As you may know, in GPGPU application, you need to keep a small amount of CPU free to ba able to feed the GPU with data to process and to retrieve the results from the GPU.

I've noticed that some CPU application in BOINC (Constellation@home for instance) tend to run at the same or higher priority than the dnetc client that Moo uses, resulting in a a GPU not being used at 100% and slow performance.

I remember when I first tried the dnetc official GPU client that I had to change the priority in the preferences.

Is it possible to apply this modification to the dnetc client that Moo uses ?

P.S : I'll give you the values I used for my dnetc official client this evening when I'll be in front of my GPU machines.

Hi,

Just to confirm here that we can certainly change the priority setting for the D.Net Client. At the moment we are using priority=4, which is the recommended value for their GPU clients. The scale is 0-9, where 0 idle and 9 almost normal priority. Even with 9 client won't use normal process priority but rather bump it's worker threads to have bigger relative priority. This is not something I can directly control since that's how the D.Net Client works.

There's probably no harm if I switch this to 9 for the GPU apps. That way they get the most CPU cycles to feed the GPUs.

-w

Sorry for the delay in answer, I forgot about this post ...

Anyway, I just checked how I set my GPU clients : it was set to 2 for my ATI client, but it was to be run with FAH clients.

Now, with the current settings of the Moo client and the Constellation application (on CPUs), I only get ~65% GPU usage of my two 9800 GTX+ and 90% on the 5770.

Bumping manually the priority of the Moo client from idle to normal in Windows Task Manager allow me to use 100% of the GPUs.

I hope it'll help you tweak you client to solve the issue.

The impact is much more important on my two 9800 GTX+. A WU takes twice the time if I don't set the priority manually.

On the ATI, the impact on performance is almost unseen ... (and as mentionned in previous post, the difference in GPU usage is lesser than on the NV cards).

Sorry for the delay in answer, I forgot about this post ...

Anyway, I just checked how I set my GPU clients : it was set to 2 for my ATI client, but it was to be run with FAH clients.

Now, with the current settings of the Moo client and the Constellation application (on CPUs), I only get ~65% GPU usage of my two 9800 GTX+ and 90% on the 5770.

Bumping manually the priority of the Moo client from idle to normal in Windows Task Manager allow me to use 100% of the GPUs.

I hope it'll help you tweak you client to solve the issue.

Doing this in the Task Manager is something Boinc probably won't support, Boinc is a process designed to use UNUSED processor cycles, not all the cycles it can to the exclusion of anything else. Boinc was never designed to be used on 'Boinc only' machines. We users do that on our own and the Developers actually discourage that when asked. Now tweaking the project, as Teemu showed everyone how to do, is okay but in the Windows Task Manager probably not. You can always do it manually but everytime you start a new workunit you have to do it again, as it resets after each one is completed.

Doing this in the Task Manager is something Boinc probably won't support, Boinc is a process designed to use UNUSED processor cycles, not all the cycles it can to the exclusion of anything else. Boinc was never designed to be used on 'Boinc only' machines. We users do that on our own and the Developers actually discourage that when asked. Now tweaking the project, as Teemu showed everyone how to do, is okay but in the Windows Task Manager probably not. You can always do it manually but everytime you start a new workunit you have to do it again, as it resets after each one is completed.

I'm using Bill2's Process Manager : I created a rule on the name of the client, so everytime it starts, B2PM bump the priority to normal automatically :)

Doing this in the Task Manager is something Boinc probably won't support, Boinc is a process designed to use UNUSED processor cycles, not all the cycles it can to the exclusion of anything else. Boinc was never designed to be used on 'Boinc only' machines. We users do that on our own and the Developers actually discourage that when asked. Now tweaking the project, as Teemu showed everyone how to do, is okay but in the Windows Task Manager probably not. You can always do it manually but everytime you start a new workunit you have to do it again, as it resets after each one is completed.

I'm using Bill2's Process Manager : I created a rule on the name of the client, so everytime it starts, B2PM bump the priority to normal automatically :)

Ahhh I see, thanks.

Since I installed 11.10 drivers, the client is no longer using my CPU ... :)

I hope this issue is definitely behind us.