Einstein@Home GPU Application for ATI/AMD Graphics Cards

[PoppaGeek]

After more than a year of work by Oliver Bock, Bernd Machenschalk, Heinz-Bernd Eggenstein and other developers, we are pleased to announce the release of the first Einstein@Home application for ATI/AMD Graphics Cards.

This OpenCL application, which searches Arecibo data for new radio pulsars, is about a factor of ten faster than the same search running on a typical CPU. The application is currently available for Windows and Linux computers with Radeon HD 5000 or better graphics cards. We hope to have a version for Macintosh (Apple OS X 10.8, Mountain Lion) sometime this summer, but there are still some problems that need to be fixed or worked around.

Volunteers who wish to run this application will need to install version 7.0.27 or later of the BOINC client. Please see this thread for more information, or if you want to ask questions.

Many thanks to the AMD/ATI team for their support in the OpenCL software development effort.

Bruce Allen
Director, Einstein@Home

[PoppaGeek]

Einstein@Home GPU/APU Application for AMD/ATI Graphics Cards: discussion thread

Hi everyone,

We're proud to announce the official release of our first application for AMD/ATI graphics cards (GPU) and accelerated processing units (APU). Please use this thread to discuss this release.

Minimum Requirements:

Windows or Linux
BOINC Client 7.0.27 (download)
ATI GPU or APU (OpenCL 1.1 compliant, equivalent to Radeon HD 5xxx)
512 MB video memory
Catalyst Driver 12.x (don't install the APP SDK!)

Notes:

Please be aware that all previous tests we ran could only do a limited amount of testing given the vast amount of different hard- and software run by our volunteers. There might still be smaller issues, in particular with validation. Please bear with us, we'll do our best to improve the application over the next weeks and months.
The same applies to application performance. This is our first official release. It might not be on par with our CUDA application just yet. We still have ideas on how to improve application performance and we're going to introduce them in due course.
Tip: performance might be improved if you set BOINC not to use all your CPU cores (e.g. all but one).
Support for Apple Mac OS X is currently targeted for OS X Mountain Lion (10.8)

Known Issues:

When running this application (all 32-bit) on 64-bit systems you might encounter the following error message (error number 255/-1):
[ERROR] Failed to get OpenCL platform/device info from BOINC (error: -1)
If that happens, please download the latest AMD/ATI Catalyst driver, reinstall it and reboot your computer.

[Snow Crash]

XS has been doinig quite well on Einstein lately ... climbing the project charts pretty quickly
We're on target to hit 100 MILLION team points very very soon!

Looks like PG decided to rejoin the team
Of course if you think I'll let you keep taking daily #1 pie from me ... think again
Things WILL turn around this weekend

[PoppaGeek]

Eh, took ya 3 days to notice. Was hoping to get a few more days in.

[PoppaGeek]

Grats on the 11M stone!



I be eating my pie and watching for you.

[PoppaGeek]

XS E@H hits 95,000,000!



1 Billion here we come!

[Snow Crash]

Nice pie PG!!!
Ya caught me not looking and made me pay for it
... not going to happen again ... really I mean it ... truly ...

[Snow Crash]

Einstein has an Intel iGPU app also ... I'm getting a bit more than 8500 ppd on my 4670K without it being overclocked (yet) :yup:

[PoppaGeek]

OH NOES!! Wut does it mean??


Feel the Pressure.....

[Snow Crash]

pressure ... why yes ... some good fun ... let's see how far you can push me
Doesn't mean I'm intimidated, nor does it mean I don't have anything else to bring to the table

[PoppaGeek]

pressure ... why yes ... some good fun ... let's see how far you can push me
Doesn't mean I'm intimidated, nor does it mean I don't have anything else to bring to the table

Now calm down.



Come on big guy, put down the hardware.


We know yer cool bro.

[PoppaGeek]

In moving cards around I discovered just how WCGs Clean Energy Project effects other tasks running. Some machines I have it set to 1 CEP at a time and some 3 and some unlimited. The machines that were running unlimited CEP and E@H had much longer Elapsed time on E@H WUs than the ones with less CEP. It is well known CEP is demanding but I was not expecting the difference I am seeing. I moved a 8800GT from a unlimited CEP to no CPU tasks yet a slower CPU and the Elapsed time has been reduced 6,900 seconds.

[Snow Crash]

I'm putting my NV gear back on GPUGrid for a while, so you have a chance, a small chance to catch me in the dailies ... we'll see how close you get before I add my 7850 back into the mix!

[stoneageman]

Time to catch a wave or three

[Snow Crash]

and glad to have you ... water's warm and especially nice for AMD - my 7950 averages just under 100k ppd (1 card PCIE3x16, OC GPU to 1250) I'm running 2 WU at a time, I tried 3 but there was no PPD difference, the GPU utilization went up but apparently that was just the overhead of managing 3 WUs.

These WUs are bandwidth sensitive, multi-card system running PCIE2 are definitely constrained.

Aricebo WUs are looking at the inner galaxy and are more likely to discover a binary pulsar but get about 10% less point than Perseus WUs that are looking outwards at ... you guessed it the Perseus Arm.

I found that running both types at the same did not perform well on NVidia w/ PCIE2x16

wplachy's in the house and making some serious noise ...

[OldChap]

and glad to have you ... water's warm and especially nice for AMD - my 7950 averages just under 100k ppd (1 card PCIE3x16, OC GPU to 1250) I'm running 2 WU at a time, I tried 3 but there was no PPD difference, the GPU utilization went up but apparently that was just the overhead of managing 3 WUs.

These WUs are bandwidth sensitive, multi-card system running PCIE2 are definitely constrained.

Aricebo WUs are looking at the inner galaxy and are more likely to discover a binary pulsar but get about 10% less point than Perseus WUs that are looking outwards at ... you guessed it the Perseus Arm.

I found that running both types at the same did not perform well on NVidia w/ PCIE2x16

wplachy's in the house and making some serious noise ...

Running 3770K on MVG @3.9 with one card (7950) @1100 now and have 2 threads in use running Win7. The other 6 threads are running a VM also with Boinc but running WCG in linux. It is my daily rig and is not 24/7.

I am not sure yet but runtimes seem to behave differently for me than for SC.

I am going back over some quick tests that I did while it was raining at the weekend. Running Perseus arm survey only The first batch is as follows:

1. running 2 wu's concurrently on 2 threads using app_config
<gpu_usage>0.5</gpu_usage>
<cpu_usage>1.0</cpu_usage>

Completion times: best to worst = 5714-5723 so ave 5719 per 2 wu's or 2859 (47m39s) per wu.


2. running 3 wu's concurrently on 2 threads using app_config
<gpu_usage>0.33</gpu_usage>
<cpu_usage>0.50</cpu_usage>

Completion times: best to worst = 7854-7922-7939 so ave 7905 per 3 wu's or 2635 (43m55s) per wu.


3. running 4 wu's concurrently on 2 threads using app_config
<gpu_usage>0.25</gpu_usage>
<cpu_usage>0.50</cpu_usage>

Completion times: best to worst = 10253-10291-10332-10371 so ave 10312 per 4 wu's or 2578 (42m58s) per wu.


4. running 5 wu's concurrently on 2 threads using app_config
<gpu_usage>0.20</gpu_usage>
<cpu_usage>0.40</cpu_usage>

Completion times: best to worst = 13114-13169-13278-13288-13526 so ave 13275 per 5 wu's or 2655 (44m15s) per wu.

It is not a lot but there is a slight advantage 4 vs 3 as there is running 3 rather than 2 whereas with 5 it is going the wrong way again. I do however wonder if running 5 or 6 wu's concurrently but on 3 threads would still support this trend.

Obviously this is a very small sample and the gains are not exactly huge but if anything at all the above may show that individuals should try similar testing on individual hardware setups if wanting to maximise any points scoring potential.

[stoneageman]

Thanks, very useful info. We are some way down the rankings, so plenty to go for. The validations are taking some time though

[PoppaGeek]

IIRC they allow 2 weeks turn around on WUs. Which works for me because 3 of my crunchers are daily machines and game machines and GPU WUs do not run while games are being played. Another machine is often shut down due to heat, from weather. And they all run all CPU cores on WCG. I never get good numbers on mine. I do thank SC for the mention of mixed WUs on Nvidia, maybe that will help a little.

Team should be climbing fast now.

[Snow Crash]

One of the nice things with the Einstein project is that you can configure how many concurrent GPU tasks to run directly on your preferences page
preferences

"GPU utilization factor of BRP apps" - enter .5 to make your machine run 2 tasks per GPU which is what I found is the best efficiency on 7950, 7850, 660Ti, 670.

Changing that setting does not take place immediately ... not until you download a new task from Einstein.

I usually leave 1 free CPU thread per GPU WU to make sure it is properly fed ... either I set my BOINC preferences to use 1 less thread than I have or I use an app_config file to change the Einstein default of using .5 cpu per GPU to 1.

[OldChap]

....app_config goes where on this?

[PoppaGeek]

<BOINC-DATA-folder>/projects/einstein.phys.uwm.edu

[OldChap]

Thanks... yep got that in but boinc still reporting .5 cpu and 1 gpu for each with the following app_config:

<app_config>
<app>
<name>einsteinbinary_BRP5</name>
<user_friendly_name>Binary Radio Pulsar Search (Perseus Arm Survey)</user_friendly_name>
<gpu_versions>
<gpu_usage>0.5</gpu_usage>
<cpu_usage>1.0</cpu_usage>
</gpu_versions>
</app>
</app_config>

???

[PoppaGeek]

Sometimes it takes Boinc awhile to apply changes in the app_config.xml file. I think it is when new WUs start because restarting Boinc does not usually make the changes. If in Linux make sure ownership and group are correct, mine is boinc:boinc.

sudo chown boinc:boinc /var/lib/projects/einstein.phys.uwm.edu/app_config.xml

I made some changes in mine and restarted it and the Event Log does not show it found an app_config.xml file and no changes are showing in Tasks. But I am used to Boinc ignoring me. I do not seem to make much of an impression on it.

[OldChap]

Most curious. It seemed to take the number of tasks running from 1 to 2 but did not report the cpu use or the gpu use as per the app_config.

Ran all night in this config then crashed during the day (I forgot to leave a window open for cooling). Now the screen shows the same as the app_config.

Just going to run 1 card while FAHV points sort their ideas out on my rigs then "as and when" and not heavily overclocked.

cpu load for each instance seems to be typically just 1-2% and watching I never saw it exceed 8%

currently running (on win7) looks like this:



So, hopefully this will get 2 wu's every 2 hours with this type of work which based on completions so far could be around 3,300 per hour or 75K plus per day

this is 7950 @ 1050 and mem 1250. Not sure what to make of this but as I am just chipping in a little to boost team pts I will leave it at that unless someone here shows me a reason to change things

[stoneageman]