On 29/9/06 14:28, "Xu, Anthony" <anthony.xu@xxxxxxxxx> wrote:
>> 5. LTP testing might run very slow in SMP VTI Domain with credit scheduler.
>> If binding VTI
>> and Xen0 vcpu, this bug won't be there.
>
> Hi keir,
>
> In credit scheduler, two vcpus in the same domain may be scheduled on the same
> CPU. For instance, vcpu0 and vcpu1 are running on the same CPU, vcpu0 is doing
> spin_lock in guest, then time slice is due, vcpu0 is scheduled out before
> doing spin_unlock, vcpu1 is scheduled in, vcpu1 is also trying to get the same
> spin lock, yes vcpu1 can't get this lock due to this lock has been got by
> vcpu0, so vcpu1 tight loop until its time slice is due, so this domain is very
> slow.
> Of cause, even vcpu0 and vcpu1 are running on different CPUs, there is still
> similar situation, but the impact is less.
>
> What do you think about this?
IBM did some tests at one point which concluded that we'd get little or no
benefit from paravirtualising spinlocks to yield-on-spin. This is mainly
because spinlock critical regions are usually very small. The chances of
preempting someone while they are in a spinlock critical region, *and*
scheduling someone else who wants to take the same lock, *and* not having
another CPU idle enough to run the lock-holding VCPU, should be very small.
-- Keir
_______________________________________________
Xen-ia64-devel mailing list
Xen-ia64-devel@xxxxxxxxxxxxxxxxxxx
http://lists.xensource.com/xen-ia64-devel
|
Home