WARNING: This website is obsolete! Please follow this link to get to the new Albert@Home website!

[New release] BRP app v1.28 feedback thread

log in

Advanced search

Message boards : Problems and Bug Reports : [New release] BRP app v1.28 feedback thread

1 · 2 · Next
Author Message
Profile Bikeman (Heinz-Bernd Eggenstein)
Volunteer moderator
Project administrator
Project developer
Avatar
Send message
Joined: 28 Aug 06
Posts: 1483
Credit: 1,864,017
RAC: 0
Message 112164 - Posted: 8 Aug 2012, 17:37:23 UTC

We just released the first app versions of a new BRP4 v1.28 app release. Over the next days, we will release this app for all supported platforms, today we released

- OpenCL (ATI/AMD), 32 bit, for Linux,
- OpenCL (ATI/AMD), 32 bit, Windows and
- OpenCL (ATI/AMD) OSX (Lion and later).

All versions require BOINC 7.0.25 or later, but we encourage volunteers to actually use the latest recommended BOINC versions, currently 7.0.28 or 7.0.31.

Versions for CPU and NVIDIA CUDA GPU will follow soon, plus some app versions that are native 64 bit apps.

New in this app version:
- Various performance improvements
- cross-validation improvements for HD 69xx GPUs

Cheers
HBE
____________

Profile zombie67 [MM]
Avatar
Send message
Joined: 10 Oct 06
Posts: 130
Credit: 30,230,982
RAC: 6
Message 112166 - Posted: 9 Aug 2012, 2:36:50 UTC

Hmm. My ATI OSX/ML machine can't get work. 7.0.31.
____________
Dublin, California
Team: SETI.USA

Profile Bikeman (Heinz-Bernd Eggenstein)
Volunteer moderator
Project administrator
Project developer
Avatar
Send message
Joined: 28 Aug 06
Posts: 1483
Credit: 1,864,017
RAC: 0
Message 112168 - Posted: 9 Aug 2012, 8:25:24 UTC - in response to Message 112166.

Hmm. My ATI OSX/ML machine can't get work. 7.0.31.


We had a temporary problem on Albert, BRP4 work should now be available again.

Cheers
HBE
____________

Profile zombie67 [MM]
Avatar
Send message
Joined: 10 Oct 06
Posts: 130
Credit: 30,230,982
RAC: 6
Message 112172 - Posted: 9 Aug 2012, 15:32:03 UTC

Got some. Thanks!
____________
Dublin, California
Team: SETI.USA

TRuEQ & TuVaLu
Send message
Joined: 11 Sep 06
Posts: 75
Credit: 615,315
RAC: 0
Message 112173 - Posted: 12 Aug 2012, 17:23:34 UTC

I did 1 task.

http://albert.phys.uwm.edu/results.php?hostid=3755

Was faster then the older one.

Good work.

Profile Stephan Goll
Send message
Joined: 13 Dec 05
Posts: 19
Credit: 1,159,367
RAC: 0
Message 112174 - Posted: 12 Aug 2012, 22:00:00 UTC

http://albert.phys.uwm.edu/workunit.php?wuid=98764
Looking good so far. Great work.
Stephan


____________

Petrion
Send message
Joined: 30 Apr 08
Posts: 1
Credit: 96,090
RAC: 0
Message 112175 - Posted: 12 Aug 2012, 22:21:16 UTC
Last modified: 12 Aug 2012, 22:24:26 UTC

Crunched roughly 50 so far with no problems; using an HD 6850 with Boinc v.7.28 and OC'd i5 2500K (4.5GHz).

Time reduced a lot from average 3,800secs to 2,100secs doing 1xWU at a time...now if you could reduce the CPU Test WUs from 16,000+ secs I wouldn't mind. :)

Profile Gary Roberts
Volunteer moderator
Send message
Joined: 9 Feb 05
Posts: 1685
Credit: 85,000
RAC: 0
Message 112176 - Posted: 13 Aug 2012, 0:27:09 UTC - in response to Message 112164.


- OpenCL (ATI/AMD) OSX (Lion and later).

I have two pretty much identical hosts - 1859 and 1868 that now have this app. They are both running the latest BOINC (7.0.31) and both have AMD 5750 1GB GPUs. The only hardware difference is that the first one has 4GB RAM whilst the second has 8GB.

The one with smaller RAM has run times around 7Ksecs whilst the other is taking 18Ksecs????

Both machines run EAH tasks on all 4 CPUs and both are doing S6LV1 tasks in aroung 17-18Ksecs. All figures quoted were for tasks crunched over the weekend when there was no user sitting at the keyboard. These machines are used during office hours on weekdays. When idle, both machines power down the screen quite quickly (maybe 3 mins) so the only thing I can now think of is whether or not the screensaver process is still running when the screen turns off. I'll have to check that next time I have access to those machines.

Anyone got any other suggestions? At some point I will release a CPU core from crunching to see what difference that makes and later I may attempt to run two tasks simultaneously - maybe not when you think of the extra heat that might be produced. The aluminium cases are running pretty hot already :-).

____________
Cheers,
Gary.

Profile Bikeman (Heinz-Bernd Eggenstein)
Volunteer moderator
Project administrator
Project developer
Avatar
Send message
Joined: 28 Aug 06
Posts: 1483
Credit: 1,864,017
RAC: 0
Message 112177 - Posted: 13 Aug 2012, 10:16:51 UTC

Hi Gary!

Freeing a CPU core is probably a good thing to try.

Another idea: What I'm always a bit suspicious about (BOINC-performance-wise) is OSX's power saving feature to automatically switch between a CPU build-in GPU (e.g. the Intel stuff in Sandy/Ivy Bridge CPUs) and the dedicated graphics card. Could this be the case that this feature is set differently on those two hosts perhaps?

Cheers
HB

____________

Alex
Send message
Joined: 1 Mar 05
Posts: 88
Credit: 398,734
RAC: 0
Message 112178 - Posted: 13 Aug 2012, 21:37:24 UTC

Congratulations!

My pc did a BRP 1.28 in ~1980 sec. HD6900 Series GPU, i7 with free cpu, 7.0.31, CCC 12.6
On HD5850 a wu takes ~3450sec.
AMD-apps are grown up now.
____________

Alex
Send message
Joined: 1 Mar 05
Posts: 88
Credit: 398,734
RAC: 0
Message 112182 - Posted: 14 Aug 2012, 11:09:38 UTC
Last modified: 14 Aug 2012, 11:13:52 UTC

Phantastic!



The amd is a HD6950 running a single wu @ 82% gpu load, the nVidia is a GTX680, unknown settings.

edit: uups, picture is missing...


wu 997334
task 279100 runtime 1,981.83 Binary Radio Pulsar Search v1.28 (opencl-ati)
task 279099 runtime 11,162.70 Binary Radio Pulsar Search v1.25 (BRP3cuda32)
____________

Profile Bikeman (Heinz-Bernd Eggenstein)
Volunteer moderator
Project administrator
Project developer
Avatar
Send message
Joined: 28 Aug 06
Posts: 1483
Credit: 1,864,017
RAC: 0
Message 112183 - Posted: 14 Aug 2012, 12:34:08 UTC - in response to Message 112178.
Last modified: 14 Aug 2012, 12:43:56 UTC

Congratulations!

My pc did a BRP 1.28 in ~1980 sec. HD6900 Series GPU, i7 with free cpu, 7.0.31, CCC 12.6
On HD5850 a wu takes ~3450sec.
AMD-apps are grown up now.


Hi,

actually the app should do even faster, did you free a CPU core on the machine with the HD 5850 as well? I have a HD 5850 that is doing a single unit in ca 1800 sec (!!) (although that is on a comparatively fast dual core 3 GHZ Core2 (no hyperthreading!) with only one other task running on the CPU.

Cheers
HB

EDIT: Ah I see....2 cards in one box I guess. That is distorting the numbers a bit of course, but I'm glad to see that you see a substantial performance increase.

HB
____________

Alex
Send message
Joined: 1 Mar 05
Posts: 88
Credit: 398,734
RAC: 0
Message 112184 - Posted: 14 Aug 2012, 13:21:17 UTC - in response to Message 112183.



Hi,

actually the app should do even faster, did you free a CPU core on the machine with the HD 5850 as well? I have a HD 5850 that is doing a single unit in ca 1800 sec (!!) (although that is on a comparatively fast dual core 3 GHZ Core2 (no hyperthreading!) with only one other task running on the CPU.

Cheers
HB





OK, its the same machine which has both gpu's installed. CPU is a i7, it's a machine for normal daily business (office and controlsystem design), which sometimes runs 100% for BOINC, actually exclusively for Albert (I'm working here right now).
CPU-load for all 4 cores/8 threads are all well below 30%.
The 'smaller' gpu is in an x8 pcie-slot, maybe this also makes a difference.
And since two generations of gpu's are installed, I never can be shure that I have the optimal drivers installed.
Anyway, I'm happy that the amd gpu's are much more usable now for DC jobs than in earlier days when just Milkyway had a pretty fast amd app.

Alexander

____________

Profile Steve Hawker*
Send message
Joined: 22 Jun 12
Posts: 1
Credit: 1,064,230
RAC: 0
Message 112185 - Posted: 14 Aug 2012, 18:14:29 UTC - in response to Message 112184.

Seeing as there are few projects that will crunch WUs on my MacBook's GPU, I'm always interested to see how it pans on a new release.

1. the time estimates are way off. initial estimate was >50 hours. task crunched in about 7 hours. IIRC this is down from 12 hours.
2. the CPU utilization was 0.564 of a core. I think this is high (Collatz grabs just 0.02 of a core). Even so I think this is a reduction
3. i still only got 500 credits. for 7 hours crunching Collatz gives me ~6,500

I'm not whining about the credits - I read, understood and accepted the notice. But this is a feedback thread after all.

I enjoy the credit game as much as the next cruncher but I don't always pick the highest paying project. But i'd like a decent whack even so. What i'd like to know is if and when these apps move to production, what will be the going rate? I have a very slight preference for astronomy projects over mathematical projects but 500 credits won't get my GPU. I know, I'm just one and there's plenty more dedicated fish in the sea. nonetheless, this is my feedback.

I'll keep coming back and crunching a WU each time there's a new app. After all, you went to the trouble of making an OSX app, the least I can do is test it.

Thanks!

Alex
Send message
Joined: 1 Mar 05
Posts: 88
Credit: 398,734
RAC: 0
Message 112186 - Posted: 14 Aug 2012, 19:22:14 UTC

I could not resist ..

I 'borrowed' my old testsystem and tried some wu's.
AMD A8 3870 APU: wu 100169 runtime 6,489.60
HD5830 runtime ~ 2,915.96
____________

Profile Bikeman (Heinz-Bernd Eggenstein)
Volunteer moderator
Project administrator
Project developer
Avatar
Send message
Joined: 28 Aug 06
Posts: 1483
Credit: 1,864,017
RAC: 0
Message 112187 - Posted: 15 Aug 2012, 7:45:29 UTC

Hi!

Thanks for the feedback!

As for credits, our yardstick used to be SETI@Home, so we tuned the credits so that the CPU apps would grant roughly as many credits as the SETI@HOME CPU apps on the same hardware in the the same time. The GPU apps would grant as many credits per UNIT as the CPU apps (that is: all results are treated equally, whether done on CPU or GPU). With SETI@Home now using the "Credit New" BOINC crediting system, things might have diverged a bit. Still my feeling is that E@H will probably try in the future to stay in line with the bigger (in terms of nr of participants) "mainstream" projects and will not try to "out-credit" all the other projects :-). Just my personal idea, tho.

Cheers
HB

____________

Profile Bikeman (Heinz-Bernd Eggenstein)
Volunteer moderator
Project administrator
Project developer
Avatar
Send message
Joined: 28 Aug 06
Posts: 1483
Credit: 1,864,017
RAC: 0
Message 112191 - Posted: 21 Aug 2012, 14:14:10 UTC

Thanks all for testing, the OpenCL app versions are now released on Einstein@Home.

Next we will release the CUDA apps based on the same codebase to Albert@Home. The perfmance increase will not be as high as for the OpenCL versions which still had to to a bit of catching up with the CUDA versions, so we should now have CUDA and OpenCL apps that are on par, performance-wise, so to speak.


CU

HBE
____________

Vlatko
Send message
Joined: 9 Aug 12
Posts: 4
Credit: 149,500
RAC: 0
Message 112192 - Posted: 23 Aug 2012, 19:10:18 UTC

Are the CUDA version released?
I am seeing the BRP4 v1.28 version is running on my BOINC menager

Profile Bikeman (Heinz-Bernd Eggenstein)
Volunteer moderator
Project administrator
Project developer
Avatar
Send message
Joined: 28 Aug 06
Posts: 1483
Credit: 1,864,017
RAC: 0
Message 112193 - Posted: 24 Aug 2012, 9:26:39 UTC - in response to Message 112192.

Are the CUDA version released?
I am seeing the BRP4 v1.28 version is running on my BOINC menager


Indeed, the Windows CUDA app is out now, I'll launch the Linux and Mac OSX versions today and then write an announcement.

From what I've seen so far the CUDA Windows app should reduce the runtime of a BRP4 task in a way that scales more with the CPU speed than with GPU parameters, e.g. on modern CPUs (nd if running just one unit at a time), the overall runtime should be reduced by roughly 400-700 seconds. For slower cards that take 2 h per unit, this doesn't amount to much, but for faster cards that take less than half an hour, this should be a substantial speed-up. Any feedback on speed-up in different configurations (2 tasks t time, multiple cards,...) are welcome.

Cheers
HB

____________

Jeroen
Send message
Joined: 25 Nov 05
Posts: 12
Credit: 638,256
RAC: 0
Message 112194 - Posted: 24 Aug 2012, 14:25:28 UTC - in response to Message 112193.

I am looking forward to testing out BRP4 v1.28 for Linux. Thanks for the updates!
____________

1 · 2 · Next

Message boards : Problems and Bug Reports : [New release] BRP app v1.28 feedback thread


Home · Your account · Message boards

This material is based upon work supported by the National Science Foundation (NSF) under Grant PHY-0555655 and by the Max Planck Gesellschaft (MPG). Any opinions, findings, and conclusions or recommendations expressed in this material are those of the investigators and do not necessarily reflect the views of the NSF or the MPG.

Copyright © 2016 Bruce Allen for the LIGO Scientific Collaboration