From what I understand about 1.2 is that a GPU can be partitioned, so that it runs multiple WUs. Can this be implemented in a new version of BOINC, so that users with GPUs can partition their cards in ....CU???.
What I would want is, for the hard core scenario :)
If I had a gpu card with 4 Compute Units, I would want to assign it like this.
1st CU for OS, and it should not be touched by BOINC (to avoid screen lagging while doing other stuff on the computer)
2nd CU for POEM
3rd CU for SETI
4th CU for WCG :P
I remember that I asked this question somewhere, and I got a response that there wasn't support in drivers. But now that there is support, this should be important for some users, like me, which don't have powerful gpu cards.

Forwarded to development.

---

I know that my first post is something like 3 years in the future, nevertheless ... Thank you Ageless.

Maybe a first approach would be to have an option in an 'experimental' tab, like:
Don't use the fist CU on the GPU.

In a second approach, another option similar to the multiprocessor systems could be introduced:
On powerful GPUs, use only x percent.
e.g. on the 4 CU card.
if I put 25%, boinc should understand to use the 'last' CU.
at most 50% ... boinc should use CU 3&4
at most 75% ... boinc should use CU 2&3&4
This option could also replace the number of applications run on the gpu, from the app_info
e.g.
- <coproc>
<count>0.5</count>

Now, what I read about this option is that the main device can be divided into sub-devices which can all be addressed separately, plus the main device. So in your example of using sub-devices for POEM, Seti and WCG, you wouldn't need a sub-device for the OS, as you can use the main device for that.

Still not sure if it's BOINC that needs to be able to do this, more like the project's applications. But I'll leave that up to development to answer, be it here or to me in email. Most of them aren't available at this time, so I am not expecting an answer soon. Hope you aren't either. ;-)

---

David Anderson wrote:

BOINC already has the capability for jobs to use a fractional GPU. To use this, the schedulers of those projects (the plan class function for OpenCL apps) need to be changed to detect when this is possible,
and specify it.

AFAIK, Milkyway@home is the only project that currently uses fractional GPUs.

-- David

---

Good, but I still would like some options in BOINC manager so that I can manage the 'rest' of the GPU, to limit and assign the resources for different projects, and not use the app_info file.

Hello,
I found this topic interesting.
Mainly because:
-most users have only one GPU in theirs computer.
- this causes the computer to lag even during the most basic tasks like reading Wikipedia(what is very annoying!!!)

So if there will be the option to reserve at least 1 Compute Unit(CU) to tse Operating System it would be nice.
Or any other way to do this (ie. scroll bar with percentage in advanced options)

There is already CPU control build in the BOINC.

This new feature could be marked as experimental , beta etc.


Later on You can add the possibility to partition the device freely(aka. virtualization) for different projects.

Thanks in advance.

How about sending double precision tasks to the gpu that actually has DP hardware and avoiding the single precision GPU? Can that be included in the device partitioning? I still have this problem with milkyway project getting assigned to a single precision gpu.

Software such as DVDFab uses CUDA for video encoding. Currently, I stop boinc when I bring up DVDFab. This is done automatically using that cc_config file. However, it would be nice if they could both run at the same time. Ideally, the the device partitioning could handle this one would think.

How about sending double precision tasks to the gpu that actually has DP hardware and avoiding the single precision GPU? Can that be included in the device partitioning? I still have this problem with milkyway project getting assigned to a single precision gpu.

In the alpha test versions of BOINC there is the ability to exclude a gpu from a project or an app. It's worked from about 7.0.8 onwards. But before you rush out to try the alpha versions I suggest you read the change log message thread, in particular the incompatibility warnings.

Further out there is an idea to get BOINC to recognize different gpu capabilities and schedule accordingly. No idea when that is slated for, but at least its on the drawing board.

---

MarkJ

How about sending double precision tasks to the gpu that actually has DP hardware and avoiding the single precision GPU? Can that be included in the device partitioning? I still have this problem with milkyway project getting assigned to a single precision gpu.

In the alpha test versions of BOINC there is the ability to exclude a gpu from a project or an app. It's worked from about 7.0.8 onwards. But before you rush out to try the alpha versions I suggest you read the change log message thread, in particular the incompatibility warnings.

Further out there is an idea to get BOINC to recognize different gpu capabilities and schedule accordingly. No idea when that is slated for, but at least its on the drawing board.

This seemed to work fine. I posted my cc_config and observed results here at the milkyway forum.