Posts by Oliver Behnke

WARNING: This website is obsolete! Please follow this link to get to the new Albert@Home website!

21) Message boards : Problems and Bug Reports : [New release] BRP app v1.22 feedback thread (Message 111973) Posted 27 Apr 2012 by Oliver Behnke Post: I did only now realise that my card is not supported because you demand a min workgroup size of 256. We don't, we just set a preferred value. If your GPU doesn't support it, the value is dynamically adjusted accordingly. Cheers, Oliver
22) Message boards : Problems and Bug Reports : [New release] BRP app v1.22 feedback thread (Message 111915) Posted 9 Mar 2012 by Oliver Behnke Post: Now that's strange. Looks like BOINC's borked runtime estimation again. Thanks for reporting... Oliver
23) Message boards : Problems and Bug Reports : [New release] BRP app v1.22 feedback thread (Message 111913) Posted 8 Mar 2012 by Oliver Behnke Post: ... and it looks like its faulty ???? Well, let's say it's non-optimal, in particular for GPU apps. The runtime estimates are determined for every application version independently. Thus after each newly released version BOINC needs some time to gather statistics to come up with a valid/reasonable runtime estimate. Don't worry, we won't be using this new system over on einstein until it proves reliable, but we need to test it here in order to improve (fix) it at all - as soon as time permits. Best, Oliver
24) Message boards : Problems and Bug Reports : [New release] BRP app v1.22 feedback thread (Message 111909) Posted 6 Mar 2012 by Oliver Behnke Post: It's not. The reason isn't a flaw in our runtime estimation (stored in the work unit definition) but BOINC's new automatic runtime estimation system (a.k.a. new credit system) we're also testing here on albert... Oliver
25) Message boards : Problems and Bug Reports : [New release] BRP app v1.22 feedback thread (Message 111906) Posted 5 Mar 2012 by Oliver Behnke Post: Since v1.21/v1.22 update, I have got a lot of validation errors. Please check my stats/failed units, maybe that will help you to find a bug. My system is BOINC 7.18 + Mac OS X 10.7.3 + AMD HD6750M 1GB Same as for "[AF>Le_Pommier] McRoger": no working OS X OpenCL app for the time being... Oliver
26) Message boards : Problems and Bug Reports : [New release] BRP app v1.22 feedback thread (Message 111905) Posted 5 Mar 2012 by Oliver Behnke Post: Same for me, my results Boinc Manager 7.0.12 OS X Lion 10.7.3 ATI Radeon 5770 (Apple) 1 Gb Besides there is this error in the log [19:14:19][1216][INFO ] Checkpoint file unavailable: status.cpt (No such file or directory). Might it explain why calculation time is so huge compared to other platforms ? 1) Please read my intro post of this thread. The Mac version is known to produce invalid results. We already disabled it. 2) Your GPU is simply not that efficient that's why it takes so long. It's not about the platform but the GPU. 3) The message you quote is an "INFO" message, so no, it's not the reason. It's normal when a fresh dataset is being analyzed for the first time - there can't be any checkpoint then. HTH, Oliver
27) Message boards : Problems and Bug Reports : [New release] BRP app v1.22 feedback thread (Message 111904) Posted 5 Mar 2012 by Oliver Behnke Post: Hi Jord, So, pretty please, can the fpops estimate be adjusted enough that they don't come in thinking to take 200+ hours? I'll forward this to Bernd but he's pretty overwhelmed with more important topics right now and the BOINC devs are of little help analyzing this right now. Please bear with us. Cheers, Oliver
28) Message boards : Problems and Bug Reports : [New release] BRP app v1.22 feedback thread (Message 111895) Posted 2 Mar 2012 by Oliver Behnke Post: Could running vbox be a problem? Don't know but could be if vbox acquires the GPU somehow... Oliver
29) Message boards : Problems and Bug Reports : [New release] BRP app v1.22 feedback thread (Message 111893) Posted 1 Mar 2012 by Oliver Behnke Post: The scheduler requests new jobs, and 3 seconds later it starts S6LV1. Well, S6LV1 tasks are not the same as BRP tasks. S6LV1 tasks re-use data already present on your host while BRP data is only used once, for a single WU. Looking at the error output of the failed S6LV1 task should tell us what happened. Please open another thread for that problem if it persists. This thread is meant to discuss BRP v1.22 only. Cheers, Oliver
30) Message boards : Problems and Bug Reports : [New release] BRP app v1.22 feedback thread (Message 111890) Posted 29 Feb 2012 by Oliver Behnke Post: No X windows here, I'm running win7 Oops, sorry :-) When you start the BOINC client it'll list all GPUs in the event log (advanced view). For AMD/ATI devices it might talk about CAL and OpenCL - we're interested only in the latter. You should find the list of GPUs more or less at the top of the event log, right before the registered projects are mentioned. Oliver
31) Message boards : Problems and Bug Reports : [New release] BRP app v1.22 feedback thread (Message 111888) Posted 29 Feb 2012 by Oliver Behnke Post: Are there actually tasks available? I'm not receiving anything for my GPU. Yes there are and your config looks fine so far. Please have a look at the BOINC event log: did BOINC recognize your AMD GPU as OpenCL device? According to our logs it doesn't seem to be the case. You might need to reinstall the Catalyst driver. Also, remember to start the X server and make sure that BOINC can access the X display. If "clinfo" exists on your system you may use it to verify that your GPU is properly enumerated by OpenCL. Cheers, Oliver
32) Message boards : Problems and Bug Reports : [New release] BRP app v1.22 feedback thread (Message 111882) Posted 28 Feb 2012 by Oliver Behnke Post: should we dump Wu's still using the 1.21 app? I've noticed a few inconclusives and errors on that app. not seeing any problems from other GPU projects Depends on what errors you saw. If they are memory-related you probably want to reset your project. You'll be resent the same tasks but they'll be crunched with the latest app version which required less memory (OpenCL only). Cheers, Oliver
33) Message boards : Problems and Bug Reports : [New release] BRP app v1.22 feedback thread (Message 111881) Posted 28 Feb 2012 by Oliver Behnke Post: Just to be sure: By "latest Catalyst driver (>=12.1)" you mean "AMD Catalystâ„¢ 11.11 - Revision number 12.1" ? I can't find any Catalyst 12.1 for Linux on the AMD site. That's just because of AMD's sloppy web editing. The one you found is the 12.1 driver. Oliver
34) Message boards : News : New BRP application released: v1.22 (CPU, CUDA, OpenCL) (Message 111877) Posted 27 Feb 2012 by Oliver Behnke Post: Discussion of the new app release takes place over in this thread. Notes: * OpenCL GPU memory requirements reduced by 128 MB (to ~360 MB) * Dropped support for OpenCL 1.0 GPUs (in favor of the above, OpenCL 1.0 only GPUs like the Radeon 4xxx are too slow anyway) * More graceful OpenCL memory error handling * Known issue: no OpenCL support for Mac OS X for the time being (we're looking into a potential Apple bug, no pun intended) * CUDA apps are enabled again for the time being * Input data download volume (per work unit) can be reduced by a factor of two (soon) * Please use the latest Catalyst driver (>=12.1) and BOINC client (>=7.0.12). Cheers, Oliver
35) Message boards : Problems and Bug Reports : [New release] BRP app v1.22 feedback thread (Message 111876) Posted 27 Feb 2012 by Oliver Behnke Post: Hi, We just released BRP4 v1.22 which adds a number of improvements. Notes: * OpenCL GPU memory requirements reduced by 128 MB (to ~360 MB) * Dropped support for OpenCL 1.0 GPUs (in favor of the above, OpenCL 1.0 only GPUs like the Radeon 4xxx are too slow anyway) * More graceful OpenCL memory error handling * Known issue: no OpenCL support for Mac OS X for the time being (we're looking into a potential Apple bug, no pun intended) * CUDA apps are enabled again for the time being * Input data download volume (per work unit) can be reduced by a factor of two (soon) * Please use the latest Catalyst driver (>=12.1) and BOINC client (>=7.0.12). Let's try and collect your feedback to this specific release (and this one only) in this thread. Thanks, Oliver
36) Message boards : Problems and Bug Reports : [OpenCL] app v1.20/v1.21 feedback thread (Message 111872) Posted 27 Feb 2012 by Oliver Behnke Post: Looks like a BOINC issue. Thanks for reporting. Oliver
37) Message boards : Problems and Bug Reports : [OpenCL] app v1.20/v1.21 feedback thread (Message 111863) Posted 20 Feb 2012 by Oliver Behnke Post: As it says, not enough memory... Update: we'll improve the error handling and messaging for such cases. The next release will state it more explicitly and the app will ask BOINC to retry again after 5 minutes, as it does for other GPU memory issues... Oliver
38) Message boards : Problems and Bug Reports : [OpenCL] app v1.20/v1.21 feedback thread (Message 111860) Posted 17 Feb 2012 by Oliver Behnke Post: I had no success in a non X setup. (I do not need X here) Stopping lighdm so that all modules are loaded and DISPLAY (COMPUTE) variable set to ":0" atiOpenCL apps do not start on the commandline in a console. Known limitation of the AMD drivers/runtime. If your machine is isolated you may also just allow all users to connect to your (running) X display, works just fine for me: export DISPLAY=:0 && xhost + Best, Oliver
39) Message boards : Problems and Bug Reports : [OpenCL] app v1.20/v1.21 feedback thread (Message 111850) Posted 10 Feb 2012 by Oliver Behnke Post: It turned out on debian64, the AMD driver is installed on /usr/lib32/libamdocl32.so and not /usr/lib/libamdocl32.so. I don't know why but I had an old version in /usr/lib. Just copied the new lib over, and it's now working. I experienced the same issue on my machine. I turns out that this is an inconsistency between the APP SDK and Catalyst. Both install this 32-bit compatibility library, but in different locations, hence the mix-up that leads to the segfault. I already filed a bug report with AMD. I recommend to not install the APP SDK, you don't need it to run OpenCL GPU applications as long as you installed Catalyst >= 11.3 (Windows, 11.11 for Linux). The SDK is for developers only. Cheers, Oliver
40) Message boards : Problems and Bug Reports : [OpenCL] app v1.20/v1.21 feedback thread (Message 111839) Posted 8 Feb 2012 by Oliver Behnke Post: Oliver i installed ubuntu supplied ati catalyst it is 12.1 but its missing opencl Are you sure you installed all relevant packages? Linux distributions might often split the original/upstream driver installation package into separate packages like the kernel module, the user space driver, the OpenCL runtime and the OpenCL ICD. Debian does this and Ubuntu could do it in a similar way since it's based on Debian. Good to see that it now works for you! Cheers, Oliver

Previous 20 · Next 20

This material is based upon work supported by the National Science Foundation (NSF) under Grant PHY-0555655 and by the Max Planck Gesellschaft (MPG). Any opinions, findings, and conclusions or recommendations expressed in this material are those of the investigators and do not necessarily reflect the views of the NSF or the MPG.

Copyright © 2024 Bruce Allen for the LIGO Scientific Collaboration