Came into this session late…apologies for the rather fragmented notes.

• When Normal Behavior isn’t good….
  • Probabilistic vs. absolute health model
• VC Ops has a domain-specific health model for vSphere
• Use absolute health noel until analytics’ probabilistic model is ready to generate DT’s
• 2 Goals of capacity management
  • Efficiency — optimization of capacity.
  • Predictability — availability of capacity.
• How to think about capacity
  • Look at usable capacity, not total capacity (account for failover, extra buffer, HA failover, etc.).
  • Think in terms of VMs…you deploy VMs — not CPU or memory.
    • Deploy CPU/memory/disk/etc. in “VM-sized units”.
• Understanding behavior
  • Understand weekly patterns — business week, weekend, workload spike at  9 AM on Mondays.
• Optimizing capacity
  • Multiple levels — VMs, hosts, clusters
  • Powered off and idle VMs
  • Right-sizing — yes, please
• How do we determine an idle VM?
  • Simple…when VM usage is less than a configured threshold.
  • But…when is it idle overall? Time-based over xx% over xx time period.
    • Two thresholds — VM idle or not, VM idle how much over last week/month/etc.
• Seriously down in the weeds stuff (in a good way)….walking through when VMs are idle and/or over utilized vs. underutilized.
• Summary
  • Virtualization and cloud present new challenges.
    • Fuild Capacity, invisible walls, separation of provider and consumer.
  • Must move beyond manual monitoring and troubleshooting.
    • Most performance management solutions today require extensive manual effort.
    • Doesn’t scale well with cloud environments.
  • One more point….

A lot of deep down stuff that I didn’t blog as was difficult to turn into a good post….