Multiple physical CPUs performance

[Rapid e-Suite]

Is there any special host,guest OS other than default need to be change to have VirtualBox provide optimal CPUs performance to guest when host have more than one physical CPUs?

We have problem like not getting optimal performance when allocated more than half of number of vCPUs available on the host to a guest. We have totally 12 cores from 2 x physical CPUs on the host which specification are as below, when we we allocated more than 6 vCPUs to the guest VM will have a weird behavior.

I'm installing Fusion Applications 11.1.7 on Oracle Linux 6.5 in VirtualBox 4.3.6 VM and host is also Oracle Linux 6.5 x86_64. I have a guest VM with Fusion IDM installed and can be started up properly when the guest VM get only 6 x vCPUs allocated. But when I try with higher number of vCPUs, the guest VM will sometimes can't finish OS start up, it look like hanging due to CPU issue somehow as I noticed CPU usage by the VM will be very high when it hang. Sometimes, the guest VM which number of vCPUs more than 6 cores be able to started properly, but application (Fusion IDM) inside the VM won't start up properly and will notice there is CPU usage still high in the guest. And there is only on guest running on the host at the time.

Below host specifications are:

CPU: 2 x E5-2620 0 @ 2.00GHz ( Hyper-Thread is disabled).
Motherboard: Supermicro, X9DR3-F
RAM: 10 x Kingston Buffered ECC DDR-3 SDRAM 16GB PC-1600 = 160GB RAM total.l
DISK: VM is located on Mdadm RAID0 of 2 x WD Black 2 TB SATA Hard Drives ( WD2002FAEX).

Guest configurations:

vCPU: 6 - 10
RAM: 144GB

[martyscholes]

I noticed similar issues on Opteron CPUs on a Solaris host whenever we allocate more cores to a guest than exist on a single socket. I suspect that this is related to the amount of inter-cpu traffic (HT on Opteron, QPI on Xeon) that takes place. The working theory is that on a NUMA host when CPUs are allocated to a guest, the guest sees them as SMP, treating them all equally. This works fine when all of the cores are running on the same socket, accessing the same memory, but when the core count must span sockets, the amount of cross-socket traffic explodes.

Or so I tell myself. In any event, we keep the guest core count below the number of cores on a single socket, for this very reason.