On Sun, Aug 05, 2012 at 12:30:49PM -0700, Eric Northup wrote:
> On Sun, Aug 5, 2012 at 5:58 AM, Gleb Natapov <[email protected]> wrote:
> > APIC code has a lot of checks for apic presence and apic HW/SW enable
> > state. Most common configuration is when each vcpu has in kernel apic
> > and it is fully enabled. This path series uses jump labels to turn checks
> > to nops in the common case.
>
> What is the target workload and how does the performance compare?
Since those patches micro optimize interrupt injection path workload
with a lot of interrupts and a lot of cpus is the target workload. I
ran a simple test that transfers big file from gust (with 16 vcpus)
to a host. The workload generates ~40000 interrupt per second (this is
with userspace networking). Using perf I checked how functions affected
by the change perform with and without the patches:
Without:
0.85% kvm_apic_present
0.45% __apic_accept_irq
0.30% kvm_apic_has_interrupt
0.20% kvm_irq_delivery_to_apic
0.16% apic_has_pending_timer
0.09% apic_mmio_write
0.04% kvm_apic_accept_pic_intr
0.04% kvm_lapic_get_cr8
0.04% kvm_lapic_enabled
With:
0.61% kvm_irq_delivery_to_apic
0.49% __apic_accept_irq
0.12% apic_has_pending_timer
0.11% kvm_apic_has_interrupt
0.04% kvm_apic_accept_pic_intr
0.03% apic_mmio_write
0.02% kvm_lapic_get_cr8
> As
> a naive question, how different is it than just using gcc branch
> hints?
>
If I add likely() annotations to kvm_apic_present() gcc generates
exactly same code.
--
Gleb.
--
To unsubscribe from this list: send the line "unsubscribe kvm" in
the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
