BCA1902: Virtualizing Exchange 2010

This session focused on virtualizing Exchange 2010 on vSphere 4.x or 5.x. Highlights of this session include:

  • Bottom line there’s no reason why you should be afraid of virtualizing Exchange 2010.
  • Exchange 2010 now uses 32Kb I/O blocks, vs. 8K for previous versions.
  • Exchange 2010 I/O is much more optimized than previous versions
  • VMware fully supports share-nothing clustering with nearly all ESXi features (HA, DRS, etc.). Share-nothing clustering is used by Exchange 2010 and SQL database mirroring.
  • You can virtualize all Exchange 2010 roles
  • You can combine DAGs with HA, vMotion, DRS, Fibre Channel, FCoE and iSCSI
  • VMDKs must be thick provisioned, not thin. Should use EZT VMDKs.
  • Not supported is NFS for Exchange data or VM snapshots (for roll-back purposes, backup is fine).
  • VMware internally uses standard load performance tools like jetstress and loadgen (more so loadgen)
  • On vSphere 5.0 you can average 1000 users per vCPU/pCore, and linearly increases to 12K users
  • Fibre Channel has the best performance, but iSCSI is fine too
  • 2-7% CPU overhead vice physical hardware
  • No I/O latency impact on virutalization
  • Best practices include
    • vCPUs <= pCores
    • Exchange is not NUMA aware so keep VM size less than NUMA node size
    • Use the Exchange processor query tool to determine users per core estimates
    • Use the Exchange mailbox role calculator for storage/network calculations
    • DO NOT over commit resources such as memory or vCPUs
    • Use LSI logic SCSI adaptor unless you’ve already standardized on the pvscsi driver
    • Use multiple vSCSI adaptors and distribute the mailbox/log load across them
    • Only use RDMs if your hardware storage array requires them to do VSS snapshots, otherwise VMDKs are perfectly fine.
    • For DRS keep the VMs smaller and ensure EVC mode is enabled
    • Enable HA for all VMs, use host admission control, and enable VM monitoring
    • Utilize host DRS groups and VM DRS groups
      • DAGs must be on separate nodes
      • “should run on” for all other roles
    • For vMotion set the clusterheartbeat setting to 2000ms, up from 1000ms default if you aren’t using jumbo frames
    • Use the Exchange profile analyzer tool
    • Design Questions: Dedicated or multi-role VMs? How much HA? DB size? Backups?
    • Deploy dedicated mailbox VMs, but can easily use other combos like HUB/CAS.
    • Processor selection has a major impact on users per core. (.e.g. Intel x5470 vs. x5660 shows a dramatic reduction in CPU utilization).
    • How big should your page file be? Check KB 889654 to reduce page size
    • Monitor % CPU RDY, KAVG, DAVG and GAVG ESXi counters

Whew…the speaker covered a lot of ground. In a nutshell Exchange 2010 runs extremely well on ESXi, but you do need to be aware of the tweaks/best practices when deploying it on vSphere. Professional services organizations that specialize in Exchange engagements really need to understand the various hypervisors and best practices, or you could run into some issues or cause customers problems.

Print Friendly, PDF & Email

Related Posts

Subscribe
Notify of
0 Comments
Inline Feedbacks
View all comments