CUDA crunching slower now?

Questions and Answers : GPU applications : CUDA crunching slower now?
Message board moderation

To post messages, you must log in.

1 · 2 · Next

AuthorMessage
Profile Mark Peters
Volunteer tester
Avatar

Send message
Joined: 5 Jul 02
Posts: 80
Credit: 588,422
RAC: 0
Belgium
Message 865937 - Posted: 16 Feb 2009, 0:33:43 UTC

Got a question about CUDA crunching. Took first about 6 to 15 min. Now it takes 4 hours, how come? Credit is even lower dan before.

Is there anybody who has the same issues?

Kind regards,

Mark
To boldly go where no human has gone before!
ID: 865937 · Report as offensive
Tribble

Send message
Joined: 21 Feb 02
Posts: 65
Credit: 7,978,002
RAC: 0
Australia
Message 865983 - Posted: 16 Feb 2009, 3:56:23 UTC - in response to Message 865937.  

I have an issue where it says it too just over 2 minutes of CPU time
to complete but infact it takes 40+ minutes of real time to complete.

What's with that? I get credit for 2mins of work yet it takes 40 minutes ...


Sooooo confusing.
ID: 865983 · Report as offensive
Profile perryjay
Volunteer tester
Avatar

Send message
Joined: 20 Aug 02
Posts: 3377
Credit: 20,676,751
RAC: 0
United States
Message 865988 - Posted: 16 Feb 2009, 4:09:28 UTC - in response to Message 865983.  

It takes a minute or so on the CPU to set up and then switches to GPU to finish. You get credit for the complete work unit. If you notice, most of your work will claim higher than your wingman running on CPU alone. Since he is claiming lower you will get the same as he claims. If you are paired with another running cuda you will both get the higher credit.


PROUD MEMBER OF Team Starfire World BOINC
ID: 865988 · Report as offensive
Tribble

Send message
Joined: 21 Feb 02
Posts: 65
Credit: 7,978,002
RAC: 0
Australia
Message 865991 - Posted: 16 Feb 2009, 4:22:17 UTC - in response to Message 865988.  

That doesnt explain the disparity of the work time and the real time though.
Any reason for that?
ID: 865991 · Report as offensive
Vager

Send message
Joined: 1 May 04
Posts: 10
Credit: 1,066,439
RAC: 0
Canada
Message 866032 - Posted: 16 Feb 2009, 6:05:34 UTC - in response to Message 865991.  

That doesnt explain the disparity of the work time and the real time though.
Any reason for that?


The BOINC Client only counts actual CPU time. Since most of the work is being done by the GPU the counter will not go up. There's still a small amount of CPU needed to feed the WU to your GPU when using CUDA.

Vager
ID: 866032 · Report as offensive
Profile BigWaveSurfer

Send message
Joined: 29 Nov 01
Posts: 186
Credit: 36,311,381
RAC: 141
United States
Message 866678 - Posted: 18 Feb 2009, 6:38:28 UTC - in response to Message 865937.  

Got a question about CUDA crunching. Took first about 6 to 15 min. Now it takes 4 hours, how come? Credit is even lower dan before.

Is there anybody who has the same issues?

Kind regards,

Mark


Mark,

Yes I am having the same issues, it started about a few days ago. I have been pushing through CUDA units in about 30 min (actual time) for the last few weeks, but now it is taking close to 2 hours to complete a CUDA unit. I figured it was a bigger file only to find out that I am getting the same credit I did for the 30 minute jobs! Something is not right here! My GPU temp has stayed around 68 at all times, however I have noticed since the painfully slow CUDA units have started my displays are lagging more than they did before, temp and fan speed are still what they were with the faster units.

One other thing I noticed was the quick units would finish in 30 min and have a CPU time of 4:30ish, but these slow units taking almost 2 hours are only showing 2:30ish on the CPU. I figured with the longer the GPU crunches the more info the CPU has to feed it.

Let me know if any of this sounds familiar.

Thanks!
ID: 866678 · Report as offensive
Profile BigWaveSurfer

Send message
Joined: 29 Nov 01
Posts: 186
Credit: 36,311,381
RAC: 141
United States
Message 866684 - Posted: 18 Feb 2009, 6:52:14 UTC - in response to Message 866678.  

Ok, did some digging into my results and thing I found a reason for it, however, would like someone to explain it to me :)

All of my quick CUDA units (ones that finish in 30 minutes) all have an angle range of 0.4xxxxx

All of the slow CUDA units (ones that finish in 2 hours) all have an angle range of 0.010xxxx

That seems to be the link between the slow units vs the fast units, however, why is this? Should I get more credit?

I remember before low work angles were crashing the program/drivers etc (it happened to me), but that seemed to be fixed, but now they just run really slow.

I never had these really low angle WU's until the last few days.

Any thoughts?
ID: 866684 · Report as offensive
Profile Jord
Volunteer tester
Avatar

Send message
Joined: 9 Jun 99
Posts: 15184
Credit: 4,362,181
RAC: 3
Netherlands
Message 866688 - Posted: 18 Feb 2009, 7:07:08 UTC

The Very Low Angle Range tasks run very slowly on CUDA. The developers know about this and are trying to fix that, but what they first wanted to get was stability. In the past when you would hit a so-called VLAR, your tasks would crash. They no longer crash, they do run, but just slower than normal.

I am not going into the credits discussion as for all I care they stop with that nonsense today.
ID: 866688 · Report as offensive
Profile BigWaveSurfer

Send message
Joined: 29 Nov 01
Posts: 186
Credit: 36,311,381
RAC: 141
United States
Message 866914 - Posted: 18 Feb 2009, 23:48:51 UTC - in response to Message 866688.  

What did you mean by this?

"I am not going into the credits discussion as for all I care they stop with that nonsense today."

I finally cleared out of the low angle units, wow, my computer is not so bogged down anymore. I agree, stability is of #1 importance. I do hope they also adress the credit issues as well down the time, either way I am still doing a LOT better than without CUDA.
ID: 866914 · Report as offensive
Profile Philip Stehno
Avatar

Send message
Joined: 17 May 99
Posts: 7
Credit: 10,241,552
RAC: 0
United States
Message 867365 - Posted: 20 Feb 2009, 15:18:41 UTC

I've noticed CUDA units take forever but I've also noticed another side effect and that is I only get CUDA work from SETI@HOME leaving no SETI work for the other processors in my machine.
ID: 867365 · Report as offensive
OzzFan Crowdfunding Project Donor*Special Project $75 donorSpecial Project $250 donor
Volunteer tester
Avatar

Send message
Joined: 9 Apr 02
Posts: 15691
Credit: 84,761,841
RAC: 28
United States
Message 867397 - Posted: 20 Feb 2009, 18:14:41 UTC - in response to Message 867365.  

I've noticed CUDA units take forever but I've also noticed another side effect and that is I only get CUDA work from SETI@HOME leaving no SETI work for the other processors in my machine.


CUDA is SETI@Home MultiBeam only, so if you are running a CUDA app, no other SETI MultiBeam apps are allowed to run (this is a current limitation of BOINC that will be addressed in a future version). To keep your other processors busy, you must either allow AstroPulse workunits or join another project. Note that one core will always be close to idle since it will be "feeding" the GPU and waiting for the data to return (this will also be fixed in a future version of BOINC).
ID: 867397 · Report as offensive
Nick

Send message
Joined: 17 May 99
Posts: 96
Credit: 17,356,094
RAC: 0
United States
Message 867479 - Posted: 21 Feb 2009, 0:09:01 UTC

I'd like to set my preferences to allow AstroPulse to run with CUDA but I only want to allow it on this machine, not my others. Where can I find the local preference file?


Nick

ID: 867479 · Report as offensive
OzzFan Crowdfunding Project Donor*Special Project $75 donorSpecial Project $250 donor
Volunteer tester
Avatar

Send message
Joined: 9 Apr 02
Posts: 15691
Credit: 84,761,841
RAC: 28
United States
Message 867491 - Posted: 21 Feb 2009, 0:47:39 UTC - in response to Message 867479.  

I'd like to set my preferences to allow AstroPulse to run with CUDA but I only want to allow it on this machine, not my others. Where can I find the local preference file?


Nick


I don't know if its set-able in the local preferences, but you can always set up a new preference Venue, and make sure the computer you want different preferences is set to use the new venue.
ID: 867491 · Report as offensive
Nick

Send message
Joined: 17 May 99
Posts: 96
Credit: 17,356,094
RAC: 0
United States
Message 867498 - Posted: 21 Feb 2009, 1:05:26 UTC - in response to Message 867491.  

you can always set up a new preference Venue, and make sure the computer you want different preferences is set to use the new venue.

Would you mind sharing how that is done?


ID: 867498 · Report as offensive
OzzFan Crowdfunding Project Donor*Special Project $75 donorSpecial Project $250 donor
Volunteer tester
Avatar

Send message
Joined: 9 Apr 02
Posts: 15691
Credit: 84,761,841
RAC: 28
United States
Message 867547 - Posted: 21 Feb 2009, 3:47:56 UTC - in response to Message 867498.  

you can always set up a new preference Venue, and make sure the computer you want different preferences is set to use the new venue.

Would you mind sharing how that is done?


You can go to SETI@Home Preferences, scroll to the bottom of the page and create a new venue. When you're done setting the prefs for the new venue, save them.

Then, go back to Your Computers and find the Host ID in question, click on its link to view the details, then near the bottom you can select its preference Venue that you'd like to assign to it.

The next time this machine contacts the servers, it will download the newer venue prefs.
ID: 867547 · Report as offensive
Profile Sutaru Tsureku
Volunteer tester

Send message
Joined: 6 Apr 07
Posts: 7105
Credit: 147,663,825
RAC: 5
Germany
Message 867563 - Posted: 21 Feb 2009, 4:36:47 UTC


Raistmer's V7 mod (MB on GPU and AP or other project on CPU) [multiple GPUs]
[available with kill VLAR-WUs mod]

Raistmer's V8 mod (MB on GPU and AK v8 MB on CPU) [only one GPU]
[I don't know if the VLAR-WU kill mod is included here]

Have a look also in the NC forum:
AK V8 + CUDA MB team work mod

ID: 867563 · Report as offensive
Tronic

Send message
Joined: 23 Mar 03
Posts: 8
Credit: 10,599,675
RAC: 0
Chile
Message 867581 - Posted: 21 Feb 2009, 5:08:23 UTC - in response to Message 867397.  

I've noticed CUDA units take forever but I've also noticed another side effect and that is I only get CUDA work from SETI@HOME leaving no SETI work for the other processors in my machine.


CUDA is SETI@Home MultiBeam only, so if you are running a CUDA app, no other SETI MultiBeam apps are allowed to run (this is a current limitation of BOINC that will be addressed in a future version). To keep your other processors busy, you must either allow AstroPulse workunits or join another project. Note that one core will always be close to idle since it will be "feeding" the GPU and waiting for the data to return (this will also be fixed in a future version of BOINC).


I saw this on another post:
You can try and add the cc_config.xml file to the Boinc data folder to get all the cores running.

I was running two CUDAs and one astropulse, leaving a large portion of my second CPU core untapped. Setting the number of cores to 3 allowed me to run both GPUs with CUDA and both CPU cores with astropulse.
ID: 867581 · Report as offensive
OzzFan Crowdfunding Project Donor*Special Project $75 donorSpecial Project $250 donor
Volunteer tester
Avatar

Send message
Joined: 9 Apr 02
Posts: 15691
Credit: 84,761,841
RAC: 28
United States
Message 867589 - Posted: 21 Feb 2009, 5:21:13 UTC - in response to Message 867581.  

I saw this on another post:
You can try and add the cc_config.xml file to the Boinc data folder to get all the cores running.

I was running two CUDAs and one astropulse, leaving a large portion of my second CPU core untapped. Setting the number of cores to 3 allowed me to run both GPUs with CUDA and both CPU cores with astropulse.


Yep, but I try not to provide such advanced information in Q&A unless the person shows technical skills. I wouldn't want anyone who doesn't know what they're doing to mess around with editing their .XML files, such as those who would use Internet Explorer or Wordpad to edit the document and mess up its formating.
ID: 867589 · Report as offensive
Nick

Send message
Joined: 17 May 99
Posts: 96
Credit: 17,356,094
RAC: 0
United States
Message 867593 - Posted: 21 Feb 2009, 5:27:36 UTC - in response to Message 867547.  

You can go to SETI@Home Preferences, scroll to the bottom of the page and create a new venue. When you're done setting the prefs for the new venue, save them.

There is no option to create a new "venue".

I see that I can select home, school or work. I will use one of those to customize the environment for this machine.

Thanks for getting me close.


ID: 867593 · Report as offensive
OzzFan Crowdfunding Project Donor*Special Project $75 donorSpecial Project $250 donor
Volunteer tester
Avatar

Send message
Joined: 9 Apr 02
Posts: 15691
Credit: 84,761,841
RAC: 28
United States
Message 867605 - Posted: 21 Feb 2009, 5:49:18 UTC - in response to Message 867593.  

You can go to SETI@Home Preferences, scroll to the bottom of the page and create a new venue. When you're done setting the prefs for the new venue, save them.

There is no option to create a new "venue".

I see that I can select home, school or work. I will use one of those to customize the environment for this machine.

Thanks for getting me close.


Those "home", "school" and "work" options are different venues. Same thing.
ID: 867605 · Report as offensive
1 · 2 · Next

Questions and Answers : GPU applications : CUDA crunching slower now?


 
©2024 University of California
 
SETI@home and Astropulse are funded by grants from the National Science Foundation, NASA, and donations from SETI@home volunteers. AstroPulse is funded in part by the NSF through grant AST-0307956.