[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [Xen-devel] VT-d scalability issue



Not regarding the other questions/objections in this thread for a
moment --- what kind of performance improvments are we talking of here
if the vcpus are pinned?  Is it close to 1 VM or is there still some
performance degradation due to IOTLB pressure?

[And talking of IOTLB pressure, why can't Intel document the IOTLB
sizes in the chipset docs?  Or even better, why can't these values be
queried from the chipset?]

        eSk


[Edwin Zhai]
> Keir,
> I have found a VT-d scalability issue and want to some feed backs.

> When I assign a pass-through NIC to a linux VM and increase the num
> of VMs, the iperf throughput for each VM drops greatly. Say, start 8
> VM running on a machine with 8 physical cpus, start 8 iperf client
> to connect each of them, the final result is only 60% of 1 VM.

> Further investigation shows vcpu migration cause "cold" cache for
> pass-through domain.  following code in vmx_do_resume try to
> invalidate orig processor's cache when 14 migration if this domain
> has pass-through device and no support for wbinvd vmexit.

> 16 if ( has_arch_pdevs(v->domain) && !cpu_has_wbinvd_exiting )
> {
>     int cpu = v->arch.hvm_vmx.active_cpu;
>     if ( cpu != -1 )
>         on_selected_cpus(cpumask_of_cpu(cpu), wbinvd_ipi, NULL, 1,

> }

> So we want to pin vcpu to free processor for domains with
> pass-through device in creation process, just like what we did for
> NUMA system.

> What do you think of it? Or have other ideas?

> Thanks,


> -- 
> best rgds,
> edwin

> _______________________________________________
> Xen-devel mailing list
> Xen-devel@xxxxxxxxxxxxxxxxxxx
> http://lists.xensource.com/xen-devel


_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxxxxxxxx
http://lists.xensource.com/xen-devel


 


Rackspace

Lists.xenproject.org is hosted with RackSpace, monitoring our
servers 24x7x365 and backed by RackSpace's Fanatical Support®.