Welcome to MilkyWay@home

GPU Credit Question


Advanced search

Message boards : Number crunching : GPU Credit Question
Message board moderation

To post messages, you must log in.

Previous · 1 · 2

AuthorMessage
Profilekashi

Send message
Joined: 30 Dec 07
Posts: 311
Credit: 149,490,184
RAC: 0
100 million credit badge13 year member badge
Message 41905 - Posted: 3 Sep 2010, 2:07:56 UTC
Last modified: 3 Sep 2010, 2:10:27 UTC

OK I've had a look. A single core of your 5970 2MB at default 725MHZ is completing a task in 105-120 seconds with a CPU time of 3.8-5 seconds when all is going correctly. However with 2 5970s installed some tasks are going very slowly and taking over 300 seconds with a very high CPU usage.

There are a number of possible reasons. Some I have thought of are:

One or more GPU cores are going to sleep when a screensaver is activating or the monitor is turned off or for some other reason. This will cause 2 tasks to be processed on one GPU core (the second task may be processed on a phantom core created by BOINC which does not show up in the stderr out) but it does not do it efficiently as when 2 tasks are configured to run per core in a app_info.xml file. Some have cured this by various combinations of adjusting settings such as extending second card onto one monitor, enabling or disabling CrossFire, changing instances of EnableULPS=1 to EnableULPS=0 in Windows registry and rebooting, attaching an extra VGA connector to a card in addition to a DVI or DisplayPort connector, attaching the second card to an additional monitor, etc. Others such as myself could only permanently cure this problem with the use of dummy plugs. It is important to note that it is necessary to stop and restart BOINC or reboot after changing any of these settings. In particular enabling and disabling CrossFire without stopping and restarting BOINC afterwards has been known to cause problems.

The driver is resetting when Flash video is played. This may cause the task to be restarted. This can be fixed by unticking "use hardware acceleration" within the Flash video settings. Easiest place to do this is on Flash website, right click the Flash content there to gain access to Flash configuration. This only affects certain Catalyst versions, it is present in 10.4 version I use.

The Catalyst driver version you are using is unstable with your configuration. Cure usually involves trying different driver versions.

Running intensive video applications is causing slowdown by hogging video resources.

Running intensive CPU applications on all 12 "cores" of CPU is restricting availability of CPU resources to support GPU processing. Can be tested by leaving all CPU cores idle if possible by stopping or suspending any running programs or BOINC CPU tasks.

Cards are installed close together and one is overheating and throttling. Increasing GPU fan speed, removing side of case and blowing a floor fan in the side or otherwise improving case cooling can help this.
MilkyWay does not benefit from high GPU memory speed, if you reduce the GPU memory speed you will reduce GPU temperature and save some power. I use a GPU memory speed of 500MHz on MilkyWay and DNETC.

This one is unlikely because it would probably cause your whole system to become slow and unresponsive or be unable to boot but 2 2MB cards can sometimes cause memory remap problems. BIOS on some motherboards has settings to adjust this. On my DFI it is called Memory LowGap, other brands use other names for this.

There are other possibilities I haven't thought of and others with a mutiple 5970 configuration may be better able to advise you and offer suggestions. It can be very annoying and frustrating when new hardware does not perform properly. However it took me a long time and a lot of patience before I was able to use multiple cards successfully. It is the dual core cards that are most likely to have issues. When you have sufficient time you need to monitor things closely via both BOINC Manager and GPU-Z to find out what is happening when the GPU tasks are running slowly. As you become more familiar with when the problem occurs and how it shows up a pattern may emerge which will help you to diagnose what is going wrong.
ID: 41905 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Len LE/GE

Send message
Joined: 8 Feb 08
Posts: 261
Credit: 104,050,322
RAC: 0
100 million credit badge13 year member badge
Message 41907 - Posted: 3 Sep 2010, 3:03:07 UTC

You did run some n-body simulations (cpu only) as well, which are the ones with 6 minutes and more.

Now to your different gpu times. I found the following:
CPU time: 4.78923 seconds,
GPU time: 107.745 seconds,
wall clock time: 220.84 seconds

There is clearly something else taking time on the gpu. Since I have no dual gpu card, I have seen such behavior only when 2 WUs are started at the same time on a gpu. One is waiting while the other is crunching.
ID: 41907 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Profilekashi

Send message
Joined: 30 Dec 07
Posts: 311
Credit: 149,490,184
RAC: 0
100 million credit badge13 year member badge
Message 41908 - Posted: 3 Sep 2010, 3:39:09 UTC
Last modified: 3 Sep 2010, 3:40:27 UTC

Yes and the tasks with wallclock time of about 220 seconds are showing only one core in stderr out, "Found 1 CAL device"

Assuming there is only one 5970 installed when these tasks are being processed this means that the other core is not being recognised by the MilkyWay application perhaps because it has gone to sleep. As pointed out by Len LE/GE the one core that is available to the MilkyWay application is probably processing 2 tasks concurrently.

Although both tasks would actually be processing on GPU Core 0, BOINC Manager would probably be showing the tasks processing on Core 0 and Core 1. This is why you need to check GPU-Z GPU load percentage to see if all installed cores are working as looking at BOINC Manager alone will give you misleading information about this.
ID: 41908 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Ba

Send message
Joined: 6 Nov 09
Posts: 12
Credit: 348,876,876
RAC: 0
300 million credit badge12 year member badge
Message 41911 - Posted: 3 Sep 2010, 16:03:11 UTC

I have the same problem, just installed my second 5970 on a clean XP64 install.

I have dummy plugs installed and have extended the desktop, boinc sees 4 gpu's and crunches 4 WU's but two take twice as long to complete.
GPUZ shows one gpu with zero activity so i assume two WU's are being crunched on one gpu.
ID: 41911 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Andy

Send message
Joined: 27 Aug 10
Posts: 19
Credit: 153,469,603
RAC: 3,768
100 million credit badge11 year member badge
Message 41912 - Posted: 3 Sep 2010, 16:07:37 UTC - in response to Message 41911.  

thanks for all the replies again...I've tried it all and given up. Most simple way for me is just use 1 GPU per PC. I've also noticed that a single HD 5870 1 GB does it the same speed as a 4 GB HD 5970!
ID: 41912 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Fred J. Verster

Send message
Joined: 22 Apr 09
Posts: 38
Credit: 27,377,932
RAC: 0
20 million credit badge12 year member badge
Message 41913 - Posted: 3 Sep 2010, 16:23:21 UTC - in response to Message 41912.  
Last modified: 3 Sep 2010, 16:46:40 UTC

Start messages from BOINC :
3-9-2010 14:44:45 Starting BOINC client version 6.10.56 for windows_intelx86
3-9-2010 14:44:45 Config: use all coprocessors
3-9-2010 14:44:45 log flags: file_xfer, sched_ops, task
3-9-2010 14:44:45 Libraries: libcurl/7.19.7 OpenSSL/0.9.8l zlib/1.2.3
3-9-2010 14:44:45 Data directory: C:\Documents and Settings\All Users\Application Data\BOINC
3-9-2010 14:44:45 Running under account Gebruiker
3-9-2010 14:44:45 Processor: 4 GenuineIntel Intel(R) Core(TM)2 Quad CPU Q6600 @ 2.40GHz [Family 6 Model 15 Stepping 11]
3-9-2010 14:44:45 Processor: 4.00 MB cache
3-9-2010 14:44:45 Processor features: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush dts acpi mmx fxsr sse sse2 ss htt tm pni ssse3 cx16 nx lm vmx tm2 pbe
3-9-2010 14:44:45 OS: Microsoft Windows XP: Professional x86 Edition, Service Pack 3, (05.01.2600.00)
3-9-2010 14:44:45 Memory: 2.00 GB physical, 3.85 GB virtual
3-9-2010 14:44:45 Disk: 195.31 GB total, 175.79 GB free
3-9-2010 14:44:45 Local time is UTC +2 hours
3-9-2010 14:44:45 ATI GPU 0: ATI Radeon HD5700 series (Juniper) (CAL version 1.4.792, 1024MB, 1360 GFLOPS peak)
3-9-2010 14:44:45 ATI GPU 1: ATI Radeon HD 4700/4800 (RV740/RV770) (CAL version 1.4.792, 1024MB, 1000 GFLOPS peak)

3-9-2010 14:44:45 Collatz Conjecture Found app_info.xml; using anonymous platform
3-9-2010 14:44:45 Milkyway@home Found app_info.xml; using anonymous platform
3-9-2010 14:44:45 SETI@home Found app_info.xml; using anonymous platform
3-9-2010 14:44:45 SETI@home Beta Test Found app_info.xml; using anonymous platform
3-9-2010 14:44:46 Leiden Classical URL http://boinc.gorlaeus.net/; Computer ID 58213; resource share 75
3-9-2010 14:44:46 Collatz Conjecture URL http://boinc.thesonntags.com/collatz/; Computer ID 11990; resource share 75
3-9-2010 14:44:46 Docking URL http://docking.cis.udel.edu/; Computer ID 29738; resource share 75


This looks like an only WIN XP 'thing'!? Needing a dummy load, haven't tried yet, if it could be a driver issue.

On SETI BĂȘta, there is a serious problem with Double GPU card's, but they are using/testing a different (OpenCL) app.
Won't access 1 0r 3 Devices, it looks like.
Is there already support for Single Precision Card's, or is that out off the question? ;^)

I take it, a (FERMI)GTX470 could run this app., too? Just curious............

Knight Who says Ni
ID: 41913 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Ba

Send message
Joined: 6 Nov 09
Posts: 12
Credit: 348,876,876
RAC: 0
300 million credit badge12 year member badge
Message 41929 - Posted: 4 Sep 2010, 22:29:05 UTC

Still have this problem and ive tried it all including bios upgrade dummy plugs extended desktop.
Ive also tried various driver versions and im still crunching 4 WU's but two of them are on one gpu with the forth gpu zero activity.

Ive running out of things to do even tried cc_config with use all gpus still nothing.

Any ideas would be great Thanks.
ID: 41929 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Ba

Send message
Joined: 6 Nov 09
Posts: 12
Credit: 348,876,876
RAC: 0
300 million credit badge12 year member badge
Message 41940 - Posted: 5 Sep 2010, 12:33:22 UTC

Got this working in the end.

I enabled crossfirex this then had me running 6 WU's one on core 0 and 1 and two on core 2 and 3!?
Then i took one of the dummy plugs off the second card and only extended the desktop to the one remaining, now it crunches one per core.
Going to try this in collatz to see if it works i hear that project has crossfire problems.
Will repot back.
ID: 41940 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
mdawson

Send message
Joined: 9 Jul 09
Posts: 13
Credit: 5,141,953
RAC: 0
5 million credit badge12 year member badge
Message 42004 - Posted: 9 Sep 2010, 5:42:04 UTC - in response to Message 41895.  

Kashi,

I read your reply and it got me to wondering about my cards. GPU-Z indicates that my "little" card has a 0% load. My "big" card is at 89%. Both of my cards (nVidia) have two outputs. The "big" card, a GTX 260 has both outputs connected to 2 physical monitors. The "little" card, an 8600GT has only one output connected to the secondary input of the right monitor. (The left monitor has only one input, the right monitor has two inputs connected, but only one in use at any given time).

If I were to follow your reasoning, the fact that the 2nd card is plugged into the 2nd input of the monitor, and is generally never activated, would be the reason why I have 0% load on the "little" card, correct? I have my desktop spread across all 3 monitor inputs. That was the only way to get the card recognized.

I'm thinking that I should invest in dummy connectors as well. You say you made them? Are there directions anywhere, or can they be purchased from Fry's or somebody like that?

Oddly, BOINC seems to indicate that both cards are operational.
ID: 42004 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
ProfileWerkstatt

Send message
Joined: 19 Feb 08
Posts: 350
Credit: 139,323,728
RAC: 383
100 million credit badge13 year member badge
Message 42005 - Posted: 9 Sep 2010, 7:55:21 UTC - in response to Message 42004.  


I'm thinking that I should invest in dummy connectors as well. You say you made them? Are there directions anywhere, or can they be purchased from Fry's or somebody like that?


Hi,
the easiest way is to take an adapter from DVI to VGA, three resistors near 75 Ohms (68 Ohm, 82 Ohm are working as well) and just stick them in as follows:
Pin1 to Pin6
Pin2 to Pin7
Pin3 to Pin8
That's it!
If you want to have something more robust, you should take a 15pin HD D-Sub male connector and solder the tree resistors to the pins as described.

Alexander
ID: 42005 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Profilekashi

Send message
Joined: 30 Dec 07
Posts: 311
Credit: 149,490,184
RAC: 0
100 million credit badge13 year member badge
Message 42008 - Posted: 9 Sep 2010, 12:00:56 UTC
Last modified: 9 Sep 2010, 12:07:52 UTC

I am not familiar with any NVIDIA configurations, so I can't speak with any experience. However I did read on one the BOINC ATI project forums of someone having trouble when connecting to 3 monitors. Can't remember if they were using ATI or NVIDIA though. I wasn't sure if NVIDIA cards show GPU load percentage correctly in GPU-Z, perhaps an NVIDIA owner would be able to clarify that for you.

As for the dummy plugs I followed the instructions from the link below and found it very easy. Used a nail clipper to snip the resistor lead to the correct length.

http://www.overclock.net/folding-home-guides-tutorials/384733-30-second-dummy-plug.html

I got 68 ohm resistors from Jaycar which is a chain of small electronics stores in Australia. I have heard of Frys, don't know what they sell but they are in America I think.
ID: 42008 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
John Clark

Send message
Joined: 4 Oct 08
Posts: 1734
Credit: 64,228,409
RAC: 0
50 million credit badge13 year member badge
Message 42012 - Posted: 9 Sep 2010, 15:10:27 UTC - in response to Message 42008.  

I am not familiar with any NVIDIA configurations, so I can't speak with any experience. However I did read on one the BOINC ATI project forums of someone having trouble when connecting to 3 monitors. Can't remember if they were using ATI or NVIDIA though. I wasn't sure if NVIDIA cards show GPU load percentage correctly in GPU-Z, perhaps an NVIDIA owner would be able to clarify that for you.

As for the dummy plugs I followed the instructions from the link below and found it very easy. Used a nail clipper to snip the resistor lead to the correct length.

http://www.overclock.net/folding-home-guides-tutorials/384733-30-second-dummy-plug.html

I got 68 ohm resistors from Jaycar which is a chain of small electronics stores in Australia. I have heard of Frys, don't know what they sell but they are in America I think.



The only multi-GPU rig I run has a mix of ATI (for graphics) and legacy nVidia. As I understand a second, or more, ATI GPU will need a dummy plug (as described in the link) and the desktop extended across to the second monitor. I do not use a dummy plug on my NV GPUs as the latest driver (258.96) will recognise them. All they need is the desktop extending in Display Properties/Settings/
Go away, I was asleep


ID: 42012 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
ProfileuBronan
Avatar

Send message
Joined: 9 Feb 09
Posts: 166
Credit: 27,520,813
RAC: 0
20 million credit badge12 year member badge
Message 42024 - Posted: 10 Sep 2010, 0:35:52 UTC - in response to Message 42012.  


The only multi-GPU rig I run has a mix of ATI (for graphics) and legacy nVidia. As I understand a second, or more, ATI GPU will need a dummy plug (as described in the link) and the desktop extended across to the second monitor. I do not use a dummy plug on my NV GPUs as the latest driver (258.96) will recognise them. All they need is the desktop extending in Display Properties/Settings/


Sadly that does not allways work John, even though i admit the newer driver is much better in controlling multi cards

Its most of the time the os which messes up if your not running sli / crossfire.
When you boot up the system clearly see the extra cards.

One of my mates have a system with 3 nvidia cards but somehow windows did put those 2 extra cards sometimes in sleep mode.

He used to connect shortly the monitor plug on them to start crunch and then switch to the other untill they all ran again. But that was not a very handy solution.
After i installed 2 dummy plugs he does no longer have to check they work allways now.

Its new, its relative fast... my new bicycle
ID: 42024 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Previous · 1 · 2

Message boards : Number crunching : GPU Credit Question

©2021 Astroinformatics Group