[vbox-dev] CPU detection problem

[Klaus Espenlaub]

Laura,

it turned out to be a core counting bug (in code which is used only on 
Linux), it got things wrong on such big systems, effectively it counted 
only one socket. Slipped through for a while, but was fixed shortly 
after 4.3.6 went out, so will be in 4.3.8 (no release date known).

"VBoxManage list hostinfo" and the /proc/cpuinfo details made this quite 
clear...

Klaus

On 30.01.2014 20:17, Klaus Espenlaub wrote:
> Hi Laura,
>
> On 30.01.2014 16:55, Baxter, Laura wrote:
>> Hi,
>>
>> I'm running VirtualBox version 4.3.6 r91406.
>>
>> The basic problem I am having is that although my (Linux) server has
>> 64 CPUs, (4 nodes each with 16 CPUs/8cores with 2 threads per core),
>> my VM is only allowing me to set 8 processors (max).
> VirtualBox can definitely deal with multi-socket systems (has been
> tested on 8 socket systems with 10 cores with 2 threads per core), so we
> need to get a bit more detail about the reality (the output of "cat
> /dev/cpuinfo") and what VirtualBox thinks (the output of "VBoxManage
> list hostinfo"), as the combination will give us a lot of hints what
> information might be wrong. lscpu is a very useful command to get a
> quick overview, but we need the lower level details. Actually the lscpu
> output you provided somewhat contradicts your claims above, you have a
> system with 8 nodes with 8 cores each - AMD doesn't do hyperthreading.
>> My question is, how can i get my VM to "see" all 64 nodes, to increase
>> the VM's processing power?
> Careful with the assumption "more VCPUs = faster VM" - even if the
> workload on real hardware scales very well, it might be different in a
> VM, as the I/O overhead is much higher, and the cross-VCPU
> synchronization can be extremely expensive.
>> Is this an issue with NUMA not being supported/implemented?
> NUMA as such isn't supported (as far as optimizing memory placement is
> concerned), but VirtualBox should definitely be able to use all CPUs and
> the entire memory.
>> Below are some details of the VM and the server, I hope someone has
>> some ideas to help, i wonder if this issue affects other users?
> We should be able to figure out what's going on... it's definitely not
> affecting everyone (and as you're using Linux it's definitely not the
> issue that modern Windows versions pretend that AMD does hyperthreading,
> assuming that the CPUs in the systems are actually threads, and that the
> core count is half the thread count).
>
> Klaus
>>
>> Many thanks,
>>
>> Laura