> there is no chance to get a relation between a partition and a core.
> APS is looking at all available cores as 100%
> 
> /hp

Ok, thanks for clearing that up. I somehow kept wanting to combine partitions and cpu affinity, while they're really two
 separate mechanisms.

After more reading and thinking, I've come to the conclusino that what I really want/need is just to set CPU affinity on
 all processes (at least all I can) to CPU0, so that my worker threads can execute without interruptions on the other 3 
cores.

Dankeschön,
~Lorenz

Thank you for your reply, Attilla.
I was indeed confused by partitions and runmasks. Now I see clearly that they're actually two completely separate things
.

> That's why I asked. I'm not aware of any easy way of setting affinity on proc 
> theads. There is a hard way: use a series 
> of DCMD_DEBUG_THREAD devctls on PROC to list out all of the thread IDs (see 
> the source code of pidin for an example) and 
> then call ThreadCtl() on each one.

Ok, that seems at least a little more elegant than the grep-ing and sed-ing I had planned ;-)

> However, there may be a way to do what you want without cpu affinity, using 
> aps only. If you want to give a bunch of 
> threads 75% total cpu and another bunch 25% of total cpu, APS can do that with
>  two partition. 

I tried removing the affinity , but then the worker threads in the partition with 75% would migrate between all 4 cpus 
all the time.. which I want to avoid for efficiency reasons.

So right now I have exactly that: System partition with 25%, Worker partition with 75% CPU. 

> The only requirement is 
> that on an <n> cpu system, each partition have at least <n> threads (other wise one 
> partition may not use up its entire 
> budget).

Yep .. that's something to consider. But as my worker partition has 3 threads that each run exclusively on their 
respective cpu and are always running (busy loop), I should always be able to reach the full budget (75% = 3 of 4 cpus).
  That's at least my design concept.

To sum up my goal, this is what I want:

CPU1 = runs *only* worker thread 1 (never interrupted!)
CPU2 = runs *only* worker thread 2 (never interrupted!)
CPU3 = runs *only* worker thread 3 (never interrupted!)
CPU0 = handles ALL other threads.

Can I really achieve this through APS without manually setting each thread's runmask?


I'm quite close of achieving this, but my worker threads get interrupted, even though they have max priority. 
However, I've noticed that the threads that interrupt my threads are mostly (all?) interrupt-driven stuff (device 
drivers), which would explain why they're allowed to preempt my threads.

So I might get away with just runmasking each device driver to CPU0, but as Attilla said before, this can get ugly.
Any other suggestions?
I mean, it's not like I can/want to mask all other interrupt handlers indefinitely long..

--
Another question on the side: What are these KER NOP enter/exit I see about every 1ms? Even when a thread is not 
rescheduled, the trace shows these events. Is that the "system tick"?

Thanks,
~Lorenz


p.s. And thanks for pointing out to me the difference between "runmasked" and "affinity" - that definitely helps 
minimize the confusion.