Hi Luca,
> Sure, I shall split the patch. Joel also suggested splitting the patch
> and I was probably wrong to think that the patch was simple to be a
> single patch :-).
>
> Since equation (3) has theoretical backing, I am perfectly fine with
> using it for our fix. Will have 2 patches as you suggested.
>
> Will get the v3 out soon..
>
Sorry for the delay. I was testing the fixes and noticed one other
issue and was working on a fix.
This is with the extra_bw for tracking the unused bandwidth in the
system. As you have shown me:
"extra_bw = Umax - ((Sum of bw of all tasks) / #cpus)"
But I noticed that tracking the extra_bw globally while rest of the
values as per-cpu causes issues with reclaiming especially when we
have tasks with small bandwidth. This is because GRUB tries to
proportionately divide unused bandwidth based on running bw on a cpu.
So even if there is only a single task in a cpu, it reclaims less if
its bandwidth is less.
I think removing the extra_bw and tracking unused bandwidth using
"max_bw - this_bw" should be okay. Since admission control will
guarantee that we don't admit more than the capacity, we should be
good with allowing tasks on a cpu to use max_bw while reclaiming.
With the above theory, the code becomes simpler and is the same as the
v2 patch.
max{u_i, (max_bw - Uinact - Uextra)}
= max {u_i, (max_bw - (this_bw - running_bw) - (max_bw - this_bw))}
= max {u_i, running_bw} = running_bw
So, "dq = -(running_bw / max_bw) dt"
v2 had passed all my tests. But having Uextra broke couple of those
when it came to multiprocessors and after identifying the root cause
and re-writing the equation, everything works now and passes all
my tests. I have more confidence in the above equation as its derived
from the SMP GRUB rule using our max_bw.
Please have a look and let me know what you think about this.
I have the v3 ready with patches split into 5(including doc patch).
I shall post v3 soon after this so that you can see the code changes
as well to have a better look.
Thanks,
Vineeth
