Open Stack

Hi OpenStack Devs,

This is Wei Du from Yahoo! OpenStack engineering team. Currently, we have some data about the actual resource usage seen by ‘proc’ on each compute node. However, the actual available resources cannot be derived from the actual usage. For example, we have a 10G RAM on a hypervisor, the current free RAM recorded in nova.compute_nodes is -3G. By running ‘proc’ on the compute node, we see the actual RAM usage is only 3G. With these data point, we cannot say the available RAM is 7G. Assume the ram_allocation_ratio is 1.5. while the correct state seen by scheduler is 2G RAM is available.

Physical RAM: 10G
Free RAM: -3G
Actual used RAM: 3G

ram_allocation_ratio: 1.5

Available RAM seen by scheduler: 10*1.5-(10-(-3)) = 2

We are planning to implement a nova api extension, which will check nova.compute_nodes for the current resource allocation state, and return the current available resources seen by nova-scheduler. We would like to get your experience/suggestions/thoughts if you happen to encounter similar situation.

Thanks.
Wei
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.openstack.org/pipermail/openstack-dev/attachments/20140506/108ccf86/attachment.html>