wasn't fixed yet...
There wasn't nothing new in 0.90. I just rebuilt an old one so only an older glibc would be required.

Can we expect a new version in 2011? Or Jan. 2012?
14) Message boards : Number crunching : MilkyWay CUDA takes up a whole core (Message 51941)
Posted 12 Dec 2011 by Amauri
Post:
I'm running the new version of cuda_opencl (0.90), but it's still using 100%+ of CPU. Ubuntu 64bit, NV 280.13. This bug wasn't fixed yet...
15) Message boards : Number crunching : MilkyWay CUDA takes up a whole core (Message 51853)
Posted 7 Dec 2011 by Amauri
Post:
I'm pretty sure I can write a workaround for it.


And I'm pretty sure you can do it. But... are you doing it?
16) Message boards : Number crunching : MilkyWay CUDA takes up a whole core (Message 51686)
Posted 13 Nov 2011 by Amauri
Post:
I had the same problem with nvidia driver 280 and 285. CUDA takes at least one whole core (if the cpu idles CUDA even takes 2 cores).

There are no problems with driver 266.


+1

Is there anything that could be changed in MW app to fix this?
17) Message boards : Number crunching : Milkyway N-Body uses all BOINC time (Message 51685)
Posted 13 Nov 2011 by Amauri
Post:
I see this behavior too...
MW is supposed to only use 50% of the available processors.
Why is it not honoring the preferences ???
JR


As Arkain said: "Multi-threaded applications are designed to use all of your cores to do the work."

And N-Body is a multi-threaded application.
18) Message boards : Number crunching : Problem with receiving credits (Message 51453)
Posted 20 Oct 2011 by Amauri
Post:
Reset will not work if app_info.xml has root owner. Boinc doesn't have root privileges, so resetting the project doesn't delete app_info.xml.

Stop boinc-client, manually delete app_info.xml, restart boinc-client and reset project. It's all.
19) Message boards : Number crunching : Please give more time to process WUs (Message 26629)
Posted 28 Jun 2009 by Amauri
Post:
I was merely trying to suggest something to help out. It has nothing to do with meeting requirements. It seems Boinc is getting too much work. You could go to a 20% share but I'm sure you have thought of that.


It's not too much work, it's my 15% (17% really) share. I want to minimize my power consumption, as I leave the computer alone for several hours. With share at 17%, the power consumption increases almost nothing running BOINC (that's what my UPS says), and my CPU stays at 30 C (85 F). Really cool, in both senses.

I'm always seeing my jobs at 90% of completion when time is about to expire, since the WUs changed from 2 hs long to 6 hs long two weeks ago...
20) Message boards : Number crunching : Please give more time to process WUs (Message 26530)
Posted 26 Jun 2009 by Amauri
Post:
I'm running just 1 WU at a time. My computer (AMD Semprom 1.66 GHz) is powered on for 12 hs/day, and I donated 15% of my CPU time to run BOINC in background. This gives me 1.8 hours/day of CPU time dedicated to BOINC. In 3 days I can process 5.4 hours of a WU - and I also run SETI...

If my config doesn't meet the requirements to run Milkyway, I'll disconnect from this project. I don't want to earn credits, I just want my work to be useful to Milkyway and not being discarded all times.


Next 20

©2024 Astroinformatics Group