| Index | Recent Threads | Unanswered Threads | Who's Active | Guidelines | Search |
 |
World Community Grid Forums
Category: Beta Testing
Forum: Beta Test Support Forum
Thread: OpenPandemics GPU Beta Test - March 12 2021 [ Issues Thread ] |
| No member browsing this thread |
|
Thread Status: Active Total posts in this thread: 160
|
|
| Author |
|
|
Eugene Zenzen
Veteran Cruncher USA Joined: Mar 31, 2006 Post Count: 888 Status: Offline Project Badges: 
|
In regard to Grumpy's comment:
----------------------------------------Worth noting is, that if you run CPU tasks at the same time as you run a iGPU task, you will slow down both the iGPU task, and the CPU tasks considerably. This was something we discovered on SETI@home, were we were running tons of different OpenCL GPU apps, and WU's. You can get away with running 1 or 2 CPU tasks on a lets say 4c/8t CPU, but no more than that, or you will slow down all WU's considerably, including the one on the iGPU. Out of 60+ latest betas run, only 3 errored out (exceeded elapsed time limit). These were all on an AMD 3400G with integrated graphics. On the other hand, 3 others ran successfully on that machine - also using the same iGPU. All betas on RX550, GT1030, GTX1050, GTX1060 ran without problems. In previous beta iterations 710s all errored out and no longer received WU in this last round. Everything running Windows 10. Here is an example of an iGPU run beta that exceeded time limit: https://www.worldcommunitygrid.org/ms/device/...og.do?resultId=1559390749  |
||
|
|
Rickjb
Veteran Cruncher Australia Joined: Sep 17, 2006 Post Count: 666 Status: Offline Project Badges: 
|
"Worth noting is, that if you run CPU tasks at the same time as you run a iGPU task, you will slow down both the iGPU task, and the CPU tasks considerably. ..."
----------------------------------------A theoretical explanation: The iGPU shares the main system RAM, and probably bottlenecks access to it, slowing down both processors. Thus discrete GPUs are the way to go for serious output. Time will tell whether a cheap older CPU system with an expen$ive newish high-end GPU will give you best bang for your buck on this project. OTOH, if no discrete GPU is available, running the iGPU with very few CPU cores doing CPU work may give more output than running all CPU cores alone. The v1.27 betas still won't run on my HD4000 iGPUs (3770Ks). And my HD7870 and GTX 770 cards are both dead  [Edit 1 times, last edit by Rickjb at Mar 13, 2021 3:40:19 PM] |
||
|
|
Michael Goetz
Cruncher United States Joined: Dec 11, 2017 Post Count: 35 Status: Offline Project Badges: 
|
A theoretical explanation: The iGPU shares the main system RAM, and probably bottlenecks access to it, slowing down both processors. Thus discrete GPUs are the way to go for serious output. Time will tell whether a cheap older CPU system with an expen$ive newish high-end GPU will give you best bang for your buck on this project. This is correct, especially when the CPU apps are often bottlenecked by memory bandwidth. This is the reason PrimeGrid codes its GPU apps such that they will explicitly not run on an iGPU. |
||
|
|
Richard Haselgrove
Senior Cruncher United Kingdom Joined: Feb 19, 2021 Post Count: 360 Status: Offline Project Badges: 
|
Worth noting is, that if you run CPU tasks at the same time as you run a iGPU task, you will slow down both the iGPU task, and the CPU tasks considerably. This was something we discovered on SETI@home, where we were running tons of different OpenCL GPU apps, and WU's. You can get away with running 1 or with luck 2 CPU tasks on a lets say 4c/8t CPU, but no more than that, or you will slow down all WU's considerably, including the one on the iGPU. It's also worth noting that these particular tests were run where the CPU tasks in question were highly optimised (hand crafted) SETI apps which pushed the CPU about as hard as it was possible to push it at the time, making extensive use of SIMD opcodes up to SSE4.1 and AVX. I suspect that using the GPU alongside those particular apps may also have stressed the overall power and thermal envelope limits of the CPU packages.It is, however, possible to find less-intensive projects where CPU and GPU can run in parallel without such serious side-effects. |
||
|
|
Grumpy Swede
Master Cruncher Svíþjóð Joined: Apr 10, 2020 Post Count: 2089 Status: Offline Project Badges:
|
Worth noting is, that if you run CPU tasks at the same time as you run a iGPU task, you will slow down both the iGPU task, and the CPU tasks considerably. This was something we discovered on SETI@home, where we were running tons of different OpenCL GPU apps, and WU's. You can get away with running 1 or with luck 2 CPU tasks on a lets say 4c/8t CPU, but no more than that, or you will slow down all WU's considerably, including the one on the iGPU. It's also worth noting that these particular tests were run where the CPU tasks in question were highly optimised (hand crafted) SETI apps which pushed the CPU about as hard as it was possible to push it at the time, making extensive use of SIMD opcodes up to SSE4.1 and AVX. I suspect that using the GPU alongside those particular apps may also have stressed the overall power and thermal envelope limits of the CPU packages.It is, however, possible to find less-intensive projects where CPU and GPU can run in parallel without such serious side-effects. You did mean "CPU and iGPU can run in parallel without such serious side-effects", didn't you? That was what my comment was about. CPU and discrete GPU, didn't usually have those effects, unless you over-taxed the CPU, and didn't take into account that the GPU had to be fed too, and left a core/thread to the GPU. [Edit 1 times, last edit by Grumpy Swede at Mar 13, 2021 5:08:41 PM] |
||
|
|
Richard Haselgrove
Senior Cruncher United Kingdom Joined: Feb 19, 2021 Post Count: 360 Status: Offline Project Badges:
|
Yes, happy to clarify that those remarks applied specifically to the Intel iGPUs. I have no experience of the AMD embedded GPUs, and external discrete GPUs are different again - they don't share the same TDP power constraints with the CPU.
|
||
|
|
[AF>Libristes]Maeda
Cruncher Joined: Sep 1, 2011 Post Count: 43 Status: Offline Project Badges: 
|
Indeed, for my HD4600 iGPU, I'm always keeping a (logical) core free.
|
||
|
|
Aegeon
Cruncher Joined: Oct 9, 2011 Post Count: 10 Status: Offline Project Badges: 
|
A machine that ran the last batch of betas with no issues has now produced errors on 5 of 7 tasks. Never seen a log file like this. All ran about 20 minutes so whatever the error was it occurred at or very near the end. Upon closer examination these exceeded the time limit. And now that machine is excluded from getting new tasks because of the errors. I can confirm I specifically got a seemingly identical set of errors for one of my computers with the latest batch on March 12th, although based on running times it appears he has a faster card than mine is. In my case it was 4 of the 7 units with an error and the error messages all include on top of everything else: "Unhandled Exception Detected... - Unhandled Exception Record - Reason: Breakpoint Encountered (0x80000003) at address 0x00007FFC486C1A52 Engaging BOINC Windows Runtime Debugger..." I can confirm the computer previously has run of bunch of the beta units from other dates without any issues, and based on the running times this is at least happening towards the end of processing these units timewise. In my case the computer has a RX 550 with 2 GB of ram, and this is a Windows 10 system. [Edit 2 times, last edit by Aegeon at Mar 13, 2021 8:19:46 PM] |
||
|
|
Richard Haselgrove
Senior Cruncher United Kingdom Joined: Feb 19, 2021 Post Count: 360 Status: Offline Project Badges:
|
Reason: Breakpoint Encountered ... BOINC deliberately exits that way when it exceeds the time limit. This time, there seem to be two problems. One is that on the big, discrete, GPUs the time limit has been set too low - that's an administrative problem, not a coding problem, and it can be sorted easily enough. The other seems to be that the iGPUs - especially the earlier ones, and perhaps only with particular drivers - never even get fully started, and make no significant progress. There's a third case - the older discrete GPUs may not be as OpenCL compliant as they claim. But that causes a different problem. The breakpoint triggered by the time limit over-run causes the extended debug log. It's intended to give the programmers something to work on, in case a program bug causes something like an endless loop. But I don't think it's that helpful in the cases this time. |
||
|
|
Grumpy Swede
Master Cruncher Svíþjóð Joined: Apr 10, 2020 Post Count: 2089 Status: Offline Project Badges:
|
The other seems to be that the iGPUs - especially the earlier ones, and perhaps only with particular drivers - never even get fully started, and make no significant progress. time. This one is interesting enough. Both of us have the iGPU HD4600, but only mine seems to crunch these Beta task OK. My iGPU is integrated into a i7-4790K CPU. I don't know which processor you have, but if it is a different one with the iGPU HD4600, it's possible that the problem is not driver related, but that it is something different in the integration of the iGPU in the processor. Edit, added: This is an interesting Nvidia site for some OpenCL test progams. NVIDIA OpenCL SDK Code Samples: https://developer.nvidia.com/opencl [Edit 1 times, last edit by Grumpy Swede at Mar 13, 2021 8:56:15 PM] |
||
|
|
|