CUDA vs nVidia OpenCL User Poll
log in

Advanced search

Message boards : News : CUDA vs nVidia OpenCL User Poll

1 · 2 · 3 · 4 · Next
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 18534 - Posted: 14 Feb 2014, 16:50:57 UTC

Testing is finishing up on the Collatz v6.xx application for CUDA. Before I release it, I'd like your opinion on how I should configure the scheduler on the server. Should it:

a) Send the CUDA if available and only send an nVidia OpenCL version if there is no CUDA version for the platform

b) Send the nVidia OpenCL version if available and only send the CUDA version if there is no nVidia OpenCL version for the platform

c) Let the BOINC client learn which is faster over time

Note: The CUDA application uses less than 1% CPU whereas the OpenCL versions use considerably more - sometimes up to 80%.

genacide
Send message
Joined: 26 Jan 10
Posts: 1
Credit: 86,952,835
RAC: 0
Message 18535 - Posted: 14 Feb 2014, 17:17:00 UTC

id say a

Michael Radix Germany
Send message
Joined: 26 Nov 12
Posts: 2
Credit: 317,127,189
RAC: 22
Message 18536 - Posted: 14 Feb 2014, 17:18:28 UTC - in response to Message 18534.

a ! :-))

Send the CUDA if available and only send an nVidia
OpenCL version if there is no CUDA version for the platform

- The CUDA application uses less than 1% CPU -

Profile Overtonesinger
Avatar
Send message
Joined: 16 Jul 10
Posts: 21
Credit: 142,646,892
RAC: 1,941
Message 18537 - Posted: 14 Feb 2014, 17:29:49 UTC - in response to Message 18536.

Exactly!

And the CUDA version will probably be faster... because CUDA is far less complicated and far more tuned and better scalable on NVidias... It should run faster at 90 percent of cards. Well theoretically :-)

I vote for a)
____________
Melwen - child of the Fangorn Forest

Profile DoctorNow
Avatar
Send message
Joined: 12 Jul 09
Posts: 30
Credit: 102,805,175
RAC: 0
Message 18538 - Posted: 14 Feb 2014, 17:32:30 UTC - in response to Message 18534.

Yeah, a) of course. Using up almost a complete core is not good for the other applications which are running. ;-)
____________
Life is Science, and Science rules. To the universe and beyond
Proud member of BOINC@Heidelberg
My BOINC-Stats

bero70
Send message
Joined: 23 Sep 09
Posts: 1
Credit: 2,329,975
RAC: 1
Message 18539 - Posted: 14 Feb 2014, 17:35:23 UTC
Last modified: 14 Feb 2014, 17:35:56 UTC

Because of the low CPU use, I would prefer a).
So another work is able to run.
____________

Profile arkayn
Volunteer tester
Avatar
Send message
Joined: 30 Aug 09
Posts: 219
Credit: 676,877,192
RAC: 17,625
Message 18542 - Posted: 14 Feb 2014, 17:44:30 UTC

A, I am seeing almost 100% CPU time on my 3 completed WU's.

You might look into using the "- use_sleep" command line option, it does a fairly good job of dropping the CPU usage.
____________

Profile [SG] Thomas Kleinhenn
Send message
Joined: 24 May 10
Posts: 1
Credit: 21,376,227
RAC: 0
Message 18544 - Posted: 14 Feb 2014, 18:00:29 UTC

A) cpu time almost 100% here too. so bring on CUDÁ.

Rayburner
Send message
Joined: 3 Aug 09
Posts: 1
Credit: 1,465,178,123
RAC: 0
Message 18545 - Posted: 14 Feb 2014, 18:04:23 UTC

a) !!

Mike
Send message
Joined: 18 Sep 13
Posts: 17
Credit: 33,587,437
RAC: 0
Message 18546 - Posted: 14 Feb 2014, 18:10:16 UTC - in response to Message 18545.

Definitely A.

Profile entigy
Send message
Joined: 1 Jul 10
Posts: 11
Credit: 155,421,783
RAC: 234,516
Message 18547 - Posted: 14 Feb 2014, 18:23:40 UTC

I find CUDA to be significantly faster on my NVIDIA card than OpenCL.

So, a) for me.

Cheers.

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 18549 - Posted: 14 Feb 2014, 19:22:23 UTC - in response to Message 18542.

A, I am seeing almost 100% CPU time on my 3 completed WU's.

You might look into using the "- use_sleep" command line option, it does a fairly good job of dropping the CPU usage.


The CUDA version already yields the CPU (unlike the BOINC samples). In addition, it calls the actual sleep function while waiting for the cudaEvent to complete.

If you know of a way to force an opencl app to yield, please let me know.

Majestix
Send message
Joined: 7 Jun 10
Posts: 7
Credit: 17,923,550
RAC: 0
Message 18551 - Posted: 14 Feb 2014, 20:20:30 UTC - in response to Message 18534.

d - let boinc (users) learn to chose the best for most credits :-)

greetings (and willing to be a candidate for c)

Profile eXtreme Warhead
Avatar
Send message
Joined: 18 Nov 12
Posts: 15
Credit: 229,064,135
RAC: 100,934
Message 18555 - Posted: 14 Feb 2014, 21:11:47 UTC

definitvely prefer cuda!

Profile robertmiles
Send message
Joined: 8 Oct 09
Posts: 45
Credit: 15,071,562
RAC: 46,008
Message 18557 - Posted: 14 Feb 2014, 21:33:46 UTC

I'd prefer c, unless you can add an option to let the more experienced users override what it would normally decide to do.

I've seen another BOINC project say that their CUDA application is faster on Nvidia GTX 5xx and older, but their OpenCL application is faster for Nvidia GTX 6xx and newer.

Eligor85
Send message
Joined: 3 May 11
Posts: 1
Credit: 1,625,191
RAC: 0
Message 18559 - Posted: 14 Feb 2014, 22:48:48 UTC

I say "c" I say "c" so to permit eventually parallel WU of other projects that doesn't use cuda and/or OpenGLso to permit eventally parallel WU.

Mark West
Send message
Joined: 7 Oct 12
Posts: 1
Credit: 338,035,978
RAC: 2,081,047
Message 18560 - Posted: 14 Feb 2014, 23:00:40 UTC

While it seems that a would be fastest, I'd be fine with c. Let the system learn which is faster.

Michael Radix Germany
Send message
Joined: 26 Nov 12
Posts: 2
Credit: 317,127,189
RAC: 22
Message 18561 - Posted: 14 Feb 2014, 23:08:05 UTC
Last modified: 14 Feb 2014, 23:08:49 UTC

It would be nice even if the user's settings in the Collatz Conjecture
could select what should be used for apps.

Dirk Broer
Send message
Joined: 20 Aug 10
Posts: 33
Credit: 176,957,861
RAC: 578,680
Message 18562 - Posted: 14 Feb 2014, 23:54:43 UTC

A, because I run some 45 other projects as well and need the resources
____________

Country Bumkin
Send message
Joined: 29 Oct 10
Posts: 4
Credit: 41,632,851
RAC: 0
Message 18563 - Posted: 15 Feb 2014, 0:09:11 UTC

I am not worried what the automatic option is, just make sure there is a manual selection procedure for at least Linux. Manual tuning also made a 30%+ difference in speed on my system for both NV and AMD cards.

regards CB

1 · 2 · 3 · 4 · Next

Message boards : News : CUDA vs nVidia OpenCL User Poll


Main page · Your account · Message boards


Copyright © 2018 Jon Sonntag; All rights reserved.