Thread Priority

[efmer (fred)]

I am using MB_6.08_mod_CUDA_V11_VLARKill_refined it looks you modified the thread priority.
I am reading -1 0, 0 of the 3 running threads.  That is THREAD_PRIORITY_BELOW_NORMAL,  THREAD_PRIORITY_NORMAL, THREAD_PRIORITY_NORMAL.
I already suspected something like that because sometimes the audioplayer just stops for a fraction of a second.

[Raistmer]

It was question?
If it was answer is yes, thread priority changed, process priority changed too.
CUDA thread priority should be higher than priority of other working threads for CPU apps in system.
But it still lower than any normal task thread priority (if app in question (audio player) doesn't use nice too). So any-non BOINC activity going on normal priority should not be affected.
(don't forget that we speak about Windoze, no real respect to priorities from this OS . Look one of Matt's posts in Thech news - how he gave explanation for BOINC deinstallation.... In real-time OS such action would never required!)

[efmer (fred)]

It was question?
If it was answer is yes, thread priority changed, process priority changed too.
CUDA thread priority should be higher than priority of other working threads for CPU apps in system.
But it still lower than any normal task thread priority (if app in question (audio player) doesn't use nice too). So any-non BOINC activity going on normal priority should not be affected.
(don't forget that we speak about Windoze, no real respect to priorities from this OS . Look one of Matt's posts in Thech news - how he gave explanation for BOINC deinstallation.... In real-time OS such action would never required!)

It was maybe a strange question. But I'm looking for the most optimum priority setting.... for one of my computers with a program I'm working with that can change the CUDA working threads priority.
What I've seen is the process running at the lowest priority, probably not doing much, and one thread is running on point below normal, looks like this one does all the real work and 2 threads running at normal priority but as far as I can see they do very little work.

[Raistmer]

yes, you got sitation correct.
these 2 other threads are control (as Joe stated on main forum, shared mem communications timer and exception catcher) ones, thread with lowest priority is worker one.
The feature of this build: CUDA worker thread runs onincreased (non-idle) priority with respect to worker thread from CPU app (idle priority).
Moreover, you can tune priority of CUDA worker thread by change process priority via task manager (for example).
It's not possible for CPU apps.
One can increase worker thead priority ONLY by setting process priority to RealTime, other options will not work (cause thread itself has idle priority). This behavior Windows-specific I suppose and described on MSDN.

[efmer (fred)]

One can increase worker thead priority ONLY by setting process priority to RealTime, other options will not work (cause thread itself has idle priority). This behavior Windows-specific I suppose and described on MSDN.

You're sure because I tested this with a load application and running the process on normal and setting e.g. 2 worker threads (on a 2 core system) to e.g. real time shuts down most of the system  as expected.

[efmer (fred)]

One can increase worker thead priority ONLY by setting process priority to RealTime, other options will not work (cause thread itself has idle priority). This behavior Windows-specific I suppose and described on MSDN.

You're sure because I tested this with a load application and running the process on normal and setting e.g. 2 worker threads (on a 2 core system) to e.g. real time shuts down most of the system  as expected.

I think you're probably right, but that means the GPU may get better filled if the process priority is raised somewhat. Because it is competition with other BOINC projects running at the same priority  (idle). 

[Raistmer]

scheduling performed by per thread basis, so thread priority matters. And it' already increased.
You could try to play with priority more though and report results here (better if it will be numbrs).