Can the developers in the mean time fix the driver detection?

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

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

Please, update your hosts with this builds. ATi and NV builds got speed increase, HD5-version added, NV version added.

CPU builds got some updates too.

Known issues:
1) As with OpenCL AstroPulse, last driver version from both vendors show increased CPU usage. AMD already aknowledged this issue and promised to fix in new Catalyst releases, NV still keeping silence about this issue.

2) OpenCL NV app can silently (i.e., w/o errors in stderr) produce incorrect results (overflows). Again, situation resembles NV AstroPulse rev521 case and usually means too long kernel call. Why NV OpenCL runtime doesn't report error code for kernel enqueue runtime call - no idea. But low-end NV GPUs could be not capable to use this app. This testing should determine GPU requirements for NV app too.

Thanks, I downloaded the WU and confirm the difference.

At least part of the problem is a long-standing issue which is seen in SETI@home Enhanced between the stock CPU and stock CUDA applications too. The most efficient order to do the various searches is different for CPU and GPU, so for this kind of task with a lot of potential signals the CPU finds a different subset than the GPU does. Eric Korpela is aware of the issue and maybe if he or Jeff Cobb get a chance the Validator code will be revised to judge quick overflow results differently.

The way that's pertinent to this case is the MB7 r365 sources are primarily targeting openCL builds so the Autocorr and Spike searches are done in a different order than stock 6.97.

At first glance, that doesn't seem to explain all the differences between the MB7 CPU result and 6.97 result. I need to analyze more to really be sure what the data indicates.

FWIW, those results will be considered "weakly similar" and both get credit when the third result is returned.

Maybe this is part of the issue regarding validation?

[quote]tex 1954

It certainly seems something is lacking in whatever SDK is being used...

:)

Another interesting observation. When boinc only uses 2 cores for the calculations on the CPU, and the app 1.19 (atiOpenCL) process is running with realtime priority we have is Screenshot.

1.19 (atiOpenCL) app process uses 25% of the CPU (what is the total load of one core) and GPU load, thus increasing to a maximum of 95%.

My opinion: Too much demand for the CPU, because of what is too low speed of calculation. For comparison, the NVidia 9500 GT sredenee calculation time of one WU 2,5 hours.

Ok...
Me babling about lots of non of use stuff here....

I really can't follow any of what you're saying, but thanks for the effort anyway ... :-)

Please forgive me if I don't follow this thread closely at all from here on, I have other fish to fry. :-)

Cheers, Mike.

That's good, very good !

As ever, humor is the answer .... :-)

Cheers, Mike.

What we have here is an issue of 'locus of control' - who is running the shop?

With all sorts of arguments presented - of varying bases - and subsequent annoyance by some of the failure of the proposal that the act of volunteering requires/demands that their wishes be met, when all it really needs to resolve is that they simply don't volunteer their equipment for use. So in practical terms it's a simple resolution, as advertised.

Also it would appear that most posters have missed the self evident practical problem that offering credit implies some time consuming and frequently fruitless interaction ( as exquisitely demonstrated here ) with people that are quite unable to take no as an answer, however expressed. And that often depends upon people's personal horizon of interest and level of maturity ( a wide array of both ). That is the base reason for not offering credit here at Albert. We need a test milieu without the endless complication of such interactions. Clearly some see that as a dangerous paradigm to allow ..... :-)

Please keep discussing though, as I'm sure some sociologist one day may wish to view this, probably to sort out the question of "what were those credits all about?". :-)

Personally I think the time will come when credits will die, for the practical reason of the type of inefficiencies as demonstrated here in this thread. An enterprise that is touting for outside help by definition also may lack resources to take on the extra burden thus entailed. To wit : the effort required to keep 'noisy' crunchers happy may not be offset by their gifts. Have any of those complaining here considered that - or do we divert, yet again, to emotional landscaping?

Cheers, Mike.

( edit ) Oh, I forgot to mention : this is a science project looking for astronomical events. ;-)

Is there any way that you can increase GPU load?

I'll try to answer this question.

Yesterday I ran the app 1.19 (atiOpenCL). My computer is AMD Phenom X4 925. Two ATI HD4850 1GB. Software is Windows 7 Ult x64. Boinc 6.13.12. AMD Driver 11.11
Screenshot

It seemed to me that the GPU load is very low. I looked at the Gigabyte HUD program and saw that the load on the GPU is ~ 40%. And progress of WU increased by 0,002% in the second.

I tried the following method. Priority of the process - real time.
Screenshot 1
GPU load was ~ 80%. And progress of WU increased by 0,003%-0.004$ in the second.

Then I returned the priority - below average. GPU load returned to ~ 40%. Progress of WU increased by 0,002% in the second.
Screenshot 2

Then I again tried real time priority.The result was repeated. GPU load ~ 80%.
And progress of WU increased by 0,003%-0.004$ in the second.
Screenshot 3

The screenshots you have to look at Image

BOINC projects are credit based projects as the only thank you a project can give to a user is via a metric called cobbnlestones or credits.

Where does it say that a project needs to give credits/cobblestones?
Where does it say how many?
Where does it say that those have to be exported?

And again, to explain further what this project does, Bernd wrote in this message:

We can and would test new applications (and application versions) either privately or on the main project. What we intend to test on Albert is server side code that can't be tested on the main project without risking serious damage. The OpenCL support of BOINC requires server side modifications, too, that's why we test the OpenCL apps here.

Albert@Home stats are currently being exported, I intend to shut this down late this week. There will be a news item (including RSS) at least 24h in advance.

But I still like to understand what happened here, how & why people that are only interested in credit ever attached a host to this project. This project was never meant to attract this kind of attention. I'd like to avoid such misunderstanding in the future.

BM

So I guess :

A simple "NO you can't" from a program developer would have been a nice reply

hasn't suited you for this thread then ? ;-)

Cheers, Mike.

I wonder if the tasks are in different sizes here...

Nope, simply depends on the GPU...