Re: [PATCH 1/2] sched/deadline: accurate reclaim bandwidth for GRUB

[Vineeth Remanan Pillai <vineeth@xxxxxxxxxxxxxxx>]

Hi Luca,

On Thu, May 11, 2023 at 3:37 AM luca abeni <luca.abeni@xxxxxxxxxxxxxxx> wrote:
>
> I've just seen v2, and (unless I misunderstand something) I see you
> removed the max{u_i/u_max, 1 - (u_inact + u_extra}} thing?
>
> I fear this might break the real-time guarantees provided by the
> algorithm...
>
I am sorry I missed sending more details before sending out v2. So, I
think there is another bug in the existing implementation. Let me try
to explain the details.

SMP GRUB paper has the equation for depreciating runtime as:
  dq_i = -max{u_i, 1 - (extra_bw + Uinact)} dt

Since we are caping at Umax, the equation would be
  dq_i = -(max{u_i, Umax - (extra_bw + Uinact)} / Umax) dt (1)

But existing implementation is:
  dq_i = -max{u_i/Umax, 1 - (extra_bw + Uinact)} dt (2)

Here in (2), we factored Umax only to the first term "u_i" and the
second term in max{} was as left as it is. What do you think?

Now with normal DL and SCHED_FLAG_RECLAIM tasks, equation (1) can be
re-written as:
  dq_i =
   -(max{u_i, Ureclaim_max - (extra_bw + Uinact)}/Ureclaim_max)dt (3)

I tested this equation (3) and it works as expected. What do you think
about the correctness of equation (3)?

I felt that, since we are using sequential reclaim mentioned in the
paper and we isolate all parameters per-cpu(except for extra_bw) we
could use the "-dq = -(U/Umax) dt" equation as it was simpler than
equation (3). Sorry that I missed discussing this. I shall send out
v3 with equation (3), if you think it's the right way to go to enforce
deadline guarantees.

Thanks,
Vineeth