Varrow madness 2013 virtualizing sql presentation

Virtualizing SQL

How to implement on virtual
infrastructure to maximize
performance and HA

The Golden Rule
• Design considerations for SQL workloads are
essentially the same between physical and virtual
environments now
– vCPU can perform the same as CPU now, even virtual
multi-core sockets
– If guaranteed I/O SLA is demanded, dedicated disks can be
provisioned to virtual infrastructure just as easy as physical
(easier in most cases). Storage vMotion means you can
always adjust later completely transparent to SQL
– Virtual memory is always backed by physical memory
unless a cluster in under pressure. So vRAM is no different
from RAM as long as you are not overprovisioning your
virtual infrastructure

Base Advantages of Virtualized SQL
• High Availability
– Provides all SQL VMs with an inherited cluster-like ability, without any
of the headache(when SLA’s do not require very low RTO)
• Snapshot
– Ability to snapshot prior to major changes (ie: Windows updates)
allows for faster rollback (to that point in time)
• Flexibility
– Easily increase memory or CPU to meet periodic demand spikes. End
of month activities, or periods of high usage for front end applications
• Portability
– Move workloads/servers to higher-power hosts without interruption
of service. Refresh hardware without reinstall. Provides for easier and
more plentiful disaster recovery options

The Way Virtualization Used to Be
• Old drawbacks – Anemic VMs from a SQL
perspective
– Storage
• Response times could get worse
• Throughput could be decreased
– Compute
• Lack of memory
• Lack of vCPU
– Cost
• Consolidation ratio blown. Host memory density not worth it

Multi-Vendor Virtualization Today
• New strengths
– Storage
• 2TB disk size, 60 SCSI targets per VM
• Improvements in storage. much better latency on storage and IO, up to 1 million IOPS
from a single VM with vSphere 5.1!
• Server-side flash storage - great fit for tempdb. low persistence but high utilization.
• New technologies in storage can enable SQL admins to gain access to persistent flash
based storage to store things like index data files (separate them into different file
groups) identify heavily read-only content and you can move it into flash. Hybrid flash is
the future
• EMC Fast Cache as an example
– Compute
• More CPU , up to 64 vCPU for a single VM
• More RAM , up to 1TB for a single VM
– Cost
• Much lower price, economy of scale kicking in
• Consolidation ratio good. Host memory/CPU density has increased massively. 384 GB+
RAM in a single 2 socket blade is common

SQL Workloads
• OLTP
– High volume web back-ends
– High IOPS requirements throughout the day
• Plan for peak daily activity
– CPU contention may require more vCPU per VM,
especially on poorly optimized (front end) systems.
– Pro-tip : web application developer <> SQL developer
in most cases. Poorly coded queries can shred a virtual
SQL instance just as easily as a physical. Virtual vs
physical doesn’t change your DBA hat.
– Dynamically scale based on known demand peaks

SQL Workloads
• OLAP
– Data warehouse
• Predictable high volume workloads
• Just like a physical server, daily ETL jobs will crush the disk.
This is not an artifact of virtualization, just the nature of the
job.
• Easier to dynamically scale a server for EOY operations. Add
CPU/Memory when needed and remove when not required.
• Storage DRS can adapt workloads to appropriate storage
dynamically.
– Reporting services
• Treat like a web server (since that is essentially what it is)

When is it not a good idea to
Virtualize?
• Vendor support
– Number 1 case that precludes virtualizing a database
workload
– Internal assessment of risk and maintenance of a
parallel physical test environment to satisfy Vendor
requirements (reproduce an issue in the physical test
environment)
• Ultra-low latency/Custom hardware
requirements (example Stock market)
– Not many real-world examples that people will often
run into.

Licensing
• SQL 2012 as point of reference. Individual VM vs
Host licensing
• Individual VM licensing
– Core license model
• 1 core license per virtual CPU/thread
– Hyper-threading counts!!
• Sold in 2-packs
• Minimum of 4 core licenses per VM
– Server/CAL licensing
– Still present for standard edition
– Only present for Enterprise edition if you currently have an active
SA contract on your Enterprise Ed. Server/CAL licenses
– Can run maximum of 4 servers per host on Enterprise

Licensing
• License mobility
– Any SQL license with SA on it can move to different
physical hosts as the VM moves.
– Available for both per-core and Server/CAL models
• Dense virtualization licensing
• License all of the physical processors on a host and you can
spin up unlimited virtual SQL servers (with SA on your
licenses)
• Great fit for environments with large number of SQL servers
• Can carve out either dedicated SQL clusters, or VM affinity
rules for the group of SQL servers to only run on a group of
hosts.

VM Configuration
• TEMPLATE
– Best thing you can do is to create a standard SQL VM
template that is tweaked (and documented!) to high
heaven
– Reduces deployment time to approximately an
hour, even with some post-patching
– Consistent high performance design reproduced
throughout the environment
– Allows you to set the standard but still let
virtualization administrators to deploy SQL VMs for
you

VM Configuration
• Memory considerations
– SQL and OS side tweaks
• “Lock pages in memory” permission for SQL service account
• Set SQL max memory to 1-2GB less than OS memory
• Rough Rule of thumb, under 8GB of memory = 1GB reserved
for OS. More than that = 2GB reserved for OS.
– Vmware Reservation for full memory amount
– Critical if you use lock pages in memory tweak, but important to
use regardless.
– This is one of the real legitimate uses of memory reservations!
Don’t be afraid to ask for it!

VM Configuration
• CPU considerations
– SQL tweaks
• Set Max Degree of Parallelism (MDOP) to be equal to the number
of vCPU assigned to the VM
• Split TempDB into a number of files equal to the number of vCPU
assigned to the VM. This will optimize access to TempDB
– VMware tweaks
• Start with a single vCPU in your template.
– Easier to go up to the multiprocessor HAL than come back down again
– All virtual resources have overhead, less is more. Only assign what you
really need.
• Remember that you can now create multi-core vCPU sockets. This
can have licensing implications.

VM Configuration
• Network considerations
– VMTools and VMXNET 3
– VMXNET 3 NIC driver is bundled into VMTools.
– Ensure the latest version of VMTools is included in
your template
– SIGNIFICANT performance advantage over default
E1000 virtual NIC driver (TCP offloading),
especially when using 10 Gb NICs in the hosts.

VM Configuration
• Storage considerations
– Sensitive workloads: design backend storage like
you would for physical SQL
– For SQL VM’s with low performance reqs: design
backend storage like you would for other server
VM’s

VM Configuration
• Storage Protocols
– FC / FCoE
• Stable, consistent, reliable. We prefer FC vs IP based storage
protocols when discussing SQL (and other business critical
apps)
– iSCSI
• Easy to start, but more complex than FC to do right
• You don’t need Jumbo Frames, that’s to reduce CPU and we
have plenty of CPU
– NFS
• No true multipathing (yet)
• Does not support RDMs, so capped at 2 TB for a volume
• Scales out very well

VM Configuration
• FC vs iSCSI
– Fibre Channel – similar to
railways
• Purpose built, connected to
predetermined specific endpoints.
• Predictable performance
– iSCSI – similar to highways
• Can be more flexible
• Endpoints are simple to add
• Traffic, latency can be a problem

VM Configuration
• RDM vs VMDK
– Performance is near identical. Very, very small
performance hit for VMDK
– VMDK is much simpler to use
– VMDK is <still> limited to just under 2 TB
– RDM is required for Physical to Virtual failover
clusters
– RDM is required for array based snapshot and
backup applications

VM Configuration
• VM SCSI controllers
– Use multiple SCSI controllers to allow parallel I/O
operations (max of 4)
– Separate controller for OS, data files, transaction
log files
– Queue depth – default is usually fine, engage your
storage team and vendor before adjusting. Lots of
knobs and dials here.

VM Configuration
• Paravirtualized SCSI controllers
– Recommended for new SQL VM’s that are storage
performance sensitive
– Will require loading the mass storage driver during
Windows if used for the system drive (use a
template!)
– About 2000 IOPS from the VM is when it starts to
make a difference

VM Configuration
• Partition alignment - host:
– In vSphere 5.X, VMFS is aligned at the 1 MB mark, this
is fine
• Partition alignment - guest:
– Align guest partitions, usually 1 MB is good (use a
template!)
• Create VMDK files as Eager Zeroed Thick
– You don’t want write I/O’s waiting for the .vmdk to be
zeroed
– VAAI enabled storage arrays will speed this up

Using PCIe SSD Cards
• I/O Latency measured in microseconds instead of
milliseconds
• EMC XtremeSF, FusionIO, others. Does not require a
specific storage array.
• Manually place tempDB here, for example, or other non-
unique data (mirrored at the application level, for instance)
• Software enables using local SSD as an extension of storage
array cache. Can be used for unique data, reads are
accelerated and writes are handled by the storage array
(EMC XtremeSW)
• Or use SQL Always-On to protect data, allowing unique data
to be written to local SSD
• Will affect vMotion (and therefore DRS), but SQL Always-on
can mitigate

Backup Strategy
• Choice between Host based application aware VM
backups, and traditional SQL maintenance plan backups
– Must choose an approach for Full Recovery Mode, as each one
wants to have control over the backup chain, and you do not
want to risk both truncating the transaction log
– Host based appropriate for single daily backups
– SQL Maintenance plan preferred for “point-in-time” type restore
SLA
– Must adjust either strategy no matter which is chosen. Neither
can do the whole job, so both must be used
• Backup admins will have policies configured to groups of
VMs. You must ensure SQL VMs are designated as a
separate backup group so policies can be tweaked no
matter what choices are made

Backup Strategy
• Host based application aware backups
– Whole VM backup that will truncate SQL/Exchange
Logs when the job is run
– Utilizes VM snapshots to freeze disk IO while the
backup is run
– Provides a crash consistent backup. Rollbacks may
occur when databases are brought online
– Appropriate for workloads that have a single daily
backups SLA
– Ensure your databases are in Simple recovery mode
not Full, to keep transaction log size down

Backup Strategy
• SQL Maintenance plan backups
– Still useful for index and statistics and general
maintenance work around the database, even if
Host based backups are used

Virtual SQL Clustering and HA
• Traditional clustering
– Very much still available. 5 node cluster max with
vsphere 5.1
– vMotion and Storage vMotion not allowed. HA not
impacted
– Virtual/Physical hybrid clustering possible
– Some restrictions on allowed storage protocols
– Does not make traditional clustering any less painful
– Does allow for a cultural transition from physical to
virtual more palatable for some

• SQL Always-On clustering
– Best of both worlds!
• Essentially get clustering without shared storage
requirement or headaches normally found in MSCS clusters
• No shared storage = less money with no loss in
stability/performance
– Super-fast failover you get from SQL Mirrored
databases without SQL NCLI (or .NET data connector)
requirement.
• Fails groups of databases over together
• Works great even for legacy apps that use old ODBC/JDBC
drivers

• SQL Always-On clustering
– Based on proven technologies
– Windows Failover Clustering
– SQL Mirroring
– Exchange DAG similarity
– Lagged nodes allow for cheap off-site DR capacity
– Read-intent queries
• Automatically offload certain query types to a read-only
node
– Backups can run on secondary copies
• Can specify preference for where backups run. Maintenance
plans are intelligent and know if another node is running a
backup. Can leverage this intelligence in scripts as well.


Video demonstration of
Always-on failover

• Site Recover Manager – Offsite replication and
recovery options
– Protects entire virtual infrastructure including SQL
– Can provide close to 0 RPO and very low RTO
– Scripted DR recovery means critical SQL servers
can come up first. VM “importance” preference.
Critical for SQL workloads that are often back-
ends for other systems.

Take-away’s for the DBA’s
• Use a TEMPLATE!!
– Seriously build a SQL VM template
– Building a template is great….updating a template
twice a year at least is better
• vCenter access
– Performance stats over time
– Console access to VM’s (virtual KVM access)
– Event history
– Capacity planning
– Security concerns from VM Admins can be handled by
granting read-only access to the DBA

Take-away’s for the DBA’s
• More <> better
– Often times overprovisioning virtual resources can
actually mean worse performance. Don't ask for more
than you need
– Start with single vCPU in your template. Easier to go
to SMP than to go back to single proc. Single VM
reboot to add vCPU’s for SMP.
– Remove extraneous devices, CD-
Rom, ISO’s, Floppy, anything that is not needed for
clean operation.
• You can always add things on temporarily if needed

Links
• Always-On Failover Demonstration – Kyle Quinby, Varrow
• MS KB 920093 - Tuning options for SQL Server when running in high
performance workloads
• MS SQL 2012 license guide for virtualization
• VMw KB 1037959 - Microsoft Clustering on VMware vSphere: Guidelines
for Supported Configurations
• VMware's SQL Server Best Practices Guide
• When to use Lock Pages in Memory with SQL Server
• How to Enable the Lock Pages in Memory Option
• VMware - 1 Million IOPS on 1 VM
• VMw KB 1010398 - Configuring disks to use VMware Paravirtual SCSI
(PVSCSI) adapters
• VMw KB 1022242 - Types of supported Virtual Disks on ESX/ESXi hosts

Varrow madness 2013 virtualizing sql presentation

More Related Content

What's hot (20)

Similar to Varrow madness 2013 virtualizing sql presentation (20)

Varrow madness 2013 virtualizing sql presentation

Editor's Notes