Posts by Oliver Behnke

WARNING: This website is obsolete!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

41) Message boards : Problems and Bug Reports : [OpenCL] app v1.20/v1.21 feedback thread (Message 111832) Posted 7 Feb 2012 by Oliver Behnke Post: First successfully completed opencl wu http://albert.phys.uwm.edu/result.php?resultid=109919 Great, so it turned out that the installation was flawed, most likely because of missing 32-bit compatibility libs. Anyway, I hope we see at least some of those errors again on other systems as we'd really like to understand them and make BOINC more robust for these types of situations. *@HV Margus:** Please run at least one or two tasks before you fix your driver setup. This would really help us! Thank you! Best, Oliver
42) Message boards : Problems and Bug Reports : [OpenCL] app v1.20/v1.21 feedback thread (Message 111831) Posted 7 Feb 2012 by Oliver Behnke Post: Maybe we can get some feedback from those successfully returning results from 64bit linux machines with ATI/AMD gpus. Guys, we just released v1.21 which adds some useful debugging info to the task/result error log. Please reset the albert project or wait until the next task is fetched. The log output of those tasks will help us to understand the problem. Thanks, Oliver
43) Message boards : Problems and Bug Reports : [OpenCL] app v1.20/v1.21 feedback thread (Message 111821) Posted 6 Feb 2012 by Oliver Behnke Post: For those of you on Linux who have the AMD APP SDK installed alongside the AMD Catalyst drivers, please make sure you follow the order as details in the AMD developer release notes (chapter 2). HTH, Oliver
44) Message boards : Problems and Bug Reports : [OpenCL] app v1.20/v1.21 feedback thread (Message 111820) Posted 6 Feb 2012 by Oliver Behnke Post: I have no idea whether poems app is a fully boinc compliant app, but they certainly use the boinc opencl detection features Ok, fair enough. As to the number of gpus each of my 3 boxes only has a single gpu. Ok, thanks. May I suggest that you release a 64bit opencl app, if that works (at least detects the opencl features) then the problem lies in the detection of 32bit capabilities on a 64bit machine. Not anytime soon due to our current time constraints. However, there are successful results from 64-bit Linux systems, so it's not a general issue. Oliver
45) Message boards : Problems and Bug Reports : [OpenCL] app v1.20/v1.21 feedback thread (Message 111817) Posted 6 Feb 2012 by Oliver Behnke Post: 3 gentoo 64bit boxes all with the same problem and they are crunching openCL wus for poem and primegrid - exit with code 255 - unable to get openCL info - error -1 FYI, those projects don't use BOINC's OpenCL features at all. As far as I know we are the first project trying to get a fully BOINC-compliant OpenCL implementation going, including OpenCL device management controlled by BOINC, not the app. The error you experience is related to that and we're already aware of it and are investigating it as we speak. HV Margus* and Trog Dog: do you have multiple GPUs in your systems or just one? Cheers, Oliver
46) Message boards : Problems and Bug Reports : [OpenCL] app v1.20/v1.21 feedback thread (Message 111804) Posted 3 Feb 2012 by Oliver Behnke Post: According to apple specs it's a MacPro with the standard graphics card ATI Radeon 5770 1GB video mem (one auxilary power cable). Maybe AMD-APP-SDK limits to 50%? Hang on, do you run Linux on your MacPro? Anyway, you're witnessing the mess we've to deal with here, AMD's driver and CAL/OpenCL runtimes don't reliably report RAM sizes on some GPUs. They don't even agree with each other. I wouldn't wonder if the driver/runtime tells our app that there's no more memory available (hence the error you see) even when there is... This is valuable hint indeed! Oliver
47) Message boards : Problems and Bug Reports : [OpenCL] app v1.20/v1.21 feedback thread (Message 111802) Posted 3 Feb 2012 by Oliver Behnke Post: Oops you're right, video memory is reported as 1024MB for both cards. With 817MB, 983MB available, but global memory for OpenCL is reported from AMD-APP-SDK as 512MB , but (strange) boinc says 817MB, 983MB available). Is there a tweak in OpenCL configuration? amdccle (Catalyst Control Center) says 1024MB Video Memory for both cards. Hmm...? This is a (sort of) known issue with AMD GPUs/drivers. Do you happen to know the actual/correct value of your card? Look at the retail box if it helps :-) Oliver
48) Message boards : Problems and Bug Reports : [OpenCL] app v1.20/v1.21 feedback thread (Message 111800) Posted 3 Feb 2012 by Oliver Behnke Post: Great, initial estimated time to completion, 286 hours. No wonder it went into HP. But that DCF of 11 is way out of whack. i3-530, Win 7 - 64bit, 8GB RAM, ATI HD6850 2GB, Catalysts 11.12, BOINC 7.0.12 I'll forward this to Bernd. Thanks, Oliver
49) Message boards : Problems and Bug Reports : [OpenCL] app v1.20/v1.21 feedback thread (Message 111799) Posted 3 Feb 2012 by Oliver Behnke Post: Last combination is Catalyst 12.1, AMD-APP-SDK-v2.6, (without OpenCL v1.2 support) boinc 7.0.12. Results are like=> http://albert.phys.uwm.edu/result.php?resultid=114016 clinfo reports:clinfo Global memory size: 536870912 Your card has 512 MB video memory. That's right on the edge of what our app requires (see above). In your case it's not sufficient - error 5 means "out of resources" (memory). Sorry. You would be one who would benefit if we drop OpenCL 1.0 support in favor of lower memory requirements... Cheers, Oliver PS: BOINC's memory information still seems to be off. I'll forward this...
50) Message boards : Problems and Bug Reports : [OpenCL] app v1.20/v1.21 feedback thread (Message 111795) Posted 2 Feb 2012 by Oliver Behnke Post: As I said, we can reduce it but only by dropping OpenCL 1.0 (i.e. AMD 4xxx) support. Oliver
51) Message boards : Problems and Bug Reports : [OpenCL] app v1.20/v1.21 feedback thread (Message 111793) Posted 2 Feb 2012 by Oliver Behnke Post: I run on an ATI 4850 with 512MB RAM and Catalyst 11.12 As discussed earlier, we (still) need ~490 MB free GPU memory. This doesn't seem to be the case here. We'd be able to reduce that by 120 MB but that would mean to exclude all OpenCL 1.0 GPUs, including your 4xxx series card. FYI, we think about doing this anyway since the OpenCL 1.0 devices don't offer a significant performance gain over the CPU version so it seems worthwhile dropping their support in favor of more relaxed memory requirements. Cheers, Oliver
52) Message boards : News : New OpenCL application released: v1.20 (Message 111791) Posted 2 Feb 2012 by Oliver Behnke Post: Discussion of the new app version takes place over in this thread. Best, Oliver
53) Message boards : Problems and Bug Reports : [OpenCL] app v1.20/v1.21 feedback thread (Message 111790) Posted 2 Feb 2012 by Oliver Behnke Post: Hi, We just released v1.20 which should fix an important bug which prevented proper validation, in particular on AMD 4xxx GPUs. Let's try and collect your feedback to this specific app version (and this one only) in this thread. FYI, please use the latest Catalyst driver (>=12.1) and BOINC client (>=7.0.12). Thanks, Oliver
54) Message boards : Problems and Bug Reports : Running on ATI (Message 111788) Posted 1 Feb 2012 by Oliver Behnke Post: pls check your PM Are you trying to keep it a secret? :-). Please post in the thread for everyone's benefit. I'm sure many others would be interested too. No, not as a secret. I offered him full access to my system via teamviewer. Since I'm not a BOINC or Project developer I'm not shure that I can supply all infos that are intresting to know. And I'm also shure you or Oliver can post the infos much better than I could. Thanks Alex. Jord provided all I wanted to know about your box already :-) Cheers, Oliver
55) Message boards : Problems and Bug Reports : Running on ATI (Message 111780) Posted 1 Feb 2012 by Oliver Behnke Post: This is the on-chip GPU of my FM1 Cpu. I got the whole system (Cpu, Mainboard and Ram) for ~250.- Euro. Now that's interesting. Can you post a link to that box? I'd like to take a look at the specs! Is this using an AMD APU or similar? BTW, the first two completed and validated, thats much better than the CPU wu's ... Good news :-) Oliver PS: we identified and fixed an issue that was causing a number of validation issues. v1.20 will be released short and should improve things.
56) Message boards : News : Sending work (Message 111779) Posted 1 Feb 2012 by Oliver Behnke Post: Ok, bug fix implemented and tested. We'll release v1.20 shortly... Oliver
57) Message boards : Problems and Bug Reports : Running on ATI (Message 111777) Posted 31 Jan 2012 by Oliver Behnke Post: Hi Alex, Hi, yesterday I installed my new system with two ATI-GPU's, latest BM 7.0.11 , latest AMD drivers and an outdated nVidia GPU (9500). I've seen that the GPU usage is very very low, so it makes sense to run more than one wu on every GPU. Is there an app_info available for that purpose? We don't provide an app_info.xml but I'm sure people will help you prepare one yourself. Also, BOINC 7.0.12 is the latest and the GPU usage is usually between 60-80%. However, this is only the case when there is enough CPU power left to feed the GPU (see previous postings). Best, Oliver
58) Message boards : News : Sending work (Message 111776) Posted 31 Jan 2012 by Oliver Behnke Post: They will. The 4xxx series doesn't support local memory, it's emulated via global memory which incurs a big impact on performance. Also, this series only allows for 64 work items per work group when local memory is used, hence the resizing. However, I doubt that the resizing actually affects the accuracy of the computation, but if it does, it needs to be fixed! Well, it turned out it does indeed! We'll fix it ASAP. Oliver
59) Message boards : Problems and Bug Reports : Running on ATI (Message 111764) Posted 27 Jan 2012 by Oliver Behnke Post: Is there a chance, that CPU-load for GPU-WUs could be lowered in the future? Yes, we have some ideas for that in the pipeline, it's just a matter of time to implement and test them. However, keep in mind that not all algorithms can be ported 100% to the GPU. This really depends on the problem and our current implementation is already not that bad after all. Right now we still need to focus on getting a first stable and numerically correct release out of the door -- and apart from our own code we largely depend on BOINC's readiness WRT OpenCL and proper AMD drivers, both are not easy to come by. Best Oliver
60) Message boards : Problems and Bug Reports : Running on ATI (Message 111763) Posted 27 Jan 2012 by Oliver Behnke Post: Is there any chance that you can lower again to 450MB? If you ask whether our app's memory requirements might be reduced to 450 MB at some point, then probably since we always strive to support as many cards as possible by keeping our requirements as low as possible. Right now we can't due to other technical constraints. 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