Mine had been asking WCG for GPU work for ages... at most every week or fortnight when the WCG tells the client to back off and every time if the Update button is operated. This was until I disabled it on this host with <no_gpus>1</no_gpus>

--//--

P.S. What is relevancy to topic?

References:
Sekerob [Mar 3, 2012 8:16:01 PM] post
skgiven [Mar 3, 2012 8:03:42 PM] post

skgiven, thanks for responding.
Support for GPU-crunching may be a feature of BOINC alright, but "requesting new tasks for ATI GPU" strikes me more like an error statement than anything resembling a form of advertisement or a call-to-attention about any BOINC feature for that matter. The existence or the promised capabilities of a feature is distinct from the actualization or use of the feature. The "requesting new tasks for ATI GPU" statement is more of a reference to a WCG GPU implementation details (or the lack of such details as the case may be) rather than a reference to anything about the GPU-related support features of the BOINC software.

SekeRob, thanks for responding.
I feel and I think that there is something not correct with the way the mechanism behind the BOINC messages work and this may have something to do with the symptom of "Waiting for GPU memory". For example, why would the other-side (BOINC or WCG-server) act as if the client is asking for GPU work enough that, in your explanation, that WCG tells the client to back-off? The way I see it, the WGG-server need not tell clients to back off, (or that BOINC is "Waiting for GPU memory") -- for they have never requested any GPU work to begin with, or did they? By the way, is the <no_gpu> tag seen as the final arbiter in deciding if a client asks for a GPU work or not? What happens to website control from website config user-settings? And we don't have a GPU-aware WU yet, I understand, so in the interim, why is WCG or BOINC assuming that clients ask for GPU-work and seemingly taking their queue from the value of the <no_gpus> tag? Is that a safe assumption while nothing GPU-aware_related-WUs are yet operational? I suggest that something be done to have the value of the <no_gpu> tag be ignored, and servers to take the position that clients are not asking for GPU work -- and hold this position until after WCG is about ready to support a GPU-crunching operation.
;
; edit1: 2012.03.03Sa.2304.utc > corrected datimes of references' post.
;

Andzgrid, in response to your second question, there is a lot to reply to...where to start.

First, BOINC cannot override the cc_config. That would defeat the purpose. Another thing to keep in mind is WCG is a project only. BOINC covers many many projects so whether WCG has a GPU project or not is irrelevant to BOINC. It will ask at the specified intervals of every project as long as there is a GPU available to crunch with and it needs work.

The rest of your argument is circular and really hard to follow. Just be assured that the reasoning behind the work fetch schemes are there for a reason. If you wish to change the algorithms, take that suggestion to the BOINC developer mailing list.

Just so people do not go ballistic about this problem let me make a few remarks.

I think, as usual in hindsight, that I may have caused the problem to appear by changing the preference file checkmark "Use GPU while system is in use" while WCG was running (I know of no other way).

Subsequently I got the message about waiting for GPU memory.
What is happening is that allthough the PC does not have any GPUs, WCG code will allocate memory based on the checkmark in the preference file.
Thus in the startup messages you get the message "No usable GPUs found" .
In other words, the code finds out that there are no GPUs allthough the checkmark says to use GPUs. But the memory is allready allocated and remains allocated based on the checkmark.

Now I change the check mark while WCG is running. I now delete the memory allocated for the GPUs that do not exist.
When the program checks for the GPU memory it fails to find the memory hence the message "Waiting for GPU memory" shows up.

The code should be changed so that we do not allocate memory for GPUs based on the checkmark in the preference file unless we first check that the GPUs are indeed present i.e. installed.
In other words ignore the preference file GPU checkmark until the presence of GPUs has been verified.
Alternatively, deallocate the memory that was allocated based on the checkmark once we realize that there are no GPUS.
I think that will fix this minor problem.

In any event, it appears based on what I have read in recent forums, that WCG is about to start using GPUs fully and the coders may correct this minor problem at that time.