Albert@home is still an unofficial, non-public test project. Don't expect anything to work here. The only type of work Albert@home is currently sending out is for a highly experimental BRP4 OpenCL application.
Copyright © 2024 Einstein@Home. All rights reserved.
Comments
Sending work
)
Great, will expect doom and hell-fire. :D
Ok, so here as well, something breaks the signature. Enigma has the same problem.
RE: Albert@home is still an
)
There is no option in the preferences to select ATI graphic cards so the OpenCL will only be done with Nvidia at the moment.
Conan
Thanks for reporting.
)
Thanks for reporting. Although I think due to the way this setting is handled internally you shouldn't be able to opt-out of ATI 'work' right now.
Anyway, we'll add this.
BM
BM
RE: Ok, so here as well,
)
Honestly I don't care much about sigs at all (I usually view posts w/o sig), and in particular on Albert. This is about the last thing I intend to test here.
The problem may occur because Albert doesn't export any stats, or something else.
Can you edit your preferences and set the signature again from scratch?
BM
BM
Nah, that doesn't fix it but
)
Nah, that doesn't fix it but as you said, it's not a priority. The infinite work duration tasks that are sent... now that's another thing. ;-)
Actually, that is possible with the option in the clients that can do that, and possibly with the option in the latest test clients.
Then it'll just do:
2011-11-01 05:13:40.7956 [PID=8972 ] [version] Checking plan class 'ATIOpenCL'
2011-11-01 05:13:40.7956 [PID=8972 ] [version] No ATI devices found
2011-11-01 05:13:40.7956 [PID=8972 ] [version] [AV#443] app_plan() returned false
2011-11-01 05:13:40.7956 [PID=8972 ] [version] Checking plan class 'NVOpenCL'
2011-11-01 05:13:40.7956 [PID=8972 ] [version] No NVidia devices found
2011-11-01 05:13:40.7956 [PID=8972 ] [version] [AV#440] app_plan() returned false
But really, the BOINC back-end code should self-detect whether an ATI application is available in http://albert.phys.uwm.edu/apps.php and then automatically enable the "Use ATI GPU" preference. Apparently it doesn't work yet with your used version of the back-end and OpenCL applications.
RE: Thanks for reporting.
)
Done.
BM
BM
RE: But really, the BOINC
)
Yah, we didn't look into that clever piece of code when we set up the plan classes. It did only work if your plan class name contained the string "ati" (lowercase).
BM
BM
Nothing but Computation
)
Nothing but Computation errors so far on the OpenCL Wu's after just a few seconds. Tried 11.6 & 11.9 Drivers 5850 & 5870 Cards ...
STE\/E
The Linux App appears to
)
The Linux App appears to work, at least on one of our machines.
I guess this is on Windows, where we are observing the same errors. FWIW the App works when running standalone, it crashes only when running under the BOINC Client.
BM
BM
[20:05:56][5684][INFO ]
)
[20:05:56][5684][INFO ] Starting data processing...
[20:05:56][5684][ERROR] Failed to get OpenCL platform/device info from BOINC (error: -31)!
[20:05:56][5684][ERROR] Demodulation failed (error: -31)!
http://albertathome.org/task/35469
RE: FWIW the App works when
)
How do you run the app stand-alone? It's not an executable.
It's called einsteinbinary_BRP4_1.09_windows_intelx86__ATIOpenCL ... no .exe Could that be the problem?
OK, I got the new AtiOpenCL
)
OK, I got the new AtiOpenCL app 1.11, got a bucket load of work for it as well, but it doesn't start. I even suspended all my other work and still the OpenCL tasks didn't start.
So I have reset my preferences, set them back to use both CPU & GPU. Doesn't matter. I took out the option from cc_config.xml as it clearly didn't work yet, but that doesn't matter either.
Now I am confused. :)
For the doubters:
03/11/2011 23:32:02 | | ATI GPU 0: ATI Radeon HD 4700/4800 (RV740/RV770) (CAL version 1.4.1417, 1024MB, 992MB available, 2000 GFLOPS peak)
03/11/2011 23:32:02 | | OpenCL: ATI GPU 0: ATI RV770 (driver version CAL 1.4.1417, device version OpenCL 1.0 AMD-APP-SDK-v2.5 (775.2), 1024MB)
03/11/2011 23:32:02 | | No NVIDIA library found
03/11/2011 23:32:02 | | ATI GPU is OpenCL-capable
Catalysts 11.6 installed.
Edit: tried running the app stand-alone.
Activated exception handling...
00:24:28 (4688): Can't open init data file - running in standalone mode
00:24:28 (4688): Can't open init data file - running in standalone mode
[00:24:28][4688][WARN ] User/host details unavailable...
00:24:28 (4688): called boinc_finish
successfully received work
)
successfully received work for 2011 iMac/10.7.2 (Darwin 11.2)/ATI Radeon HD 6770M/
Start-up messages:
Wed Nov 2 23:05:55 2011 | | OpenCL: ATI GPU 0: ATI Radeon HD 6770M (driver version 1.0, device version OpenCL 1.1, 512MB)
Wed Nov 2 23:05:55 2011 | | ATI GPU is OpenCL-capable
Work Unit Start messages:
Thu Nov 3 20:27:03 2011 | Albert@Home | Starting task p2030.20100913.G48.73+01.03.S.b5s0g0.00000_1968_3 using einsteinbinary_BRP4 version 110 (ATIOpenCL)
WU started with an estimated remaining time of ~ 1:20:00. It has been processing for 2:10:00 and now estimates 3:46:00 remaining. Status reports: (0.20 CPUs and 1.00 ATI GPUs).
Jon
I'll post updates as processing continues/aborts/etc.
Hmm, OK, so all that work
)
Hmm, OK, so all that work mysteriously ran and erred within seconds. Th error is still the same as with 1.09
My system has gone through several reboots yesterday, prior to it attacking the 1.11 work.
Appear to have completed the
)
Appear to have completed the first WU successfully:
Fri Nov 4 04:59:49 2011 | Albert@Home | Computation for task p2030.20100913.G48.73+01.03.S.b5s0g0.00000_1968_3 finished
~8:20:00 of elapsed time. A second work request was made for ATI work and received 47 WUs (more than I really wanted but I'll let it run out as long as there are no major issues).
Jon
Great! Status
)
Great!
Status update:
Currently the Mac OS ATI App appears to be the only OpenCL App that's working properly:
* Windows OpenCL Apps still crash very early when running under BOINC
* The Linux OpenCL App crashed after 6/8 (on our machine), probably due to a memory access problem
* NVidia OpenCL App run but give way wrong results, possibly originating from the same memory access problem
We've disabled the Apps we know to give unusable results, will push out new Apps when ready for testing.
CPU results are still helpful for comparison.
The "zero flops" / "negative FLOPs left -1.#IND00" issue that occurred for CPU plan-class Apps with recent Clients should be fixed now.
BM
BM
[21:26:46][3292][INFO ]
)
[21:26:46][3292][INFO ] Starting data processing...
GPU type not found in init_data.xml
[21:26:46][3292][ERROR] Failed to get OpenCL platform/device info from BOINC (error: -161)!
[21:26:46][3292][ERROR] Demodulation failed (error: -161)!
[21:26:46][3292][DEBUG] Shutting down BOINC... Bye!
21:26:46 (3292): called boinc_finish
http://albertathome.org/task/37885
I wouldn't expect OpenCL to
)
I wouldn't expect OpenCL to work with clients older than 6.13.
Reminds me: we should probably add a client version restriction for these apps.
BM
BM
I got the infamous "GPU type
)
I got the infamous "GPU type not found in init_data.xml" errror with the recommended BOINC client version 6.12.34. Then I decided to update my BOINC client to version 6.13.10 (pre-release).
Now the "GPU type not found in init_data.xml" error is gone, but the computation ends with an access violation, see here.
RE: I wouldn't expect
)
FYI, look at this: http://setiathome.berkeley.edu/result.php?resultid=2146353363
and all other helpful info about OpenCL detection with BOINC 6.10.60
So? Just because Lunatics
)
So? Just because Lunatics built a 3rd party application for Seti only that does the OpenCL detection, does not mean that the BOINC you're running does the OpenCL detection. This project uses the OpenCL detection built into the 6.13 clients to determine all OpenCL capable co-processors.
Added to that, as far as I know, the application here does not require an SSE3 capable CPU. The Lunatics app does.
10/11/2011 00:32:01 | | ATI GPU 0: ATI Radeon HD 4700/4800 (RV740/RV770) (CAL version 1.4.1417, 1024MB, 992MB available, 2000 GFLOPS peak)
10/11/2011 00:32:01 | | OpenCL: ATI GPU 0: ATI RV770 (driver version CAL 1.4.1417, device version OpenCL 1.0 AMD-APP-SDK-v2.5 (775.2), 1024MB)
10/11/2011 00:32:01 | | ATI GPU is OpenCL-capable
RE: So? Just because
)
Ok. Fine. Clear.
I'll try 6.13 some later, despite the fact that I don't like > 6.10 interface without Messages tab.
Then keep the boincmgr.exe
)
Then keep the boincmgr.exe and boinc.dll from your present installation. You can run a newer client with an older BOINC Manager.
But the messages aren't gone, they are just in a window of their own, reachable by clicking advanced->Event Log, or pressing CTRL+SHIFT+E.
As for 6.13, do know that even the latest client, 6.13.10 is highly unstable. I have managed to get it to crash 3 times already by just running this project. The 6.13s aren't available for long time use, they're test versions on the way to BOINC 7.
FYI, we just updated the AMD
)
FYI, we just updated the AMD OpenCL binaries for Windows, Mac OS 10.7 and Linux (all OpenCL 1.0, minimum BOINC version: 6.13.x).
Current status:
* Windows: finally works, no known issues
* Mac OS X: no known issues (10.7 support only!)
* Linux: might use full CPU core, please confirm
Thanks for testing!
Oliver
RE: But the messages aren't
)
I know, but it's inconvenient to use.
RE: Current status: *
)
Yes, look at that. v1.17 is finally running longer than 4 seconds. :-)
Still... not out of the woods yet. At 90 seconds now I have a computation error.
http://albertathome.org/task/38165
http://albertathome.org/task/38123
And now they all end with error at 43 seconds.
6.13.10
There was an error while deleting the color transform. (0x7e3) - exit code 2019 (0x7e3)
Activated exception handling...
[17:10:04][3420][INFO ] Starting data processing...
[17:10:05][3420][INFO ] Using OpenCL platform provided by: Advanced Micro Devices, Inc.
[17:10:05][3420][INFO ] Using OpenCL device "ATI RV770" by: Advanced Micro Devices, Inc.
[17:10:07][3420][INFO ] Checkpoint file unavailable: status.cpt (No such file or directory).
------> Starting from scratch...
[17:10:07][3420][INFO ] Header contents:
------> Original WAPP file: ./p2030.20100913.G48.73+01.03.S.b6s0g0.00000_DM24.80
------> Sample time in microseconds: 65.4762
------> Observation time in seconds: 274.62705
------> Time stamp (MJD): 55453.031586060431
------> Number of samples/record: 0
------> Center freq in MHz: 1214.289551
------> Channel band in MHz: 0.33605957
------> Number of channels/record: 960
------> Nifs: 1
------> RA (J2000): 191718.8153
------> DEC (J2000): 143053.3507
------> Galactic l: 0
------> Galactic b: 0
------> Name: G48.73+01.03.S
------> Lagformat: 0
------> Sum: 1
------> Level: 3
------> AZ at start: 0
------> ZA at start: 0
------> AST at start: 0
------> LST at start: 0
------> Project ID: --
------> Observers: --
------> File size (bytes): 0
------> Data size (bytes): 0
------> Number of samples: 4194304
------> Trial dispersion measure: 24.8 cm^-3 pc
------> Scale factor: 0.114462
[17:10:17][3420][INFO ] Seed for random number generator is -1018057414.
[17:10:48][3420][INFO ] Derived global search parameters:
------> f_A probability = 0.08
------> single bin prob(P_noise > P_thr) = 9.93986e-009
------> thr1 = 18.4267
------> thr2 = 21.5421
------> thr4 = 26.5915
------> thr8 = 35.0049
------> thr16 = 49.3672
[17:10:49][3420][ERROR] Error during OpenCL kernel setup: TSMR-1 (error: -54)
[17:10:49][3420][ERROR] Demodulation failed (error: 2019)!
17:10:49 (3420): called boinc_finish
]]>
So stop deleting the color transform. ;-)
Task list: http://albert.phys.uwm.edu/results.php?userid=7430
While running that work, GPU load is zero according to GPU-Z:
Thanks Jord! That's really
)
Thanks Jord! That's really helpful.
I need to change some settings as your GPU is not able to handle our current parallel thread block structure. I need to check the specs of AMD's GPUs again...
I'll deploy an updated binary as soon as possible.
Cheers,
Oliver
I rebooted, just to make sure
)
I rebooted, just to make sure it wasn't something stuck in my GPU.
After reboot, I first started GPU-Z, then BOINC. No load on the GPU. Tasks don't run longer than a minute so far.
In comparison, running Furmark will do this:
See the load?
(And yes, I also see that the fan is either not recognized or not working...)
The same result as Ageless
)
The same result as Ageless already has showed:
http://albertathome.org/task/38235
RE: The same result as
)
Yep, you got the same device series (HD 4xxx) as Jord. The issue is understood and I'm working on a fix, but it's not that easy...
You guys should opt-out of GPU tasks until the next version.
Hang in there,
Oliver
After Upgrading to v6.13.10
)
After Upgrading to v6.13.10 on my ATI Box I now have 2 Wu's running, 1 on a 5870 & 1 on a 5850 ... Very inefficient though, using only 29% & 25% Of the GPU ... Need a App File to possibly run 2 or maybe 3 @ a time if possible ...
STE\/E
RE: FYI, we just updated
)
Oliver, Mac OS X 10.7 does OpenCL 1.1 already, for OpenCL 1.0 OS X 10.6 is also fine.
Greetings Running a Win 7
)
Greetings
Running a Win 7 x64 box with an HD 5850 on Catalyst 11.9, Boinc 6.13.10 x64, rxing ATI OpenCL work (done so via my preferences), completed my first work unit
http://albertathome.org/task/38178 (Computer ID 1323) Pending
So far no errors, have still 47 tasks that I am working my way through,
Am glad to jump onboard and help out
Keep the work coming and Happy Crunching
Regards
Tazzduke
RE: Oliver, Mac OS X 10.7
)
I know.
a) We have an OpenCL 1.1 version, but it would constrain the list of compatible GPUs.
b) 10.6 is buggy and Apple doesn't care about it, hence we can't support it.
Oliver
Hi guys, RE: Yep, you got
)
Hi guys,
We deployed updated binaries (v1.19) that should adjust dynamically to the specific hardware they run on. This should fix the issues you've encountered and ensure support for lower-end GPUs like the HD 4xxx series.
*Finger's crossed*
Oliver
RE: *Finger's crossed* Must
)
Must be difficult to type with crossed fingers. I just tried it. So please uncross. It seems to be working. ;)
GPU-Z:
Though, can you explain to me why the OpenCL BRP4 app runs at below normal priority?
Not yet. What's and set to
)
Not yet.
What's and set to in sched_reply?
BM
BM
0.150000 1.000000 The
)
0.150000
1.000000
The task is running, but going slowly. At present rate, with it taking 2 hours to do 22%, it'll take ~9 hours to do the whole 100%. A tad slow, when compared to a similar task at Einstein done on my CPU takes ~10 hours.
I've set BOINC to use 3 CPUs, leaving one open for the GPU and everything else non-BOINC.
RE: The task is running,
)
You're mistaken, these are not similar tasks. If you want to compare GPU vs CPU performance you need to crunch a few of the tasks over here on your CPU. Also, the actual GPU performance varies quite a bit from the Radeon HD 4xxx to the 5xxx and the 6xxx series.
Some approx. runtime numbers:
HD 6970 (Windows): 1 hour
HD 6970 (Linux, no PGO): 1.2 hours
HD 5770 (Mac OS 10.7): 7.2 hours
You see that your 4xxx GPU blends in just fine...
Best,
Oliver
Current status as of v1.19: *
)
Current status as of v1.19:
* Windows: no known issues
* Mac OS X: no known issues
* Linux: might use full CPU core (known bug/feature in AMD Catalyst drivers)
Requirements:
* AMD GPU (NVIDIA will be supported by our CUDA apps for the time being)
* OpenCL 1.0
* BOINC: at least 6.13.10
* Mac OS: 10.7 only
Oliver
RE: Though, can you explain
)
Because it uses less than one CPU (avg_ncpus < 1).
If it would use a full CPU core (like CPU Apps do), it would run in "Idle" priority.
BM
BM
RE: RE: So? Just because
)
You are right.
The message tab is annoying.
I am now using ATI5850
)
I am now using ATI5850 Driver11.9 and SDK2.5
Boinc version 6.13.12 and app 1.19
Progress 38% in 4hours ; hopefully completed in the morning...
GPU load is rarely above 0% sometimes it uses 100% for half a second
Running on win Vista 32bit
Now I will sleep. Goodnite all
Could you mention a little
)
Could you mention a little more about what GPUs are suitable for the current A@H application? What I've seen so far suggests that it may require AMD/ATI GPUs, and is not ready for Nvidia GPUs yet.
For example, must the GPU have hardware support for double precision?
Also, is the current CPU application compiled from OpenCL, somewhat similar the the way the current POEM@HOME application is?
I've seen rumors that the 7.0 version of BOINC will be available soon, with at least some support for OpenCL GPU workunits, but almost nothing more on just how much OpenCL GPU support.
RE: For example, must the
)
No, the floating point speed of the GPU has nothing to do with OpenCL.
For AMD/ATI GPUs, only the GPUs at the requirements page on ATI's SDK web site, will have OpenCL capability. In numbers, from the HD4300 onwards and every GPU thereafter, plus the Firepor V3800 and everyone thereafter.
For nVidia it's easier, there any GPU capable of CUDA will be able to do OpenCL. In numbers, from the GEFORCE 8300 onwards and every GPU since.
BOINC 7 (being tested as BOINC 6.13) will support both ATI and nVidia GPUs, with a possibility for Intel GPUs to follow once Intel gets an API out. But really, BOINC doesn't need to support OpenCL, as it doesn't do any of the work. The science application will need to do OpenCL.
All that BOINC will do is detect if your GPU is OpenCL capable and if so, which version it is compliant to. Just as it already did detect if the nVidia GPU is CUDA capable and the ATI GPU CAL/Brook+ capable.
RE: What I've seen so far
)
As a starter please see this post of mine.
Apart from that, we do have a dedicated CUDA app so we don't focus on running our OpenCL app on NVIDIA devices for the time being.
Memory-wise 512 MB cards might be already too "small" for the current app. We always strive to provide apps that run on as many GPUs as possible but there are technical and algorithmic constraints that simply don't allow us to reduce the GPU memory footprint of the OpenCL app right now.
No.
Cheers,
Oliver
RE: Apart from that, we do
)
On that though, the whole point of OpenCL is to make a new standard in GPGPU and other processors, where presumably one code can work on many pieces of hardware. So in essence, you ought to be able to use the OpenCL app for an ATI GPU on an nVidia GPU, an Intel GPU, an AMD or Intel CPU, and presumably on any Tesla you throw it at. In a perfect world...
1 task
)
1 task completed!!!
:)
http://albertathome.org/host/1353/tasks&offset=0&show_names=0&state=2&appid=
http://albertathome.org/task/34526
RE: On that though, the
)
Sorry, that's a misconception. While the OpenCL framework itself indeed facilitates the general approach one still has to take into account the differences of each architecture. OpenCL doesn't help you to overcome that out of the box. While a single codebase should run on most OpenCL capable devices, it won't do so efficiently (e.g. thread parallelism vs. data parallelism on NVIDIA and AMD GPUs). Thus it boils down to the fact that you can stick to a single framework for coding/building/deployment but you still have to write architecture-aware implementations for an efficient solution.
Best,
Oliver
Not just that. On my travels,
)
Not just that. On my travels, I've seen that due to ATI's different approach to GPGPU, from Close-to-Metal through BrookGPU to OpenCL, that their drivers are mature enough to support OpenCL on their hardware.
However, for you as a project engineer to be able to make it work on all the different hardware, you need a different API, SDK, IDE, debugger and probably compiler for each different piece of hardware by vendor. So one of each for ATI, one of each for nVidia, one of each for Intel, etc.
Hence why I ended the previous comment with "In a perfect world..." :-)