How to fix IO problems for faster SQL Server performance

IO, IO! Why’s my SQL slow?
Brian Flynn
Product Manager
© 2014 SOLARWINDS WORLDWIDE, LLC. ALL RIGHTS RESERVED.

Define Speed?
» Acceleration
» Burst/Top Speed
» Sustained Speed
» Constraints
» Total Trip Time
» Average Speed
2

» Imagine transporting things a long distance.
» Just you and an over night bag vs everything you own…
3
Latency, Speed, & Throughput
High Acceleration (Low Latency) Medium Acceleration (Medium Latency) Low Acceleration (High Latency)
High Top Speed Medium Top Speed Low Top Speed
Small Cargo Medium Cargo Large Cargo
Low Capacity Medium Capacity High Capacity

» Electricity:
 Volts / Ohms = Amps
 Volts * Amps = Watts
» Data:
 Buffer Size / Latency = Transfer Rate
 Buffer Size * Transfer Rate = Throughput
4
Analogy Size Current Resistance Capacity
Electricity Volts Amps Ohms Watts
Water Pressure Flow Rate Inverse of channel
diameter
Volume/Time
Data Buffer Size
* compression
* jumbo frames
Transfer Rate
*IOPS
Latency
* Function of Distance
* Function of storage type
Throughput

» When speed is
constant, packet size
governs throughput.
» Larger packets increase
throughput.
» Greater distance
increases latency which
lowers perceived
throughput.
5

True speed is all about latency
» 1 byte or 1000 bytes,
same trip time.
» Overhead accounts
for more constraint
than sum of the
serialized trips.
6

Aren’t we talking about storage?
» HDD latency
 Rotating the platter
 Moving the head
» SSD latency
 One controller is a gateway for multiple chips
» RAID latency
 Mirror – Improves read IOPS if split reads is supported.
 Stripe – Improves read & write IOPS.
 Distributed Parity – Write IOPS penalty for calculating parity bits.
 Raid 10 – Mirror & Stripe. No write penalty. Twice the price of stripe only.
» SAN latency
 One controller is the gateway to many disks.
 Limited bandwidth to arrays
» NAS latency
 Network is usually the weakest link
7

Enough already, just make it faster!
» There’s more than straight line acceleration.
» Complex systems force us to tease out the latency bottlenecks
» Balancing 3 Options
1. Get the same work done with fewer operations
2. Do more in parallel
3. Make individual operations faster
8

Get the same work done with fewer operations
» In our transportation example, same work with fewer trips
… The big semi does this… larger cargo or payload.
» With data, however…
 Eliminate redundant or superfluous database work
• Create indexes that reduce table scans
• Keep tables defragmented to increase read-ahead
• Keep frequently used data in application memory
 Caching techniques to reduce re-reading from storage
• SAN cache
• Server memory
• SQL 2014 Buffer Pool Extension
 Compress data
• More pages in memory increases page life expectancy
• Compressed data will fit into fewer Physical I/Os
• ** CPU latency must be lower than storage latency
9

Do more in parallel
» Adding disks to an array increases maximum IOPS
 Helps where I/O can be parallelized
• Multiple CPU issuing synchronous I/Os
• SQL Server Pre-fetch & Read-ahead
• Reference: MSDN: Sequential Read Ahead
» Defragmentation increases Read-ahead, which is asynchronous
» Evaluate for parallel capabilities end to end
 Multiple file groups on isolated storage
 Multiple, isolated LUNs
 Multiple OS disk queues
 Consider SAN paths & controllers
» Bottom line, you can’t parallelize everything
 Business Analogy: The Mythical Man Month
 Humorous Analogy: The Pregnancy Problem
10

Make individual operations faster
» Almost always, a winner! But often, the most expensive solution.
» Lower latency storage
 HDD Moving parts  Mechanical latency
 Random reads  More mechanical latency
 Thus to optimize, maximize sequential reads before random reads
 Of course, with RAID arrays, almost every read is random
 SSDs have no moving parts  No mechanical latency
 Thus all reads are equal. No benefit to sequential vs random
» Who’s IO are we talking about?
 Typical disk I/Os are usually 4K
 SQL Server I/Os range from 8K to 512K
» REF: http://www.confio.com/logicalread/fragmentation-reorganization-rebuilds-sql-server-table-scans-bf01/#.U1aJKvldU9I
11

How Do I Choose? It depends… :-
» It depends… At best, all three. Next best, any combination.
 3rd party software may not permit your query tuning
 Budget constraints may not permit low latency storage
 The size of I/Os may not lend to parallelizing workloads
» General rules:
 Reducing total # I/Os is always a winner.
 For large result sets or concurrency  Increasing IOPS can help a lot
 For small result sets  Raw speed (Latency) may be your only hope.
12

Disk Queue Depth
» I hate queues, especially at the DMV! Waiting in line stinks!
» Disk Queue Length is a measure of I/O constraint.
» How high is OK? It depends on your environment.
 Different performance expectation between OLTP and OLAP systems
» If it’s consistently high, adding disks to the array may help.
» You should be watching this one and DPA does that…
13

Other WMI Counters
» Latency-Like
 Avg. Disk Sec/Read is the average time, in seconds, of a read of data
from the disk.
 Avg. Disk Sec/Write is the average time, in seconds, of a write of data
to the disk.
» Current-Like
 Avg. Disk Reads/Sec is the rate of read operations on the disk.
 Avg. Disk Writes/Sec is the rate of write operations on the disk.
» Physical Disk: %Disk Time is the percentage of elapsed time
that the selected disk drive was busy servicing read or write
requests.
REF: http://blogs.msdn.com/b/dpless/archive/2010/12/01/leveraging-sys-dm-io-virtual-file-stats.aspx?Redirected=true
14

Virtual File Stats DMO
15
http://technet.microsoft.com/en-us/library/ms190326.aspx

Crystal Disk Mark
» A great, free tool that tells you the throughput of your
storage under a few mixtures of I/O size, queue depth
and sequential or random I/O. You may not get your
storage admin’s blessing to run this.  But it’s a decent
way to measure throughput.
16

SQLIO
17
REF: http://www.brentozar.com/archive/2008/09/finding-your-san-bottlenecks-with-sqlio/

18
Query Plans & Process Monitor
» Query plans tell you when you’re scanning and seeking
» ProcMon tells you how many and how big are the storage I/Os

Server memory, SAN Cache, Tierd Storage etc
» You’ll have to clear a few obstacles.
» SQL Server caches pages in RAM so when
testing you’ll need dump the buffer cache to
ensure you are going to disk.
» And unless your storage admin is a gift from
the DBA gods, you’ll have to get real creative
to control for SAN cache in any experiments.
» Tiered storage can cause unpredictability too.
19

Free Tools for SAN & VM environments
20
http://www.solarwinds.com/products/freetools/san_monitor/
http://www.solarwinds.com/products/freetools/storage-response-time-monitor.aspx

Design Choices : Fragmentation Impacts
» Design for Read Ahead
 Understand & avoid physical and logical fragmentation
 Clustered indexes that monotonously increment on isolated storage
 Pre-allocate files in larger swatches rather than auto-grow
» Use file groups and different storage based on I/O demand
 Rarely accessed things on less performant media
 Frequently accessed things on highly performant media
» Plan ahead!
21

DMV Wait Types
» Multiple wait types indicate storage I/O
» DPA breaks down query execution by wait type over time.
 http://database.demo.solarwinds.com/
22

DEMONSTRATION TIME!
23

Think you have Database I/O issues?
» DPA can identify and recommend action on I/O issues.
» Check out the new Storage I/O feature in DPA 9.0!
24
http://www.solarwinds.com/resources/videos/utilize-dpa-storage-io-to-find-performance-issues.html

Quickly resolve Oracle performance issues
» Try Database Performance Analyzer FREE for 14 days
» Improve root cause of slow performance
 Identify issues that impact end-user response time
 Isolates root cause in just four clicks
 See historical trends over days, months, and years
 Understand impact of VMware® performance
 Agentless architecture with no dependence on Oracle Packs, installs in
minutes
www.solarwinds.com/dpa-download/

Thank you!
» Brian Flynn
» brian.flynn@solarwinds.com
» Twitter : @brianpaulflynn
» LinkedIn: brianpaulflynn
26

Questions?

Thank You!
The SOLARWINDS and SOLARWINDS & Design marks are the exclusive
property of SolarWinds Worldwide, LLC, are registered with the U.S.
Patent and Trademark Office, and may be registered or pending
registration in other countries. All other SolarWinds trademarks, service
marks, and logos may be common law marks, registered or pending
registration in the United States or in other countries. All other
trademarks mentioned herein are used for identification purposes only
and may be or are trademarks or registered trademarks of their respective
companies.

How to fix IO problems for faster SQL Server performance

More Related Content

Viewers also liked (20)

Similar to How to fix IO problems for faster SQL Server performance (20)

More from SolarWinds (20)

Recently uploaded (20)

How to fix IO problems for faster SQL Server performance