Virtual SQL Servers. Actual Performance.

© Heraflux Technologies® 2
@kleegeek
davidklee.net
heraflux.com
linkedin.com/in/davidaklee
Specialties / Focus Areas / Passions:
• Performance Tuning & Troubleshooting
• Virtualization
• Cloud Enablement
• Infrastructure Architecture
• Health & Efficiency
• Capacity Management Founder & Chief Architect

 What is Virtualization?
 Resources and queues
 Physical servers
 Storage and interconnects
 VM architecture
 Inside the VM

 Added layer
 Physical resources
 Resource queues
 Very small queue delays
 When done right…
 Shared everything architecture

150 GHz
CPU 4 TB
Memory
4x10GbE
Network
20 TB
Tier 1
Storage
40 TB
Tier 2
Storage
VM
16 vCPU
128 GB vRAM
VM
8 vCPU
64 GB vRAM
VM
2 vCPU
16 GB
vRAM
VM
2 vCPU
16 GB
vRAM
VM
2 vCPU
16 GB
vRAM
VM
2 vCPU
16 GB
vRAM
VM
2 vCPU
16 GB
vRAM
VM
2 vCPU
16 GB
vRAM
V I R T U A L I Z A T I O N

Hypervisor
CPU Scheduler
CPU
Execution
CPU Scheduling Queue
Memory Allocator
Mem
R / W
Mem Allocation Queue
Disk Scheduler
Disk
R / W
Disk Scheduling Queue
Network Scheduler
Network
Tran / Rec
Network Scheduling Queue
VM TASK
VM TASK
VM TASK
VM TASK
VM TASK

• Finite!
• Single compute node hardware
• Total cluster compute capacity
• Storage speed (IOPs, throughput)
• VM maximums
• Interconnect path speed
Hard Limits (Resources) Soft Limits (Queues)
• Measurable but not finite
• Memory oversubscription
• CPU scheduler contention
• Shared resource utilization
• “Noisy Neighbors”
© Heraflux Technologies® 7• 7
• 7

Networking
Interconnects
Storage
Physical Server
Virtualization
Operating System
SQL Server Instance
SQL Server DB
Application

 Most shared
 Most critical
 Most complex
 Most problematic
 Slowest piece of the stack
 Many individual points of contention

Controller
Controller
LUN
LUN
LUN
LUN
Disk Pool
T1
T2
T3

 Test raw performance
 SQLIO
Deprecated!
 DiskSpd Batch
heraflux.com/go/diskspd
 Collect metrics:
 I/Os per second (IOPs)
 Latency (ms)
 Throughput (MB/s) 0.00
10000.00
20000.00
30000.00
40000.00
50000.00
60000.00
70000.00
1 2 4 8 16 32 64 128
IOps Thread Intensity
IOps Per Operations per Thread
Sequential Read
Random Read
Sequential Write
Random Write

14
(Img src: http://www.asacomputers.com/2U-Server-Sandy-Bridge.html)

 Manufacturer
 Intel
 AMD
 Sockets
 Cores
 Speed vs. GHz
 Logical vs. Physical
 Hyperthreading

(Img src: http://frankdenneman.nl/2011/01/05/amd-magny-cours-and-esx/)
CPU CPU CPU CPU
Memory

(Img src: http://www.learnyourtech.com/hardware/)

 Sample Cisco UCS
memory config
 Slot & chip
placement
(Src: http://www.cisco.com/c/dam/en/us/products/collateral/servers-unified-computing/ucs-b-series-blade-servers/b200m4-specsheet.pdf page 44)

 Resource limits are easy to detect / work around
 Queue contention much harder
 Time in queue = time lost from VM
 Silent performance killer
 Everything in a VM must be scheduled
 … including idle resources
 Queue processing is not always FIFO

 Get physical machine configuration
 Try to fit VM inside one NUMA node
 Otherwise, balance across number
of NUMA nodes
 Test configurations for
best results

 Example: 16 vCPU VM
 What’s better?
 2 vSocket x 8 vCore?
 Varies by workload,
hardware
 Test it for yourself!
0
100000
200000
300000
400000
500000
600000
700000
800000
900000
8 16 64 256
Transactions/min Concurrent HammerDB Users
vNUMA SQL Server Scalability - 16 vCPUs - HammerDB
4socket x 4CPU 8socket x 2CPU 2socket x 8CPU

 1 vCPU to 1 pCPU is
a poor recommendation
 Queues
 Ready Time
 Co-Stop
 Ex: Client environment
 74 running VMs
 VMs @ 2-8 vCPU
 72 total pCores
 Fantastic performance

 “Right” amount of vCPU and vRAM resources
 Physical world = Size for end of life
 Virtual world = Size for right now
 Idle vCPUs will slow application VMs performance
 Repeat “right-sizing” analysis routinely

 Virtual disks whenever possible
 Multiple SCSI controllers
 VMware – PVSCSI
 Spread out the workload

 C: - Operating System
 D: - SQL Server Instance Home
 E: - System Databases (master, model, msdb) *
 F: - User Database Data (1 of X)…
 G: - User Database Log (1 of Y)…
 H: - TempDB…
 Y: - Windows Page file **
 Z: - Backups Adjust as necessary
(but stay standardized)

 64KB NTFS block sizes
 Set power settings to “High Performance” (CPU-Z)
 Set antivirus exclusions for SQL Server (tinyurl.com/sqlav)
 Ongoing OS-level performance metric collection
 No greater than five minute interval
 Windows Perfmon, Microsoft SCOM,
or other third-party utility
 heraflux.com/go/perfmon

 Goal: Maximize performance while reducing resource scheduling
 Parallelizable workloads
 Determine how parallel the workload is
 Set MaxDOP = vNUMA node core count (?)
 Cost threshold for parallelism = Not default
 Jonathan Kehayias – Tuning CTOP from Plan Cache
 bit.ly/1rTs9UX

 Spread out the I/O
 File groups, data files, partitions
 Parallelism with multiple active storage paths
 Reduce I/O
 Table / index compression vs. SAN compression
 In-memory constructs
 More RAM
 SSD read / write caching
 Faster I/O – All flash SAN
 And then clean up bad schemas & queries!

 Virtualization works.
 Equivalent performance if done right
 Efficiency in data handling

@kleegeek
davidklee.net
heraflux.com
linkedin.com/in/davidaklee

Virtual SQL Servers. Actual Performance.

More Related Content

What's hot (20)

Similar to Virtual SQL Servers. Actual Performance. (20)

Recently uploaded (20)

Virtual SQL Servers. Actual Performance.

Editor's Notes