Exadata VM CPU Pinning
(Doc ID 2421827.1)
Last updated on JULY 24, 2018
Applies to:Oracle Exadata Storage Server Software - Version 184.108.40.206.0 to 220.127.116.11.0 [Release 12.1 to 12.2]
Information in this document applies to any platform.
OVM by default has CPU hard paritioning based on socket. This will limit VM vCPU expansion within socket and thus lead to cpu stealing.
The issue is evident in below example.
In Exadata VM environment customers are allowed to create many VMs with different vCPU configurations.
As an example consider below Exadata VM.
It has 4 VMs with vCPU count 32/24/20/16 + 4 for Dom0 on each DatabaseNode, where physical CPU count is 96 total. (48 on each socket). Due to default hard paritioning in OVM, VMs with 32 and 24 vCPUs may end up in socket1 and VMs with 20 and 16 vCPUs in Socket 2.
Now socket 1 is over provisioned and socket2 is under provisioned.
Under load you may see CPU stealing on the VMs in socket1 due to over provisioning and may lead to CPU stall and eventually Node(VM) kill due to CSS hang. The document explains a manual procedure to perform CPU pinning of the VMs so that they are always distributed evenly across sockets and not have stealing.
To view full details, sign in with your My Oracle Support account.
Don't have a My Oracle Support account? Click to get started!