View
2.439
Download
2
Embed Size (px)
DESCRIPTION
Citation preview
Scott Schnoll
Exchange Server 2013 Virtualization Best Practices
Agenda• Does virtualization make sense for you?• Supported/Unsupported virtualization features• Sizing recommendations for virtualized
Exchange deployments• Common problem areas and how to avoid them
Customers virtualize Exchange because of
• Internal standardization of virtualization platform• Deployment optimizations• Management optimizations• Monitoring optimizations
• Hardware utilization• Cost
Customers don’t virtualize Exchange because of• Complexity
• Additional deployment steps• Additional management layer• Additional monitoring layer
• Performance impact• Workload incompatibility/unsupportability• Cost
Exchange Team’s Recommendation
• We support virtualizing Exchange because it makes sense for some of our customers
• Customers should pick the simple solution• Physical is often the simple solution, but not for every
customer
• Customers that virtualize Exchange should have a clear idea as to what they get out of virtualizing
EXCHANGE SERVER 2013 VIRTUALIZATION SUPPORT
Supported• Hypervisors
• Any version of Windows Server with Hyper-V technology or Microsoft Hyper-V Server
• Third-party hypervisors validated under SVVP
• Exchange roles• Both Exchange roles supported
• Storage• Block-level• Same requirements as Exchange
2010*
• Host-based clustering• Both Exchange roles
supported
• Migration• Both Exchange roles
supported
• Jetstress testing in guests• Only on supported Windows
hypervisors or ESX 4.1 or newer
http://aka.ms/e2013virt
Not Supported
• Dynamic memory, memory overcommit, memory reclamation• Configure static memory for
all Exchange VMs
• Significant processor oversubscription• Limited to 2:1, best practice
is 1:1
• Hypervisor snapshots• Differencing/delta
disks• Apps on the root
• Only deploy management, monitoring, AV, etc.
Exchange and Hyper-V• The most tested hypervisor for Exchange
• Every day, thousands of Exchange test machines run on Hyper-V• Strong feedback loop between Exchange and Hyper-V teams• Ongoing cross-group engineering relationship
• Hyper-V 2012 adds many new features• Removal of 4 vCPU per-VM limit fantastic for Exchange• Increased memory per-VM important for Exchange 2013
• Customers who virtualize Exchange (and size correctly) will have a great experience on Windows Server 2012 Hyper-V
Windows Server 2012 and SMB 3.0• Great platform for inexpensive, simple
storage of Hyper-V virtual machines• Scalable to meet various levels of demand• Use large low-cost disks and take full
advantage of functionality in the virtualization stack
Windows Server 2012 and SMB 3.0
Virtualized Exchange and SMB 3.0• Exchange 2013 supports VHD storage on SMB 3.0 file shares
• Can be shares presented from Windows Server 2012 or other implementations of SMB 3.0
• Specific to VHD storage – no direct access to shares from Exchange• No change to our support of downlevel SMB & NFS
• SMB 3.0 provides the ability to survive hardware failures that would otherwise impact file access• Still need to design for HA & failure handling
NFS is not supported because• Particular NFS implementations have shown
• Significant performance issues historically, and Exchange is very sensitive to high IO latencies
• Reliability issues that can result in database corruption and the need to perform database reseeds
• NFS is a standard – there are many implementations, some better than others
• Given that there are a number of alternatives for presenting storage to a hypervisor, we don’t support NFS (or older versions of SMB)
Host-based Migration• We support Live Migration and similar 3rd-
party technologies with Exchange• We don’t support Hyper-V’s Quick Migration or
any other solution that saves point-in-time state to disk
• VM has to remain online during migration
Hyper-V Replica• Replica provides DR for a VM via log shipping
to a remote hypervisor• Makes sense for applications that don’t have DR
capability built-in to the product
• Not supported for Exchange• Use DAGs for HA, SR and DR
SIZING RECOMMENDATIONS
CPU Resources• Plan to add CPU overhead to Exchange VMs
• ~10% for Hyper-V• Follow vendor guidance for SVVP hypervisors
• Use the Server Role Requirements Calculator• http://aka.ms/e2013calc• Note “Server Role Virtualization” and “Hypervisor
CPU Adjustment Factor”
General Considerations• Memory sizing generally unchanged between physical and
virtualize deployments• Exchange is not NUMA aware, but will take advantage of OS
optimizations provided to host
• Storage should be optimized for low IO latency and high service and data availability• Take advantage of hypervisor networking flexibility to provide
availability and performance
• In general: size using guidance for physical, apply to virtual
Co-location with other VMs• As a best practice, never overcommit any
resources in a way that could impact Exchange VMs
• Use reservation options to ensure that Exchange gets the resources it needs• Allows other workloads to take advantage of
overcommit
Users per Host• Avoid extreme scale-up to ensure high
availability of Exchange service• Modern hypervisors are capable of hosting
hundreds of VMs per host, but plan for a small number of Exchange 2013 mailbox VMs per-host
• Use remaining capacity for other workloads
PROBLEM AREAS…AND HOW TO AVOID THEM
Failure Domains• Understanding failure domains is critical for virtualized Exchange
design• Stuff fails – embrace failure and prepare for it with redundancy and
multiple paths to infrastructure• Placing multiple copies of the same mailbox database on the same
infrastructure lowers availability• Placing any dependencies of Exchange on the same infrastructure
also lowers availability• Active Directory• Witness servers
Oversubscription• Hypervisors don’t make CPU resources appear out of thin air• Oversubscription can help with hardware consolidation, but it doesn’t
help provide reliable high-performance Exchange services• Proper Exchange sizing ensures that resources are available on-
demand, so don’t allow hypervisors to yank those resources away• CPU constrained Exchange servers will experience reduced
throughput:• Delivery throughput reduction = queue growth• Content indexing throughput reduction = increased IOPS• Store ROP processing throughput reduction = RPC latency & end-user pain
Workload Management• Dynamically adjusts background tasks to
ensure that resources are being consumed efficiently
• Monitors resource consumption and makes decisions based on resource availability
• Inconsistent resource assignment results in bad WLM decisions
Dynamic Memory• Hyper-V’s Dynamic Memory and VMware’s
Ballooning are fantastic for lab environments• Not supported for production Exchange
servers• Statically assign memory resources to
Exchange VMs
Host-based Failover Clustering• Enables you to automatically failover VMs to another server in
the event of a hardware failure• This
• Is not an Exchange-aware solution• Only protects against server hardware/network failure• Does not provide HA in the event of storage failure / data corruption• Requires a more expensive and complicated shared storage deployment
• If you are going to deploy host-based failover clustering, deploy the Exchange mailbox server VMs in a DAG
Host-based Migration• Migration of Mailbox servers can result in cluster
heartbeat timeouts• Result is eviction of server from cluster (DAG) and
failover of active database copies• Consider carefully adjusting heartbeat timeout settings
Import-module FailoverClusters(Get-Cluster).SameSubnetThreshold=5(Get-Cluster).SameSubnetDelay=1000
Hypervisor Snapshots• Hypervisor snapshots are great for labs
• Multiple machines need to roll back simultaneously
• Not supported for production servers• Exchange system components cannot travel
backwards in time (log shipping, etc.)
Hyperthreading• It’s OK to use hyperthreading on VM hosts,
but size for physical processor cores (not virtual)
Hyperthreading
Processors (4 CPU)
Physical processor cores(16 cores)
Logical processor cores(32 logical cores)
Use these for sizing!
SUMMARY
Summary• It’s important to understand when to virtualize
Exchange – it’s not always the right choice• Don’t oversubscribe; Ensure that Exchange
always gets resources it needs• Not all hypervisor features make sense for
Exchange
QUESTIONS?
Scott [email protected]: @SchnollBlog: http://aka.ms/schnoll
Please evaluate the session before you leave