Re: [Qemu-devel] [PATCH V8 00/11] Add support for binding guest numa nodes to host numa nodes

[Wanlong Gao]

On 08/21/2013 05:00 PM, Paolo Bonzini wrote:
> Il 21/08/2013 03:22, Wanlong Gao ha scritto:
>> On 08/20/2013 09:43 PM, Paolo Bonzini wrote:
>>> Il 20/08/2013 03:07, Wanlong Gao ha scritto:
>>>>  -numa node,nodeid=0,cpus=0, \
>>>>  -numa mem,size=1024M,policy=membind,host-nodes=0-1 \
>>>>  -numa node,nodeid=1,cpus=1 \
>>>>  -numa mem,size=1024M,policy=interleave,host-nodes=1
>>>
>>> What nodes would the memory be in, for this command line?  Does it just
>>
>> The original concept here is that if the nodeid is omitted, it will be
>> set node by node from node0. Here I also keep the original concept, so the
>> memory will be in node0 and node1.
>>
>>> compute the total and split it evenly across the nodes (so that the
>>> "-numa node" options could omit nodeid and cpus too)?
>>
>> If no memory size is given for any nodes, the memory will split across
>> all nodes like (ram_size / nb_numa_nodes). And yes nodeid and cpus options
>> can also be omitted from the original concept.
>>
>>>
>>> Also, do you still need a "-m" option if you use "-numa mem"?
>>
>> The "-m" options will be used to compute the memory size of each node
>> if the memory size of each node is not set by "-numa mem" option. This
>> is also be consistent with the original concept.
> 
> Ok.  You didn't answer my exact question though---could you run the
> above command line without "-m 2048"?

Sorry :(
Sure not. And if the memory size assigned by "-m" option is not equal
the total memory size assigned by "-numa mem,", the ACPI will write the
wrong table, then the guest kernel will detect this and ignore the wrong
ACPI numa table and assume that all of the memory are belong to one node.

Would you like to run without "-m" option in mind?

Thanks,
Wanlong Gao

> 
> (Of course with hotplug like in Igor's case you'll still need "-m
> maxmem=4096,slots=4" or something like that)
> 
> Paolo
>