Have your cake and eat it too!
log in

Advanced search

Message boards : News : Have your cake and eat it too!

Author Message
Profile Slicker
Volunteer moderator
Project administrator
Project developer
Project tester
Project scientist
Avatar
Send message
Joined: 11 Jun 09
Posts: 2525
Credit: 740,580,099
RAC: 1
Message 18632 - Posted: 18 Feb 2014, 0:34:59 UTC

While most people wanted only CUDA applications, some had some good reasons for using OpenCL applications which means it really can't be one or the other. As a result, you can use whichever you want or even both by selecting your preferred plan class(es) for each venue.

The Windows and Linux CUDA applications, both 32 and 64 bit, have been released so your preferences should take effect the next time your computer requests work. The default is that all plan classes are selected.

EmSti [BlackOps]
Send message
Joined: 1 Apr 12
Posts: 5
Credit: 1,084,481,298
RAC: 0
Message 18634 - Posted: 18 Feb 2014, 1:16:00 UTC - in response to Message 18632.

When you re-edit the preferences, it looks like everything gets rechecked.

Profile arkayn
Volunteer tester
Avatar
Send message
Joined: 30 Aug 09
Posts: 219
Credit: 676,877,192
RAC: 14,459
Message 18635 - Posted: 18 Feb 2014, 3:01:02 UTC - in response to Message 18634.

When you re-edit the preferences, it looks like everything gets rechecked.


And it deleted my other venue as well.
____________

Profile nenym
Send message
Joined: 21 Jul 09
Posts: 11
Credit: 781,284,010
RAC: 314,800
Message 18636 - Posted: 18 Feb 2014, 3:33:17 UTC - in response to Message 18635.

Re-edit delete all profiles and all apps are rechecked in default profile.

Profile Slicker
Volunteer moderator
Project administrator
Project developer
Project tester
Project scientist
Avatar
Send message
Joined: 11 Jun 09
Posts: 2525
Credit: 740,580,099
RAC: 1
Message 18638 - Posted: 18 Feb 2014, 4:45:29 UTC

I guess that when BOINC finds a missing ">" on one XML element, the only thing it can do is delete everything. It should be fixed now.

Profile arkayn
Volunteer tester
Avatar
Send message
Joined: 30 Aug 09
Posts: 219
Credit: 676,877,192
RAC: 14,459
Message 18639 - Posted: 18 Feb 2014, 5:19:01 UTC - in response to Message 18638.

I guess that when BOINC finds a missing ">" on one XML element, the only thing it can do is delete everything. It should be fixed now.


Works fine now.
____________

Profile [AF>Amis des Lapins] Phil1966
Send message
Joined: 25 Dec 11
Posts: 17
Credit: 275,179,579
RAC: 3,104,118
Message 18640 - Posted: 18 Feb 2014, 6:00:49 UTC

Thank You Very Much !

Profile Slicker
Volunteer moderator
Project administrator
Project developer
Project tester
Project scientist
Avatar
Send message
Joined: 11 Jun 09
Posts: 2525
Credit: 740,580,099
RAC: 1
Message 18641 - Posted: 18 Feb 2014, 6:10:08 UTC - in response to Message 18640.

Thank You Very Much !


Thanks for the feedback which prompted the preferences change in the first place.

Acey Pilot
Avatar
Send message
Joined: 13 Jan 14
Posts: 7
Credit: 54,938,691
RAC: 0
Message 18645 - Posted: 18 Feb 2014, 13:17:10 UTC - in response to Message 18632.

While most people wanted only CUDA applications, some had some good reasons for using OpenCL applications which means it really can't be one or the other. As a result, you can use whichever you want or even both by selecting your preferred plan class(es) for each venue.

The Windows and Linux CUDA applications, both 32 and 64 bit, have been released so your preferences should take effect the next time your computer requests work. The default is that all plan classes are selected.




EXCELLENT WORK! THANK YOU SLICKER!

Profile eXtreme Warhead
Avatar
Send message
Joined: 18 Nov 12
Posts: 15
Credit: 229,064,135
RAC: 82,800
Message 18667 - Posted: 19 Feb 2014, 17:31:49 UTC - in response to Message 18645.
Last modified: 19 Feb 2014, 17:34:13 UTC

gpu: 2,394.75s cpu: 95.55s points:7,285.78 solo_collatz v6.03 (cuda55)
gpu: 3,925.08s cpu: 3,868.61s points:7,323.67 solo_collatz v6.04 (opencl_nvidia_gpu)
no optimized apps, only two wu at the same time on a 660ti.

i think my choice for cuda was the right one...i don't think that opencl on nvidia has any advantages...

Profile Zydor
Avatar
Send message
Joined: 19 Aug 09
Posts: 364
Credit: 840,811,292
RAC: 0
Message 18668 - Posted: 19 Feb 2014, 17:46:55 UTC - in response to Message 18667.

gpu: 3,925.08s ......


Suggest you take a peek at your CPU setting, likely its at circa 0.05 (ish).

It needs to be up at 0.25 to 0.5, may as well be 0.5 if no other demands on CPUs.

Keeping the figure down at the classic 0.05 (etc) will chain the apps here to the deck, they need the higher CPU figure, eg:

<gpu_versions>
<gpu_usage>1.0</gpu_usage>
<cpu_usage>0.5</cpu_usage>
</gpu_versions>

Set like that and you likely will go to circa +/- 800/900 secs all other things being equal in the configuration.

The other issue is often the third config file not being set and left blank, see:

http://boinc.thesonntags.com/collatz/forum_thread.php?id=1009

Profile Slicker
Volunteer moderator
Project administrator
Project developer
Project tester
Project scientist
Avatar
Send message
Joined: 11 Jun 09
Posts: 2525
Credit: 740,580,099
RAC: 1
Message 18669 - Posted: 19 Feb 2014, 21:13:05 UTC

I found a bug that when multiple venues existed, if any venue was set to ignore a plan class, then all venues would ignore the plan class. A fix has been released. Let me know if you run into any issues.

Profile eXtreme Warhead
Avatar
Send message
Joined: 18 Nov 12
Posts: 15
Credit: 229,064,135
RAC: 82,800
Message 18670 - Posted: 19 Feb 2014, 21:22:40 UTC - in response to Message 18668.
Last modified: 19 Feb 2014, 21:48:46 UTC

gpu: 3,925.08s ......


Suggest you take a peek at your CPU setting, likely its at circa 0.05 (ish).

i gave the wu ONE whole cpu-core and not only 0,05...

if you mention the question in the other thread, i'm the one with the 660ti probably @stock config vs a faster, but optimized hd5850

appinfo looks like:
<gpu_versions>
<gpu_usage>0.5</gpu_usage>
<cpu_usage>1.0</cpu_usage>
</gpu_versions>


PS: runtimes from cuda-wu rised up to 3400s, but still only <100s cpu. and still this contradicts the info from the cuda/opencl poll-thread

Source of the below info : PrimeGrid Preferences :

CUDA is usually faster on pre-Kepler GPUs (GTX 5xx and older).
OpenCL is usually faster on Kepler GPUs (GTX 6xx and newer). Use whichever is faster.

Profile Slicker
Volunteer moderator
Project administrator
Project developer
Project tester
Project scientist
Avatar
Send message
Joined: 11 Jun 09
Posts: 2525
Credit: 740,580,099
RAC: 1
Message 18677 - Posted: 20 Feb 2014, 2:55:22 UTC - in response to Message 18670.

gpu: 3,925.08s ......


Suggest you take a peek at your CPU setting, likely its at circa 0.05 (ish).

i gave the wu ONE whole cpu-core and not only 0,05...

if you mention the question in the other thread, i'm the one with the 660ti probably @stock config vs a faster, but optimized hd5850

appinfo looks like:
<gpu_versions>
<gpu_usage>0.5</gpu_usage>
<cpu_usage>1.0</cpu_usage>
</gpu_versions>


PS: runtimes from cuda-wu rised up to 3400s, but still only <100s cpu. and still this contradicts the info from the cuda/opencl poll-thread

Source of the below info : PrimeGrid Preferences :

CUDA is usually faster on pre-Kepler GPUs (GTX 5xx and older).
OpenCL is usually faster on Kepler GPUs (GTX 6xx and newer). Use whichever is faster.



I think what he is saying is that if you edit the file

c:\ProgramData\BOINC\projects\boinc.thesonntags.com_collatz\solo_collatz_windows_x86_64_cuda.config

and add the lines:

verbose=1
threads=9
items_per_kernel=20
kernels_per_reduction=8
sleep=1

Your times will decrease even more. Those same settings should also work for the other opencl_nvidia_gpu.config file as well.

Profile eXtreme Warhead
Avatar
Send message
Joined: 18 Nov 12
Posts: 15
Credit: 229,064,135
RAC: 82,800
Message 18682 - Posted: 20 Feb 2014, 21:52:56 UTC - in response to Message 18677.

as told in the other thread: it's not possible to run collatz with these settings on the nvidia, because the video gets "slimy". those optimizations runs perfectly on a hd5850, but those system is not my personal desktop and no one use it, so a possible video stuttering doesn't interest anyone ;)

the only thing i want to show was the difference between opencl and cuda on a 660ti and with it that the primegrid opinion isn't correct

Profile Slicker
Volunteer moderator
Project administrator
Project developer
Project tester
Project scientist
Avatar
Send message
Joined: 11 Jun 09
Posts: 2525
Credit: 740,580,099
RAC: 1
Message 18683 - Posted: 20 Feb 2014, 22:40:26 UTC - in response to Message 18682.

as told in the other thread: it's not possible to run collatz with these settings on the nvidia, because the video gets "slimy". those optimizations runs perfectly on a hd5850, but those system is not my personal desktop and no one use it, so a possible video stuttering doesn't interest anyone ;)

the only thing i want to show was the difference between opencl and cuda on a 660ti and with it that the primegrid opinion isn't correct


Are you running GPU-Z whle running each to make sure that the GPU load remains the same whether running CUDA or OpenCL? If one is at 40% load and the other at 90% load with the default settings, it really isn't a fair comparison. Also, unless both are running the same WU, it isn't a fair comparison either since the WUs vary a little in length.

That having been said, it may depend upon which CUDA version was used to compile the PG apps. Using an old version (e.g. CUDA 2.3) will allow the appto work with just about any nVidia drivers but it won't have the compute capability 1.2, 1.3, 2.0, 2.1, 3.0 or 3.5 optimizations. Since the OpenCL app is compiled by the driver at runtime, it will have the best optimizations for the GPU. I would think that compiling with CUDA 5.5 and having the binary include builds for Compute 1.0 thru 3.5 so each GPU can choose the one that matches their capabilities would allow the PG CUDA app to be faster than the OpenCL app but it would require users to upgrade to recent drivers.

Profile eXtreme Warhead
Avatar
Send message
Joined: 18 Nov 12
Posts: 15
Credit: 229,064,135
RAC: 82,800
Message 18686 - Posted: 20 Feb 2014, 23:54:07 UTC - in response to Message 18683.
Last modified: 21 Feb 2014, 0:03:42 UTC

Are you running GPU-Z whle running each to make sure that the GPU load remains the same whether running CUDA or OpenCL? If one is at 40% load and the other at 90% load with the default settings, it really isn't a fair comparison. Also, unless both are running the same WU, it isn't a fair comparison either since the WUs vary a little in length.


the gpu is everytime far above 95% because two wu running @same time and so the utilization for the gpu is about 97-99%. that's point 1. i see this the whole time i look on my desktop, thanks to the gpuobserver gadget

point 2 is, as everyone can see or not, don't know, that the time is shorter on cuda with a MASSIVE improvement on cpu-utilization @cuda. that's a fact.

actually two cudas parallel with 1 cpu-core each need about 3000-3500s with under 100s cpu for finalisation and two opencl ones parallel in the past were about 4000s with 4000s cpu also, still with a whole cpu core each.

i don't have to know about optimizations in the code with newer driver or the lenght of the wu. it is still a fact, that cuda is on my system faster und doesn't need the cpu anymore. thats the point i wanna told you...

further optimizations for collatz would be nice, but i need the system for the daily freetime and with those lines from above in the configfiles, the whole system is horrible laggy, so it's absolutely not useable on this system. i tested a bit around, but still with

threads=6
items_per_kernel=20
kernels_per_reduction=6

and even there were skipped frames in the videos. then cleared the config file and everything runs fine again. you can read this at the optimize thread: http://boinc.thesonntags.com/collatz/forum_thread.php?id=1117&postid=18623

Profile Slicker
Volunteer moderator
Project administrator
Project developer
Project tester
Project scientist
Avatar
Send message
Joined: 11 Jun 09
Posts: 2525
Credit: 740,580,099
RAC: 1
Message 18694 - Posted: 21 Feb 2014, 3:43:45 UTC - in response to Message 18686.

I knew the reduced CPU utilization for the CUDA applications would be a major reason many would prefer it to the OpenCL application.

My point is that as far as overall speed, you are comparing the CUDA app which is using:

threads=6
items_per_kernel=17
kernels_per_reduction=6
sleep=1

to the OpenCL app which is using:

kernels_per_reduction=5
items_per_kernel=16
kernels_per_reduction=6
sleep=1

I have no doubt the CUDA app is faster. But, with different config settings, it really isn't fair to compare 2K seconds on the CUDA app to 3K seconds on the OpenCL app. If you changed the OpenCL app to 6/17/6 it would be faster. But, it would still use more CPU. Since I like to run CPU apps as well, I agree with you that the CUDA app is better. Those who don't run CPU apps and find the optimal settings for OpenCL and run multiple apps on the same GPU at one via an app_config may find it runs a little faster. Or not. Since everyone can run whichever they choose now, it doesn't really matter.

Profile eXtreme Warhead
Avatar
Send message
Joined: 18 Nov 12
Posts: 15
Credit: 229,064,135
RAC: 82,800
Message 18696 - Posted: 21 Feb 2014, 17:48:08 UTC - in response to Message 18694.

compare the cuda and the opencl app at this time is fair, because, and that's the point: it gives the same amount of points. so for me cuda is much better

If you changed the OpenCL app to 6/17/6 it would be faster. But, it would still use more CPU.
how should this works with "more cpu" if the cpu is @opencl at a whole cpu-core for a wu and the app can't use two cores for one wu at a time?

Profile Slicker
Volunteer moderator
Project administrator
Project developer
Project tester
Project scientist
Avatar
Send message
Joined: 11 Jun 09
Posts: 2525
Credit: 740,580,099
RAC: 1
Message 18697 - Posted: 21 Feb 2014, 18:28:03 UTC - in response to Message 18696.

compare the cuda and the opencl app at this time is fair, because, and that's the point: it gives the same amount of points. so for me cuda is much better

If you changed the OpenCL app to 6/17/6 it would be faster. But, it would still use more CPU.
how should this works with "more cpu" if the cpu is @opencl at a whole cpu-core for a wu and the app can't use two cores for one wu at a time?


You are right. It would not work. To get it to work you would have to disable the OpenCL for CPU apps in the preferences so you would only be sent single core non-OpenCL workunits.

I hope that the Khronos group which decided on OpenCL standards will add some type of control over what percent or how many processorscan be used at a time. At present, there is no control with OpenCL. It is all or nothing.


Post to thread

Message boards : News : Have your cake and eat it too!


Main page · Your account · Message boards


Copyright © 2018 Jon Sonntag; All rights reserved.