Hello, we had to configure the metadata in the flavors, in order to distribute the memory usage across the two NUMA nodes (and hugepages). { name = "general.2xlarge", ram = "327680", cpu = "96", disk = "80", extra_specs = { "hw:mem_page_size" = "2MB", "hw:numa_nodes" = "2" } }, El 17/10/25 a las 04:12, brijesh.mishra@five9.com escribió:
Hi Team
We have use Kolla ansible for openstack Caracal 2024.1 deployment. Having 2 production sites, we have observer that the Nova scheduler is allocating virtual machines (VMs) only on a single NUMA socket of the compute host. Since the flavor does not include the property hw:numa_nodes='2', vCPUs spread on two NUMA nodes scheduling is not enabled. As a result, all vCPUs and memory for a VM are placed on a single NUMA node, while the other NUMA socket(s) remain underutilized. So we tested this property in our test flavor and found that now scheduler is able to spread vcpu/vm across two numa sockets but performance on memory side is degraded for about 20% to 30%. So we are looking for suggestions on this if there is better way to accommodate more VMs or spread vcpu on two numa socket without having performance issues on memory or cpu side, or something which have less impact?
Regards Brijesh