Einstein Broken? Need help
6 Dec 2013 21:25:38 UTC
Topic 197298
(moderation:
Hi,
I think somehow the Einstein part of BINC got broken. Since I only recently added GPU computation I decided to read further into it and yesterday night wanted to try to run 2 Tasks on one GPU, so I set my Desktop Computer to a different Preference set, set the Value of the GPU utilization factor to 0.5 and voila, 2 Tasks running on my Graphics Card. GPU Temperature went up from 66 to 75°C and Power Consumption of the System increased by 20 Watts. Thought that's a good thing as this means it's using more of the GPU.
Problem: the next days both GPU tasks were at a runtime of 12 hours and only 20% each. That's quite worse than ~1.5 hours per Task it took before. So I did reset the Value for GPU untilization back to 1. And this messed everything up. There are still 2 tasks running on the GPU, but after reporting a few tasks now the CPU Tasks don't make any progress, they are stuck at whatever percentage they are, the new ones stay at 0%. Running, and running but never increase.
I tried a restart, changing the settings online, reinstalling the boinc client.
Tasks from the WCG run fine and so do GPU tasks.
System is a 2010 MacPro 6 Core 3.33 Ghz Xeon CPU
Running OSX 10.8.5
ATI 5870
Any advice to get the machine "healthy" again?
Einstein Broken? Need help
)
If you still want to try running x2 then try to free one or two CPU cores by setting "On multiprocessors, use at most xx% of the processors" to either 84% to free 2 cores or 92% to free 1 core.
That should get the GPU tasks running faster.
Otherwise when you change the utilization factor the setting is not applied until you download new GPU tasks. So if you want to go back to one task at a time quickly then you might need to increase you cache settings a bit so that Boinc downloads new tasks for the GPU and the new utilization factor, remember to set you cache back afterwards.
I agree: you probably needed
)
I agree: you probably needed to also free at least another CPU core to see a benefit from running 2 GPU WUs in parallel. Otherwise.. I have no idea how good and current the AMD Mac drivers are and how Mac OS handles GPU task scheduling - there could be some strangeness going on here.
To get Einstein working again you might need to reset the online settings (since you probably already reset everything locally be reinstalling BOINC). Or wait - did you remove / deinstall the old BOINC installation prior to reinstalling? Otherwisse the "broken state" might just remain in the data files, while the reinstallation just replaced the executables with the same versions.
MrS
Scanning for our furry friends since Jan 2002
Thanks for the advice, I
)
Thanks for the advice, I already had set the CPU to 90%, so even with GPU Computing I had another whole free core for the office stuff when not doing graphics work on the machine. Seems like the 5870 doesn't like running two tasks.
Will try to get new WUs then to update the settings, hopefully then CPU computing will work again as well.
I didn't remove the whole installation as I didn't want to loose the Tasks in progress, if with the new GPU settings things don't work fine again I'll empty the cache and then remove the settings on the machine prior to reinstalling.
How to reset the online settings beside setting them back to default and then update?
"Setting them back to
)
"Setting them back to default" is what I meant. Well.. if you only changed the BRP utilization factor, then it should already be back to where it was. But you shouldn't have these problems anyway ;)
MrS
Scanning for our furry friends since Jan 2002
GPU Tasks work quite fine
)
GPU Tasks work quite fine again. What I learned from that: the warning not to modify the utilisation factor is there for a reason, although I thought I know what I'm doing I obviously didn't. Also either something is wrong with my card or generally the ATI 5870, from what I read a Workunit uses 300MB of the GPU Memory, with the 5870 having 1GB this should be fine for 2 Tasks at once but for some reason the card uses more power while being way more inefficient.
Even with only one task running it took the downloaded Tasks twice as long as usual but things seem to go back to normal again.
According to monitoring tools the GPU is using about 80 watts, while it is specified with ~220 watts max. It manages about 7000 credits a day at the same time the 2 years newer Nvidia 650M on my laptop is capable of 14 000 credits a day. I know 2 years is a lot on the GPU market but I'm still surprised how easily a low power notebook card beats the Desktop one.
Wether the normal CPU Task will work again I couldn't test yet as I filled my Cache with WCG tasks to take part at the christmas race there.
Usually a reboot will get
)
Usually a reboot will get those tasks running again on your Boinc Manager.
(and at times just a "project Update" on the Boinc Manager)
And yes you have to wait for a task update before it switches from 0.5 to the new setting.
I do know those Ati's are much slower than a GeForce card.
One of mine is a Radeon HD 7500 and it takes about 20 hour to run 2 tasks (.05)
I read once that a ATI 5870 can be overclocked.
It is a bit hot but not too bad so just keeping it in a cool room with maybe a fan next to it.
But as you know it would be much faster just using another nVidia card (I guess the one you have is in a laptop)
OC a 5870 <---
Unfortunately CPU Tasks
)
Unfortunately CPU Tasks (Gravitational Wave S6) still don't work. They are running but making no progress at all. I did abort the old tasks, downloaded new ones, but no chance.
I'll let the cache run empty and then recover the Einstein project from an older backup, hopefully that helps. Otherwise I'd be left clueless.
Maybe a driver Issue with the ATIs? After all NVIDIA is officially supporting GPU Computing for quite some time, while ATI made it possible only quite late.
Overclocking is not an option as I'm on a Mac, and since Apple is offering no more Computers with PCIe slots it's unlikely that new GPUs will arrive for the platform. While Saphire made the ATI 7950 available for Mac, as far as I know there's no NVIDIA GPU available. Well work performance is more important than crunching performance on that machine anyway, still would be nice to have both.
RE: Unfortunately CPU Tasks
)
The Gravitational Wave S6 tasks run OK it's just that they don't checkpoint frequently which makes it appear they aren't running. Depending on the speed of your system they might checkpoint every 60 minutes, maybe more, maybe less. Just let them run and they'll finish.
I'm not sure of what all is going wrong with your system but if it's just that the Gravitational Wave S6 tasks don't seem to progress then try just letting them run. Might save you the hassle of a restore/recover op.
Could be a driver issue but if the driver version you're using is the same as what everybody else is using or you don't find any warnings about it in these fora then consider it a good version. If nothing else works then try experimenting with different driver versions.
That sucks big time!!
Well then ignore all the Apple spin that nitwit Jobs spews and know that you can have exactly the same hardware as they put in those overpriced, machined, brushed aluminum apple crates at a far cheaper price from numerous sources. Then buy OSX and install it. Finally, take a bite from an apple, take a picture of it and glue it on your case. There you go... an Apple for the price of a regular PC. With an expansion slot, room for cooling upgrades and overclockable... the way a manly man's computer should be.
RE: as I'm on a
)
Alright, I suppose we've found your problem! According to still somewhat recent measurements a HD5870 should be able to reach ~27k RAC. A HD5850 should do ~24k, with moderate gain from a 2nd WU to ~25k RAC.
These are only rough guidelines, but a far cry from your 7k. This makes me suspect something's really broken with AMDs OpenCL driver for Mac or maybe MacOS itself - who knows, there's a lot that could go wrong.
MrS
Scanning for our furry friends since Jan 2002
Indeed that S6 Tasks did run
)
Indeed that S6 Tasks did run in the end. But there are still problems, some task finish and upload as usual. Most Tasks finish with "Waiting to run" as Status. I manually have to force them to run again and only after running for another while they upload. All tasks take longer than usual to finish. I guess I have to try to solve things with an old backup though.
MacPro, 3.33 Ghz Xeon:
http://einsteinathome.org/host/3426951/tasks&offset=0&show_names=1&state=3&appid=24
compared to my other machine,
MacBook Pro 2.6 Ghz i7:
http://einsteinathome.org/host/6045593/tasks&offset=0&show_names=1&state=3&appid=24
Edit: Runtime for S6 on:
2006 Core 2 Duo 2.2Ghz ~80,000 sec
2010 Xeon 3.33 Ghz ~55,000 sec
2012 i7 2.6 Ghz ~23,000 sec
I don't quite get how a 4 year step between a CPU is smaller than a 2 years step, especially since one of them is a Server CPU while the others are Notebook ones.
Yeah, this probably are the two last Macs I bought, the new stuff isn't made for Pros anymore, still since I need the computers for Work support and reliability is a big concern when installing MacOSX on a stock PC Hardware.
Woah, 27k RAC would be something to dream about. Well fits to the power usage. A ATI 5870 can go up to 220 Watts. My card is using about 50 Watts. so let's do the math:
50 x 4 = 200
7000 x 4 = 28000
At the same time when GPU Tasks are running switching for example between windows is already much slower, so I think there is a general problem with the Card when running at 25% of it's capacity it can't handle a 24" Screen, which much weaker cards have no problem with. Have no idea how to deal with that though...