Sending work
Submitted on 31 Oct 2011 8:42:08 UTC
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.
Language
Copyright © 2024 Einstein@Home. All rights reserved.
Comments
It's not that bad after all.
)
It's not that bad after all. However, while our current OpenCL app runs on NVIDIA GPUs indeed it doesn't yet produce valid results on all their models. Most likely due to some subtle differences (optimizations) in NVIDIA's runtime OpenCL compiler... We're investigating but we'll focus on OpenCL@AMD first since we got a CUDA app anyway...
Oliver
Looks like the OpenCL app for
)
Looks like the OpenCL app for ATI can't validate against the CPU app: http://albertathome.org/workunit/11889
Oh and thanks for making the initial replication 3, but leaving the third one unsent. ;-)
Separate post for this, can
)
Separate post for this, can you stop sending ATI OpenCL work to non-6.13 clients?
I see one of my other tasks is waiting for a wingman, because two people got "Failed to get OpenCL platform/device info from BOINC (error: -161)!" (6.12.34 and 6.10.60), the third had it running on the failed CUDA app.
Since you're checking for the information from the client, you may just as well make the client mandatory.
RE: Separate post for this,
)
Guess what we do already? :-)
We're aware of that problem but there's no fix right now... Still investigating...
Oliver
RE: Looks like the OpenCL
)
We're still tuning the validator. That's part of this test, we need to sample the numerical stability/inaccuracies across various platforms and devices...
Hm?
RE: RE: Oh and thanks
)
Sneaky... now it's sent to a third party.
It wasn't for several hours last night, before I made the comment... I just needed more patience then. ;-)
RE: Guess what we do
)
One further request then, decrease the deadline here? I see wingmen who have seemingly abandoned the cause, so now I'll have to wait 14 days before anything is resent. That's too much for a test project. Just make the deadline 3 to 5 days, that's time enough to crunch the work, even on a multi-project system, and send it back.
You want the results back fast, don't you? :-)
RE: We're aware of that
)
Should work now...
RE: RE: I've seen rumors
)
Looks adequate for computers with only one GPU.
On other BOINC projects, I've seen one thing mentioned that BOINC really needs to do if you want to be able to run both an OpenCL GPU workunit and a non-OpenCL GPU workunit at the same time on computers with more than one GPU - provide a standard way of mapping the way BOINC identifies which GPU it has assigned to a workunit to the way OpenCL identifies those GPUs. Otherwise, it's likely that those two workunits will often attempt to use the same GPU at the same time, with both failing.
Except for this, I've seen nothing on features for OpenCL that need to be in BOINC instead of the application. Adding more features would be helpful, but not required, for BOINC projects wanting OpenCL GPU workunits.
I currently don't have any AMD/ATI GPUs, so I won't be able to much for Albert@Home soon, at least until I find time to measure the size requirements for putting such a graphics board into my newest computer.
RE: Albert@home is still an
)
I've recently received some SSE and CUDA workunits. Should I consider them outdated workunits left over from Einstein@Home, or should I consider them Nvidia and CPU versions of the OpenCL workunits?
RE: I've recently received
)
The latter. All BRP4 work units on albert can be considered equal and it doesn't matter whether they're run on a CPU, a CUDA device or via OpenCL. This allows us to check cross-platform/device validation and get a feeling for their relative performances. It also allows us to test BOINC's behavior in mixed-GPU (NVIDIA and AMD GPUs in one box) setups - so far the results look good.
Best,
Oliver
ATI OpenCL and CUDA32 don't
)
ATI OpenCL and CUDA32 don't validate together either.
http://albertathome.org/workunit/12465
RE: ATI OpenCL and CUDA32
)
Not universally true but dependent on the NVIDIA GPU. My observation: a GTX 285 (Tesla) does, a GTX 580 (Fermi) doesn't. We'll investigate and will most likely tune the validator. It just takes some time since we're running low on manpower right now and got the download server issue at first prio.
Stay tuned,
Oliver
Another one then,
)
Another one then, http://albertathome.org/workunit/11889.
That's one OpenCL vs two CPU, guess whose work was deemed invalid? ;-)
Also looks like you need more
)
Also looks like you need more stringent GPU capability detection.
this host has an OpenCL capable GPU, but for some reason on has 384MB memory. Thus all its tasks err.
As far as I know you can set up the scheduler to check for OpenCL capability, memory on the GPU, BOINC client used, etc. GPUs like this need to be locked out, before they're sent work as else it's using unnecessary bandwidth and confuses the user to no end.
RE: you can set up the
)
Yep, already on our TODO list (mostly done, memory requirements pending)...
Oliver
HD4890: 1WU = 62,806.60 sec =
)
HD4890: 1WU = 62,806.60 sec = 17.5 hours and... 500.00 credits only, I guess? )
I understand, that A@H is a test project, but for my opinion it's a waste of time for HD4890.
Don't say me that HD4890 is a low-end GPU )
HD4890 brought me 1Ms in Collatz, MilkyWay, Moo! Wrapper.
In these projects HD4890 is more effective than GTX460, which usually spend only 1 hour for the same WU at Einstein@Home.
credits for this alpha
)
credits for this alpha project are 500 per WU that validates, irregardless of how long it takes to run. Also, last I read this project is soon not going to export stats, if they haven't yet turned that off.
CAL ATI Radeon HD 4700/4800
)
CAL ATI Radeon HD 4700/4800 (RV740/RV770) (1024MB) driver: 0.1
It seems Ageless and I have the same video card. Thus far, I haven't been able to validate a single WU:
http://albertathome.org/workunit/12534
http://albertathome.org/workunit/12549
http://albertathome.org/workunit/12550
http://albertathome.org/workunit/12551
http://albertathome.org/workunit/12294
Also (as mentioned above) it looks like there are problems with the nVidia cards validating with themselves (wu 12294):
The GTX 260 validates with the GTX 590, but not with the GTX 580.
ATIOpenCL v1.19 doesn't
)
ATIOpenCL v1.19 doesn't validate with another ATIOpenCL v1.19 either:
TankMaster (HD 69xx) versus me (HD 4850) == http://albertathome.org/workunit/12850 (Completed, validation inconclusive) with a third task going out.
The next two tasks I have will be against other HD 48xx Radeons. I aborted all tasks that had CUDA or CPU as wingmen. They don't validate anyway so it's kinda useless to spend electrons on that at this time. Perhaps later again when there's been time to fine tune the validator even further.
RE: ATIOpenCL v1.19 doesn't
)
After 5+ hours the third task finally went out, now to a CUDA GPU. I can tell you up front that this isn't going to validate.
Is HR a thought?
Hmmm... Forget HR. Forget
)
Hmmm... Forget HR. Forget using any HD4850 it seems, or any OpenCL 1.0 only GPUs? I noticed that Infusioned now does have credit, but he did so with a CPU task, not his HD4850.
http://albertathome.org/workunit/12832 is my HD4850 versus Alexone's (presumably) HD4850 or 4870. Validation is inconclusive.
The difference here being my Intel i3 versus his AMD XII perhaps? Or my 8GB of RAM versus his 4GB?
A third task went out to a CUDA, we all know how that's going to end. I'm ending my extra electron burn here, I aborted the other similar task, since it would put my GPU against Alexone's again anyway. Either my GPU is broken, or I use a driver version your app doesn't like (Catalysts 11.6), or it's the direction of the ley lines crossing under my computer room that break things. I stop testing for now and turn my attention fully to Skyrim. ;-)
RE: RE: you can set up
)
Due to a bug in the BOINC client this has to wait until at least 6.13.13 (should be released soon).
Stay tuned,
Oliver
RE: Either my GPU is
)
You need at least Catalyst 11.7 (8.872) since we use APP SDK 2.5 until further notice.
Oliver
RE: Due to a bug in the
)
Looks like there ain't gonna be a 6.13.13, but that instead it's going to be BOINC 7.0.x :-)
RE: RE: Either my GPU is
)
Darn, we just updated the validator (using less strict tolerances). We won't know whether the 11.6 you used before updating to 11.7 played a role in your tasks being considered invalid :-/
Oliver
What's it you're saying,
)
What's it you're saying, Oliver? You want me to return to 11.6?
That's no problem. I'm here to test work for you, remember? Totally not in it for the credits. :-)
I'll run the one ATIOpenCL v1.19 I have started already to completion on 11.7, then switch back to 11.6
RE: What's it you're
)
No, you don't have to. You may of course do so as it wouldn't require a full core anymore. Maybe the relaxed validator settings let your tasks through now (we had to tune it anyway)...
Thanks for supporting this effort!
Oliver
Hello, HD5770 fails after
)
Hello,
HD5770 fails after ~30 seconds
http://albertathome.org/task/51833
It is Debian unstable, fglrx 11.11, amd-app 2.5, BOINC 6.13.12. The installation of amd-app is essential. Without it, neither primegrid nor albert@H binaries can be executed.
Cheers,
Steffen
RE: [19:27:27][4969][ERROR]
)
Sorry, not enough GPU memory.
Why? What happens if you don't install it? The runtime libOpenCL.so should already be installed with the driver (as of 11.9 IIRC). Hm, maybe you still need to register the ICD...
Oliver
RE: RE: What's it you're
)
I did return to 11.6, but left SDK 2.5 on my system.
Immediately the use of the one core went back to 02-10%, instead of the full 25% it was using before that on 11.7.
Skyrim is now also back to being a bit more stable. I had many more CTDs (crash to desktop) with 11.7 than I have had with 11.6; where with 11.6 it would be perhaps once a day, with 11.7 it was 7 times yesterday alone. So after the last CTD I reverted back to 11.6 ;-)
So... will need an eye on what http://albertathome.org/workunit/13760 will go do. It's me versus two CUDA that can't decide between themselves who's right. I doubt I'll be the clincher for them. ;-)
RE: So... will need an eye
)
As I thought, I wasn't the clincher. That task was crunched with catalysts 11.6
Looks like these don't validate to ATIOpenCL yet. Further fine tuning of the validator? I have suspended the last 3 tasks I have until I hear more. Although, I could of course abort them and see what the newer scheduler in 7.0.2 thinks I should get for loads of amounts of work. With a REC of 11000, too much anyway. ;-)
Remember, work that got credit is bad work. Work that didn't validate is good work. It tells the developers here their validator isn't ready yet to work in the outside angry world. :)
(and really developers, how many set it and forget it users do you have on here? ;))
RE: So... will need an eye
)
I see it was another dud. It took another CUDA32 to get validated, me and another CUDA32 finishing outside the points.
I have 4 new tasks.
12537 is paired against a BRP3cuda32. I think I won't even try.
15980 is two ATIOpenCL. I wish driver detection here was working so I could make a reasonable guess as to what driver the other guy is using. Driver: 0.1 is useless.
15971 also has me paired against a BRP3cuda32.
15943 also has me paired against a BRP3cuda32.
15980 it is then.
Can the developers in the mean time fix the driver detection?
Completed, marked as
)
Completed, marked as invalid
WTF?
Easily explained:
)
Easily explained: http://albertathome.org/workunit/12082
You're ATIOpenCL, you were paired against two CUDA which walked away with the credits. The validator isn't tuned enough yet to see that these results may well have been the same.
This is from the Seti@Home
)
This is from the Seti@Home Beta message boards (developing an ATi OpenCL App):
http://setiweb.ssl.berkeley.edu/beta/forum_thread.php?id=1867
Maybe this is part of the issue regarding validation?
RE: This is from the
)
I think the problem might be in the FFT and CuFFT variations....
But I am not sure. I saw something about such a discussion in another thread...
RE: Maybe this is part of
)
Nvidia cards are still using the CUDA app though, not OpenCL.
It doesn't mean that NVidia
)
It doesn't mean that NVidia cards don't silently generate overflows in general.
Also, the second post I quoted details how even the CUDA app was not validating against a CPU due to the order of calculations and the validator needed tweaking to regard them as weakly similar.
RE: Can the developers in
)
Sorry, not up to us. I'm not sure whether the BOINC devs can do anything about it since this might even be an AMD driver issue.
Oliver
RE: RE: Can the
)
It seems to work in previous boinc manager versions. Maybe not in 6.13xx yet though....
RE: Sorry, not up to us.
)
Of course it's up to you. Apparently the server back-end version that you use doesn't store the ATI/CAL driver version, but it is sent to you.
From my 7.0.2 sched_request_albert.phys.uwm.edu.xml file:
You can even use the OpenCL information.
Then with the CAL version we can figure out which Catalysts they are. E.g. CAL 1.4.1417 is Catalysts 11.6
RE: RE: Sorry, not up to
)
Jord has a point here.
And it can help users detect possible wrong drivers when they compare with other users. On the question, "which driver is the best driver for my ATI card?"
If you want to be totally
)
If you want to be totally confused, it does work on Einstein. See my account there.
You can't blame it on the client version either, it's merely all ATI that is affected. Examples: this host uses 6.10.58 and shows driver 0.1; this host uses 6.12.41 and shows as driver version 0.1
RE: RE: Can the
)
You may talk about two different things here.
Jord, what exactly do you think should be fixed?
I do see that displaying the ATI CAL/driver version on the host web pages appears broken (on Albert), and possibly the string in the DB is, too.
In the scheduler the ATI "driver" version is stored as "char version[50]" and "int version_num" in coproc_ati, and in "char opencl_driver_version[32]" in opencl_device_prop. These could in principle be used in app_plan(), though we don't check this yet.
BM
BM
RE: Jord, what exactly do
)
Showing of the CAL driver version on ATI cards on the account pages here.
Yes, sorry, I said it wrong. I asked for a fix for the driver detection. I know you don't do that, that that's up to the client. I meant that all the driver versions showing for Nvidia GPUs is correct, for all ATI GPUs it's always 0.1, which isn't correct.
Hm. On your host page I
)
Hm. On your host page I currently read:
AMD ATI Radeon HD 4700/4800 (RV740/RV770) (1024MB) driver: 1.4.1417I don't see anything wrong with that. Maybe the previous entry was from an old Client version?
The only thing I changed this morning was (parts of) the web page code, but nothing related to the pages involved here.
BM
BM
Well, whatever you did fixed
)
Well, whatever you did fixed that bug. It now shows on all hosts I checked which CAL driver version these people use. It may have been in there all this time, just not showing as such. So thanks. :)
Btw Oliver, BOINC 7.0.1 (the minimum requirement now) was never compiled and stored anywhere. The minimum minimum anyone could download was 7.0.2; only people who got the source code of branch_7.0 and compiled that on the 30th of November will have 7.0.1, all else will have 6.13.12 or 7.0.2. ;-)
RE: Btw Oliver, BOINC
)
I know, I'm just using the exact tag/version that contains the required bug fix :-)
Oliver
RE: RE: RE: Can the
)
I just read something related on the boinc_dev mailing list. It seems that BOINC 7.01 and 7.02 don't allocate enough digits in one of the places they store
ATI version numbers, and are therefore likely to get at least some of the version numbers wrong.