Welcome to MilkyWay@home

Posts by Matt Arsenault

41) Message boards : Number crunching : All work Units giving "Computational Error" (Message 53208)
Posted 16 Feb 2012 by Matt Arsenault
Post:
I updated to 12.1 and got 2 valid WU's turned in today-otherwise lots of failures.

http://milkyway.cs.rpi.edu/milkyway/show_host_detail.php?hostid=233052
Apparently the unroll problem wasn't fixed until 1 driver later than I thought. You shouldn't get it anymore for anything newer than R700 with drivers older than 11.6
42) Message boards : Number crunching : All work Units giving "Computational Error" (Message 53206)
Posted 16 Feb 2012 by Matt Arsenault
Post:
CCC: 10.12
"Message from server: Catalyst Driver Version is Not OK for OpenCL app with this GPU"
10.12 is ancient; you'll need something more recent. The most recent 12.1 should be fine.
43) Message boards : Number crunching : All work Units giving "Computational Error" (Message 53204)
Posted 16 Feb 2012 by Matt Arsenault
Post:
Care to elaborate?

This error does not occur if I run other GPU projects. Put another way .
The drivers aren't particularly unreliable. They sometimes get into a bad state where nothing works.
44) Message boards : Number crunching : All work Units giving "Computational Error" (Message 53188)
Posted 15 Feb 2012 by Matt Arsenault
Post:
Nope. Still doesn't work. Still get the same computational error. I'm done until this BS issue is fixed.
That wasn't for your error. Your error suggests the OpenCL library is broken or missing.
45) Message boards : Number crunching : All work Units giving "Computational Error" (Message 53184)
Posted 15 Feb 2012 by Matt Arsenault
Post:
After approx 01.00 UTC 15/2/2012 all 0.82 (ati14) GPU tasks deadline dated 27/2/2012 02:03++ failed with computation error.
All remaining unstarted GPU tasks have been aborted.
The option for no new tasks is initiated.

FWIW all other GPU tasks completed succesfully today AFAIK.

Please advise.
Your errors suggests rebooting.
46) Message boards : News : Separation updated to 1.00 (Message 53046)
Posted 10 Feb 2012 by Matt Arsenault
Post:
It depends on the driver version and GPU. The most problematic GPU are Radeon 4xxx series. There are only 4 driver releases in the past year that I'm certain work (11.7, 11.8, 11.9 and 12.1+) so unless you're using one of those you won't get work for it.
47) Message boards : Number crunching : All work Units giving "Computational Error" (Message 53041)
Posted 10 Feb 2012 by Matt Arsenault
Post:
I believe it should work if you try 1.02 with your current drivers. My workaround for the Nvidia problem seems to have worked.
48) Message boards : Number crunching : Not getting work (Message 53038)
Posted 10 Feb 2012 by Matt Arsenault
Post:
It looks like my workaround did work on a variety of drivers so I dropped the limit back to what it was before.
49) Message boards : News : Separation updated to 1.00 (Message 53035)
Posted 10 Feb 2012 by Matt Arsenault
Post:
No, it is really high cpu-usage according to the windows task-manager which shows usage up to 25 % on my quad-core for each milkyway process. Boinc shows 0.05 as expected. Please show also in my example wu in the original post. The cpu-time is very high!
I know the AMD drivers had the high CPU problem in 11.7 and 11.8 but I guess that could be different on different hardware. Does it happen with the current 12.1 drivers for you?
50) Message boards : News : Separation updated to 1.00 (Message 53032)
Posted 10 Feb 2012 by Matt Arsenault
Post:
I found out where to put it, but what numbers should I use to lower CPU usage?
Raising the GPU wait factor (--gpu-wait-factor ) and/or changing the polling mode (--gpu-polling-mode ). The default wait factor is 0.75, to wait for 75% of the estimated kernel run times before waiting. Raising it towards 1 should lower CPU usage. If you want to play with that try raising the wait factor first.

Can I upgrade to BOINC 7 without wiping out all of mine existing/cached/running WU's?
I think I saw that it wipes it if the version changes, so no.
51) Message boards : Number crunching : Not getting work (Message 53029)
Posted 10 Feb 2012 by Matt Arsenault
Post:
I can't afford to burn a full CPU core just to keep the GPU busy so I'd like to revert to the 260.19.44 driver
The new version has a workaround for it. By default it isn't particularly aggressive about trying to reduce it, but for me it keeps around 50%.

Right, BOINC doesn't seem to recognize the driver:
NVIDIA GPU 0: GeForce GTX 260 (driver version unknown, CUDA version 4010, compute capability 1.3, 895MB, 560 GFLOPS peak)
I added the fallback version check against the CUDA version which is still reported.

Perhaps you could drop the version check?
A bug showed up in the older drivers. I'm not sure which ones but I knew that 280.xx+ was OK. I'll loosen the version restrictions as I know more about which versions are bad. 1.02 possibly avoids the issue but I'm not sure yet.
52) Message boards : News : Separation updated to 1.00 (Message 53026)
Posted 10 Feb 2012 by Matt Arsenault
Post:
Opposite to my first post above, I have encountered problems now. I have very high cpu-usage with 1.00 and 1.02 app and two 5850 GPU.
Do you actually have high CPU usage, or is BOINC just telling you it is using 0.98 CPUs? I see that one one of my systems (Using BOINC 7) and not on another (Using 6.12)
53) Message boards : News : Separation updated to 1.00 (Message 53025)
Posted 10 Feb 2012 by Matt Arsenault
Post:
I have been running Milkyway@home projects on my ATI 4850. Are these cards no longer supported?
They are
54) Message boards : Number crunching : GPU Requirements (Message 53011)
Posted 10 Feb 2012 by Matt Arsenault
Post:
With the changes implemented to the new wu's I have now started to receive the following message from mw@home: "Catalyst driver version is not okay for OpenCL application with this GPU."
For an R700 GPU like you say you have, it should accept it if your Catalyst version is between 11.7 and 11.9 or >= 12.1. The OpenCL version isn't important; only OpenCL 1.0 features are used. Which version were you trying?
55) Message boards : News : Separation updated to 1.00 (Message 53010)
Posted 10 Feb 2012 by Matt Arsenault
Post:
I was on 7.0.12, upgraded to 7.0.14, still the same problem.
Maybe we need to open a ticket at boinc/trac ?
I'm not sure. I've been mutilating the scheduler trying to quickly dodge driver version issues so it's quite likely not working, though it does seem to be working for me using 7.0.14


I'm on 12.1 since they went out.
I just upgraded to 7.0.15 but still got the same message that I lack the necessary extension. I wrote something here, but I don't know if it will be read by boinc devs http://boinc.berkeley.edu/dev/forum_thread.php?id=7273
It was scheduler problem. I don't have any pure cl_amd_fp64 systems so it was working for me. It should start working for you now.
56) Message boards : News : Separation updated to 1.00 (Message 53008)
Posted 10 Feb 2012 by Matt Arsenault
Post:
I was on 7.0.12, upgraded to 7.0.14, still the same problem.
Maybe we need to open a ticket at boinc/trac ?
I'm not sure. I've been mutilating the scheduler trying to quickly dodge driver version issues so it's quite likely not working, though it does seem to be working for me using 7.0.14
57) Message boards : News : Separation updated to 1.00 (Message 53007)
Posted 10 Feb 2012 by Matt Arsenault
Post:
I updated the GPU stuff to 1.02. This should fix the multiple vendor selection for some people, and hopefully works around the common error on older Nvidia drivers with the nonsense "call to 'mad' is ambiguous" error.

The opencl_amd is now known as opencl_amd_ati because I'm pretty sure the lack of the substring 'ati' in it was the cause of the no-ATI work problem yesterday (I hate the plan class system so, so much).

The scheduler should be enforcing more correct driver limit ranges now also.
g.g. for Radeon 4xxx: more broken than I thought. It only works (and you should only get it) if your catalyst version is between 11.7 and 11.9 or >= 12.1
58) Message boards : News : Separation updated to 1.00 (Message 53004)
Posted 10 Feb 2012 by Matt Arsenault
Post:
Started to get the several errors today on 3 different systems with GPUs.
Suggestions?

-------------------------------------
http://milkyway.cs.rpi.edu/milkyway/result.php?resultid=99668860

Coprocessors AMD ATI Radeon HD 5800 series (Cypress) (2048MB) driver: Operating System Microsoft Windows XP

Using SSE3 path
Error getting number of platform (-1001): CL_PLATFORM_NOT_FOUND_KHR
I've seen this problem mentioned before with a few different causes. The Catalyst installer/uninstaller at some points didn't clean up the ICD entries in the registry properly. Try removing the drivers and installing newer ones.
59) Message boards : News : Separation updated to 1.00 (Message 53003)
Posted 10 Feb 2012 by Matt Arsenault
Post:
I'm using Nvidia but CPU usage is still quite high.

Where do I put --gpu-wait-factor and --gpu-polling-mode, and how to lower CPU usage even more?
In the <cmdline></cmdline>

edit: And do I need to change the <type> tag?
No
60) Message boards : Number crunching : Not getting WU's (Message 52989)
Posted 10 Feb 2012 by Matt Arsenault
Post:
I think I fixed it.

Maximum CPU % for graphics
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.

Both of those were set to a "non number" or something. Setting them back to a real number got me 24 tasks!
Those aren't related


Previous 20 · Next 20

©2024 Astroinformatics Group