Open Stack

On Wed, 2014-02-26 at 13:20 +0000, Pääkkönen Pekka wrote:
> Hi,
> 
> I am planning OpenStack configuration for a server rack with the
> following spec:
> 
> Dell PowerEdge R820: 
> 
> + 4 processors (Intel Xeon E5-4620), 8 cores/processor
> + 512 GB RAM
> + 8 HDD, 1 TB/disk
> + 2*1Gb, 2*10Gb network cards
> 
> How would it be reasonable to configure controller and compute nodes
> in terms of vCPUs and memory?

Are you talking about making virtual machines out of the above R820
hardware nodes to serve as controllers? Or are you talking about
actually using one or more full R820 nodes as controllers?

> I am planning to use one controller due to shortage of public IP
> addresses, and I am wondering how many compute nodes I should use?

A shortage of public IP addresses should have no effect on your choice
in number of controllers. That is what load balancers are for...

> A use case of interest is execution of database instances in OpenStack
> with a write heavy workload.

Note that no OpenStack service except for Ceilometer has a write-heavy
workload (exception: Keystone's tokens table if you are using the SQL
backend for tokens -- not recommended, use memcache).

I do not recommend putting Ceilometer's database on the same node as you
use for other databases for OpenStack services. I would personally
recommend using MySQL Galera or PostgreSQL for non-Ceilometer OpenStack
service databases, and using MongoDB or a well tuned standard MySQL
master/slave setup for Ceilometer.

> I was hinted in the “Ask OpenStack”-service that 32 GB, 2  vCPUs per
> virtual machine may be a good starting point.

Our controllers at AT&T were 16GB 12-core machines with <1 TB of disk.
But, we did not put databases and message queues on the same machines as
controllers. Our controllers were stateless so that we could expand API
throughput separately from database and message queue throughput.

Best,
-jay