CapacityPlanningForOpenstack.md

capacity planning

## Capacity Planning Notes


#### Is your target Public cloud or private cloud?
- Private Cloud
  - easier
  - you know your workload and purpose
  - Fewer variables
  - Cost / Revenue is an important consideration
  - often single tenant
  - Provisioning normallly easier / slower churn

- Public
  - Designing for a generic use case and workload.

- Important Openstack Considerations
  - network model
  - neutron nodes vs. DVR
  - Assuming 10G Networking?
  - Floating IP's (None Now)
  - Future SDN provider?
  - Are we providing HA Compute?
    - Trade off is performance.

  - Volumes are mounted from SAN?

#### Hopes and dreams
  - We want to encourage every project being developed from the ground up for redundancy.

#### Questions For US
  - What configuration management and provisioning tools are in use besides Bosh.
  - What other monitoring tools are in place in the organization besides waht I've seen in the Grafana stack.
  - What hypervisor is being used? KVM? HyperV?
  - Overcommit fractions * physical cores / virtual cores per instance to figure out how many VMS of whatever flavor we can host on a compute node.

#### Questions for BC
- Look up performance characteristics of different VCPU to core ratios.

#### Cloud Controller Notes
- Box that hosts yoru db/messageque/and endpoits
- Is our CC HA?
- How active is use of the OS API?

#### Neutron notes?
- How overloaded are the network nodes? (not very)

#### Thoughts  

- getting good and reliable data in a virtual environment can be a tremendous effort.

- assume cloud users will follow similar patters to what we've witnessed already? But this is not a guarantee. so we need to seperate forecaasts in to behavior-driven and process driven forecasts.

- CF planning and service planning, different dashboards?

- How well known are high utilization periods
  - Make a list per app
  - per service
  - note the statistics of the app (framework/end user/etc.. )
  - for network utilization concerns we can collocate apps with different utilization periods on the same platform.

- load tests new apps to get usage patterns, ensure dashboard and metrics are being collected during the test, we can get a footprint here.

- real time isn't the key,  historical and aggregated data is what matters.
  - but being able to report on and spin up capacity quickkly is important
  - our work would be the basis for autoscaling activities in the platform

- most infrastructures are under peak load for under an hour or 2 per day, a widget that displays the peak load hour over time would be interesting.

- How about a visualizaton that shows peak usage windows per application
  - over time
  - on average
  - network utilization
  - memory and cpu utilization?
  - can we even get at this reliably?

- do we care about quotas as well as capacity? yes,

- discriminatory pricing (lookup airline analogy)

- difference between bare metal, vm, and container, start from top or start from bottom?

- we need to come up with a list of forecasting models.

- Do we have any projections for future utilization, 6 mo? 12mo?

- Overcommit Ratio * physical cores / virtual cores per instance to figure out how many VMS of whatever flavor we can host on a compute node. We could put this in the dashboard.

- assign a cost to a flavor and we have value to display

- We could theorically show how many containers we could provide for at a given point it time.

- We could show if we have a full AZ available red/green button.

- Make sure ephemeral storage plan has enough capacity to support numbers from the VM capacity number.

-  If we know what flavor is used for the the Container host this would be an interesting calculation WRT CF.

- Mixing lots of flavors can hurt performance under KVM, something about time slice utilization. Maybe a flavor variance number?

- Do we have Ceilometer Available in Juno?

- The number of variables in play is huge, we need to keep it simple to start.

- KVM with AMD overcommit 2 or 4:1 gives great density but you won't have the iops to support all of those VMS hitting local storage.  References the hardware from my opentsack notes.

- Should we run Unix Bench on a test stack? Has there every been a stress test?

- Putting Like VM's on like compute nodes? Flavor + nova Schedular can ensure flavor goes to a specific host aggregate.

- How do we get meaningful metrics for network utiilization from OS? Baremetal? An applicatoin that is a resource hog isn't necessarily one with a very active API. How do we segregate these?

- what ways can we assign dollar values to dashboard?
  - Want to show totaal percentage of aggregate resources in use by a given app? CF? Bosh job?

https://www.ctl.io/blog/post/capacity-planning-cloud-platform/