[Libvir] Extending libvirt to probe NUMA topology

[Daniel Veillard]

On Wed, Jun 13, 2007 at 10:40:40AM -0500, Ryan Harper wrote:
> Hello all,
> 
> I wanted to start a discussion on how we might get libvirt to be able to
> probe the NUMA topology of Xen and Linux (for QEMU/KVM).  In Xen, I've
> recently posted patches for exporting topology into the [1]physinfo
> hypercall, as well adding a [2]hypercall to probe the Xen heap.  I
> believe the topology and memory info is already available in Linux.
> With these, we have enough information to be able to write some simple
> policy above libvirt that can create guests in a NUMA-aware fashion.

  Let's restart that discussion, I would really like to see this
implemented within the next month.

> I'd like to suggest the following for discussion:
> 
> (1) A function to discover topology
> (2) A function to check available memory
> (3) Specifying which cpus to use prior to domain start
> 
> Thoughts?

  Okay following the discussions back in June and what seems available
as APIs on various setups I would like to suggest the following:

1) Provide a function describing the topology as an XML instance:

   char *	virNodeGetTopology(virConnectPtr conn);

which would return an XML instance as in virConnectGetCapabilities. I
toyed with the idea of extending virConnectGetCapabilities() to add a
topology section in case of NUMA support at the hypervisor level, but
it was looking to me that the two might be used at different times
and separating both might be a bit cleaner, but I could be convinced
otherwise. This doesn't change much the content in any way.
I think the most important in the call is to get the topology informations
as the number of processors, memory and NUMA cells are already available
from virNodeGetInfo(). I suggest a format exposing the hierarchy in the
XML structure, which will allow for more complex topologies for example
on Sun hardware:

---------------------------------
<topology>
  <cells num='2'>
    <cell id='0'>
      <cpus num='2'>
        <cpu id='0'/>
        <cpu id='1'/>
      </cpus>
      <memory size='2097152'/>
    </cell>
    <cell id='1'>
      <cpus num='2'>
        <cpu id='2'/>
        <cpu id='3'/>
      </cpus>
      <memory size='2097152'/>
    </cell>
  </cells>
</topology>
---------------------------------

  A few things to note:
   - the <cells> element list the top sibling cells

   - the <cell> element describes as child the resources available
     like the list of CPUs, the size of the local memory, that could
     be extended by disk descriptions too
     <disk dev='/dev/sdb'/>
     and possibly other special devices (no idea what ATM).

   - in case of deeper hierarchical topology one may need to be able to
     name sub-cells and the format could be extended for example as
     <cells num='2'>
       <cells num='2'>
         <cell id='1'>
           ...
         </cell>
         <cell id='2'>
           ...
         </cell>
       </cells>
       <cells num='2'>
         <cell id='3'>
           ...
         </cell>
         <cell id='4'>
           ...
         </cell>
       </cells>
     </cells>
     But that can be discussed/changed when the need arise :-)

   - topology may later be extended with other child elements 
     for example to expand the description with memory access costs
     from cell to cell. I don't know what's the best way, mapping
     an array in XML is usually not very nice.

   - the memory size is indicated on an attribute (instead as 
     the content as we use on domain dumps), to preserve extensibility
     we may need to express more structure there (memory banks for
     example). We could also add a free='xxxxx' attribute indicating the
     amount available there, but as you suggested it's probably better
     to provide a separate call for this.
     
I would expect that function to be available even for ReadOnly connections
since it's descriptive only, which means it would need to be added to the
set of proxy supported call. The call will of course be added to the driver
block. Implementation on recent Xen could use the hypercall. For KVM I'm
wondering a bit, I don't have a NUMA box around (but can probably find one),
I assume that we could either use libnuma if found at compile time or
get informations from /proc. On Solaris there is a specific library as
Dan exposed in the thread. I think coming first with a Xen only support
would be fine, others hypervisors or platforms can be added later.

2) Function to get the free memory of a given cell:

   unsigned long virNodeGetCellFreeMemory(virConnectPtr conn, int cell);

that's relatively simple, would match the request from the initial mail
but I'm wondering a bit. If the program tries to do a best placement it
will usually run that request for a number of cells no ? Maybe a call
returning the memory amounts for a range of cells would be more appropriate.

3) Adding Cell/CPU placement informations to a domain description

That's where I think things starts to get a bit messy, it's not that
adding 
   <cell>1</cell>
or
   <cpus>
     <pin vcpu='0' cpulist='2,3'/>
     <pin vcpu='1' cpulist='3'/>
   </cpus>
along
   <vcpu>2</vcpu>

would be hard, it's rather what to do if the request can't be satisfied.
Basically I still think that the hypervisor is in a better position to 
do the placement, and doing the requirement here breaks:
   - the virtualization, the more you rely on the physical
     hardware property the more you loose the benefits of virtualizing
   - if CPU 2 and 3 are not available/full or if the topology changed since
     the domain was saved the domain may just not be able to run, or run
     worse than if nothing had been specified.
CPU pinning at runtime means a dynamic change, it's adaptbility and makes
a lot of sense. But saving those dynamic instant values in the process
description sounds a bit wrong to me, because the context which led to them
may have changed since (or may just not make sense anymore, like after a
migration or hardware change).
Anyway I guess that's needed, I would tend to go the simplest way and
just allow to specify the vcpu pinning in a very explicit way and hence
mapping directly to the kind of capabilities already available in
virDomainPinVcpu() with a similar cpumap syntax as used in virsh vcpupin
command (i.e. comma separated list of CPU numbers).

   <cpus>
     <pin vcpu='0' cpulist='2,3'/>
     <pin vcpu='1' cpulist='3'/>
   </cpus>

If everyone agrees with those suggestions, then I guess we can try to get
a first Xen-3.1 based implementation

Daniel

-- 
Red Hat Virtualization group http://redhat.com/virtualization/
Daniel Veillard      | virtualization library  http://libvirt.org/
veillard at redhat.com  | libxml GNOME XML XSLT toolkit  http://xmlsoft.org/
http://veillard.com/ | Rpmfind RPM search engine  http://rpmfind.net/