On Wed, 2021-01-20 at 14:15 +0100, Nagaraj Akkina wrote:
I think we have a big problem, scheduler is able to allocate more memory than actual physical memory in the compute node, the same statistics are also there in database. For example : if we have a physical memory of 125GB in a node, the total memory allocated to all the instances(virsh dommemstat of all vms) in that host is more than physical memory say around 147GB, being ram_allocation_ratio defined as 1 and reserve_host_memory_mb defined as required.
How can nova allocates more memory than actual?
Statistics from database
+-----------+----------------+-------------+----------------------+ | memory_mb | memory_mb_used | free_ram_mb | ram_allocation_ratio | +-----------+----------------+-------------+----------------------+ | 128897 | 154624 | -25727 | 1 | +-----------+----------------+-------------+----------------------+ we are using openstack stein
You're probably losing memory due to huge pages or reserve_host_memory_mb still isn't configured correctly. What does placement report for total inventory and usage? You can find the UUID of the resource provider like so: $ openstack resource provider list Once you have that, you can view available inventory like so: $ openstack resource provider inventory list $RP_UUID You should ensure that allocation_ratio, max_unit and reserved values match what you'd expect, namely, 1.0, the total amount of memory on the host, and the value configured for reserved memory on the host. Once you've verified those, you can view usage like so: $ openstack resource provider usage show $RP_UUID This will summarize the consumption of usage by all allocations against this resource provider. Finally, you can view the individual breakdown of resources by inspecting the inventory consumed for each instance on the host, like so: $ openstack server list --host devstack-1 -c ID # for each server $ openstack resource provider allocation show $SERVER_UUID Hope this helps, Stephen
Regards, Akkina