>
> It looks like DomR's vCPU does get 50% of CPU time, so it's not that
> other vCPUs are preventing it to exploit all its own reservation. If
> that would have not been the case, there'd be a bug in the scheduler.
>
> By giving the vCPU 100% (either via "budget == period" or with
> extratime), we will figure out if the real-time applications inside can
> actually meet their deadline. If they can't even with such setup, it
> would mean the problem is somewhere else (virtualization overhead, IRQ
> latency, etc).
>
> If the applications can meet their deadline with 100% CPU time, but not
> with 50%, then there are two possibilities:
> 1) they need more than 50%;
> 2) you're having "period synchronization" issues, as Meng was
> describing.
>
> Figuring out if you're on 1, should be as easy as trying to five DomR
> 55%, then 60%, then 65%, etc. And see when it happens that the deadline
> misses are gone forever and for good.
>
> If you can't get rid of deadline misses, it probably means you are in
> case 2, and you need to find a way to make sure that your real-time
> applications inside the domain are able to actually exploit the
> domain's vCPU's budget when it's there. I.e., you don't want them to
> activate when the budget is about to finish, and hence suffer from the
> "blackout" shown in Meng's diagram.
>
> Unfortunately, that's not really trivial to do. If the workload is
> really periodic, it may be enough to find a way to do some kind of
> "calibration" at the beginning. But I'm not sure how robust this will
> actually be.
>
> Perhaps Meng has some more ideas on this as well. :-)

If the RT VCPU has only one RT task on it, we can synchronize the
release time of the VCPU and that of the RT task. In other words, the
release offset of both the VCPU and the RT task are the same in terms
of the wall clock. Then we can assign the task's parameter to the VCPU
and guarantee the task has no deadline miss if the VCPU has no
deadline miss.
However, this observation only works when the assumption that one VCPU
has only one task in the RT domain. I'm not sure how practical it is
because the observation cannot be generalized for multiple tasks on
one VCPU.

Andrii and Dario,
Do you think the assumption that one VCPU runs only one RT task is
reasonable in practice?
If it is, is there some use cases for this assumption?

Thanks,

Meng

_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xenproject.org
https://lists.xenproject.org/mailman/listinfo/xen-devel

Reply via email to