Hi.
I'm trying to configure two or more VMs with 8, 16 or 32 vCPUs on a 4 socket DELL r820 with 64 logical CPUs.
I've configured the first VM with 16 vCPUs with:
cpuid.coresPerSocket = 16
numa.vcpu.preferHT = TRUE
sched.cpu.vsmpConsolidate to TRUE
This makes the VM use one NUMA node and all 8 cores with HT (16 logical CPUs), as the application running in the VM prefers. That said, running only this VM on the host it seems like the ESXi 5.0 hypervisor migrates the VM from node to node. I'm suspecting the NUMA scheduler doing its magic. So I've tried different VM and Host attributes without any luck.
Anyone out there have some experience in this area? Obviously I could set numa.nodeAffinity, but I'm really trying to avoid it! Also disabling the numa scheduler just makes things worse. Any suggestions? My application doesn't use much memory, but it really needs to be local to the current home NUMA node.
Thanks, best regards
Haakon Sporsheim