Re: [Qemu-devel] [RFC:kvm] export host NUMA info to guest & make emulated device NUMA attr

[Andrew Theurer]

On 05/22/2012 04:28 AM, Liu ping fan wrote:
> On Sat, May 19, 2012 at 12:14 AM, Shirley Ma<address@hidden>  wrote:
> > On Thu, 2012-05-17 at 17:20 +0800, Liu Ping Fan wrote:
> > > Currently, the guest can not know the NUMA info of the vcpu, which
> > > will
> > > result in performance drawback.
> > >
> > > This is the discovered and experiment by
> > >          Shirley Ma<address@hidden>
> > >          Krishna Kumar<address@hidden>
> > >          Tom Lendacky<address@hidden>
> > > Refer to -
> > > http://www.mail-archive.com/address@hidden/msg69868.html
> > > we can see the big perfermance gap between NUMA aware and unaware.
> > >
> > > Enlightened by their discovery, I think, we can do more work -- that
> > > is to
> > > export NUMA info of host to guest.
> >
> > There three problems we've found:
> >
> > 1. KVM doesn't support NUMA load balancer. Even there are no other
> > workloads in the system, and the number of vcpus on the guest is smaller
> > than the number of cpus per node, the vcpus could be scheduled on
> > different nodes.
> >
> > Someone is working on in-kernel solution. Andrew Theurer has a working
> > user-space NUMA aware VM balancer, it requires libvirt and cgroups
> > (which is default for RHEL6 systems).
> Interesting, and I found that "sched/numa: Introduce
> sys_numa_{t,m}bind()" committed by Peter and Ingo may help.
> But I think from the guest view, it can not tell whether the two vcpus
> are on the same host node. For example,
> vcpu-a in node-A is not vcpu-b in node-B, the guest lb will be more
> expensive if it pull_task from vcpu-a and
> choose vcpu-b to push.  And my idea is to export such info to guest,
> still working on it.

The long term solution is to two-fold:
1) Guests that are quite large (in that they cannot fit in a host NUMA 
node) must have static mulit-node NUMA topology implemented by Qemu. 
That is here today, but we do not do it automatically, which is probably 
going to be a VM management responsibility.
2) Host scheduler and NUMA code must be enhanced to get better placement 
of Qemu memory and threads.  For single-node vNUMA guests, this is easy, 
put it all in one node.  For mulit-node vNUMA guests, the host must 
understand that some Qemu memory belongs with certain vCPU threads 
(which make up one of the guests vNUMA nodes), and then place that 
memory/threads in a specific host node (and continue for other 
memory/threads for each Qemu vNUMA node).

Note that even if a guest's memory/threads for a vNUMA node are 
relocated to another host node (which will be necessary) the NUMA 
characteristics of guest are still maintained (as all those vCPUs and 
memory are still "close" to each other).

The problem with exposing the host's NUMA info directly to the guest is 
that (1) vCPUs will get relocated, so their topology info in the guest 
will have to change over time. IMO that is a bad idea.  We have a hard 
enough time getting applications to work with a static NUMA info.  To 
get applications to react to changing NUMA topology is not going to turn 
out well. (2) Every single guest would have to have the same number of 
NUMA nodes defined as the host.  That is overkill, especially for small 
guests.
> > 2. The host scheduler is not aware the relationship between guest vCPUs
> > and vhost. So it's possible for host scheduler to schedule per-device
> > vhost thread on the same cpu on which the vCPU kick a TX packet, or
> > schecule vhost thread on different node than the vCPU for; For RX packet
> > it's possible for vhost delivers RX packet on the vCPU running on
> > different node too.
> Yes. I notice this point in your original patch.
>
> > 3. per-device vhost thread is not scaled.
> What about the scale-ability of per-vm * host_NUMA_NODE? When we make
> advantage of multi-core,  we produce mulit vcpu threads for one VM.
> So what about the emulated device? Is it acceptable to scale to take
> advantage of host NUMA attr.  After all, how many nodes on which the
> VM
> can be run on are the user's control.  It is a balance of
> scale-ability and performance.
>
> > So the problems are in host scheduling and vhost thread scalability. I
> > am not sure how much help from exposing NUMA info from host to guest.
> >
> > Have you tested these patched? How much performance gain here?
> Sorry, not yet.  As you have mentioned, the vhost thread scalability
> is a big problem. So I want to see others' opinion before going on.
>
> Thanks and regards,
> pingfan
>
>
> > Thanks
> > Shirley
> >
> > > So here comes the idea:
> > > 1. export host numa info through guest's sched domain to its scheduler
> > >    Export vcpu's NUMA info to guest scheduler(I think mem NUMA problem
> > >    has been handled by host).  So the guest's lb will consider the
> > > cost.
> > >    I am still working on this, and my original idea is to export these
> > > info
> > >    through "static struct sched_domain_topology_level
> > > *sched_domain_topology"
> > >    to guest.
> > >
> > > 2. Do a better emulation of virt mach exported to guest.
> > >    In real world, the devices are limited by kinds of reasons to own
> > > the NUMA
> > >    property. But as to Qemu, the device is emulated by thread, which
> > > inherit
> > >    the NUMA attr in nature.  We can implement the device as components
> > > of many
> > >    logic units, each of the unit is backed by a thread in different
> > > host node.
> > >    Currently, I want to start the work on vhost. But I think, maybe in
> > >    future, the iothread in Qemu can also has such attr.
> > >
> > >
> > > Forgive me, for the limited time, I can not have more better
> > > understand of
> > > vhost/virtio_net drivers. These patches are just draft, _FAR_, _FAR_
> > > from work.
> > > I will do more detail work for them in future.
> > >
> > > To easy the review, the following is the sum up of the 2nd point of
> > > the idea.
> > > As for the 1st point of the idea, it is not reflected in the patches.
> > >
> > > --spread/shrink the vhost_workers over the host nodes as demanded from
> > > Qemu.
> > >    And we can consider each vhost_worker as an independent net logic
> > > device
> > >    embeded in physical device "vhost_net".  At the meanwhile, we spread
> > > vcpu
> > >    threads over the host node.
> > >    The vrings on guest are allocated PAGE_SIZE align separately, so
> > > they can
> > >    will only be mapped into different host node, so vhost_worker in the
> > > same
> > >    node can access it with the least cost. So does the vq on guest.
> > >
> > > --virtio_net driver will changes and talk with the logic device. And
> > > which
> > >    logic device it will talk to is determined by on which vcpu it is
> > > scheduled.
> > >
> > > --the binding of vcpus and vhost_worker is implemented by:
> > >    for call direction, vq-a in the node-A will have a dedicated irq-a.
> > > And
> > >    we set the irq-a's affinity to vcpus in node-A.
> > >    for kick direction, kick register-b trigger different eventfd-b
> > > which wake up
> > >    vhost_worker-b.
-Andrew Theurer