Credits and WU Fetch Management of OpenCL vs non-OpenCL CPU tasks
log in

Advanced search

Message boards : Number crunching : Credits and WU Fetch Management of OpenCL vs non-OpenCL CPU tasks

Author Message
Jon Fox
Send message
Joined: 6 Sep 09
Posts: 36
Credit: 350,982,829
RAC: 255,926
Message 20305 - Posted: 24 Feb 2015, 18:12:12 UTC

Configuration: 2014 iMac 21"/OS X 10.10/BOINC Client 7.4.36/Nvdia GT 750M

Recently, I've seen a significant increase in the number of OpenCL (multi-threaded) CPU WUs to the exclusion of non-OpenCL (single-threaded) CPU WUs. I've also noticed that the credit earned, on a "credit per CPU second" basis, of the OpenCL CPU WUs is approximately 70% that of a non-OpenCL CPU WU.

As these WUs run longer on a per credit basis, I'm considering deselecting the plan class. I would also need to manipulate the Venue/Location preferences so as to allow other computers in my group to continue to receive this type of WU.

So before proceeding down that path, I thought I would raise the question of "is the disparity between OpenCL and non-OpenCL for CPU WUs intended or consequential?

Thanks in advance.

--
Jon

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: 2
Message 20306 - Posted: 25 Feb 2015, 19:47:27 UTC - in response to Message 20305.

Configuration: 2014 iMac 21"/OS X 10.10/BOINC Client 7.4.36/Nvdia GT 750M

Recently, I've seen a significant increase in the number of OpenCL (multi-threaded) CPU WUs to the exclusion of non-OpenCL (single-threaded) CPU WUs. I've also noticed that the credit earned, on a "credit per CPU second" basis, of the OpenCL CPU WUs is approximately 70% that of a non-OpenCL CPU WU.

As these WUs run longer on a per credit basis, I'm considering deselecting the plan class. I would also need to manipulate the Venue/Location preferences so as to allow other computers in my group to continue to receive this type of WU.

So before proceeding down that path, I thought I would raise the question of "is the disparity between OpenCL and non-OpenCL for CPU WUs intended or consequential?

Thanks in advance.

--
Jon


It isn't intentionally slower, but that's what happens when you add OpenCL and multi-threading on top of an already highly optimized app. Contrary to popular opinion, OpenCL performance kind of sucks. I developed an OpenMP version which was around 97% of the performance but even there, the overhead needed to manage multiple threads/processes uses up precious CPU time.

Feel free to disable it. I'd be tempted to just disable it altogether but someone will be sure to complain that it is no longer available even though it is slower. ;-)

Jon Fox
Send message
Joined: 6 Sep 09
Posts: 36
Credit: 350,982,829
RAC: 255,926
Message 20307 - Posted: 25 Feb 2015, 20:44:23 UTC

As always, thanks Slicker for the quick response and background.

Having read several forum threads on the topic of OpenCL vs. CUDA, et.al. I wasn't surprised or overly concerned about the raw performance, I just wanted to make sure that there wasn't anything I was missing on the "credit per" front before taking the rather draconian approach of deselecting the plan class entirely.

Thanks again.

--
Jon


Post to thread

Message boards : Number crunching : Credits and WU Fetch Management of OpenCL vs non-OpenCL CPU tasks


Main page · Your account · Message boards


Copyright © 2018 Jon Sonntag; All rights reserved.