[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: [PATCH] allow vcpu to move between pcpus (was Re: [Xen-ia64-devel] credit sched on ia64)
Le Lundi 03 Juillet 2006 06:22, Isaku Yamahata a écrit :
> This patch allows vcpu to move between pcpus.
> I couldn't boot domU on credit scheduler without this patch.
> This patch is only for correctness. This patch isn't efficient.
The obvious question is why not using _dirty_cpumask; but the obvious reason
is of course performance issues :-) So I *definitly* agree with your patch.
The patch is not that inefficient: only vcpu migration is inefficent !
> Maybe a consideration on VHPT per pcpu is needed.
> - allocation of region id per domain.
> Should rid region be allocated per vcpu?
* lack of rids as you noted
* linux assumes rid can be shared: if rid are allocated per vcpu, process
migration can be expansive.
> By this, it would be unnecessary to flush when vcpu movement.
> But I think this isn't very preferable because the size of region id
> given to vcpu would be too small.
> - Currently allocation of VHPT per physical cpus.
> Allocate VHPT per vcpu like vt-i domain?
Vcpu migration is a big pro for per VHPT vcpu, but NUMA is a cons :-)
> I think this is preferable, but some measuremeat should back this.
> - other?
One may ask wether or not vcpu migration is interesting wrt NUMA. I think for
NUMA machines, vcpu migration must go with memory migration.
Also, note that __cacheline_aligned_in_smp is defined as nothing on xen/ia64.
You'd better either not to use it or define it.
Xen-ia64-devel mailing list