95013690 a Practical Introduction to Storage Disk Subsystem Performance V1 55
description
Transcript of 95013690 a Practical Introduction to Storage Disk Subsystem Performance V1 55
© 2010 IBM Corporation
IBM Power Systems and Storage Symposium, Wiesbaden, Germany – May 10-12, 2010
A practical Introduction toDisk Storage System Performance
Gero Schmidt, ATS System StorageIBM European Storage Competence Center
© 2010 IBM Corporation
A practical Introduction to Disk Storage System Performance
2 IBM Power Systems and Storage Symposium, Wiesbaden, Germany – May 10-12, 20102010-09-13
Agenda
• Disk Storage System Selection & Specs
• Application I/O & Workload Characteristics
• Hard Disk Drive (HDD) Basics – It‟s all mechanic
• HDD Performance & Capacity Aspects (SATA vs FC)
• RAID Level Considerations (RAID-5 / RAID-6 / RAID-10)
• New Trends & Directions: Solid State Drive (SSD)
• Basic Principles for Planning Logical Configurations
• Performance Data Collection and Analysis
© 2010 IBM Corporation
A practical Introduction to Disk Storage System Performance
3 IBM Power Systems and Storage Symposium, Wiesbaden, Germany – May 10-12, 20102010-09-13
Agenda
• Disk Storage System Selection & Specs
• Application I/O & Workload Characteristics
• Hard Disk Drive (HDD) Basics – It‟s all mechanic
• HDD Performance & Capacity Aspects (SATA vs FC)
• RAID Level Considerations (RAID-5 / RAID-6 / RAID-10)
• New Trends & Directions: Solid State Drive (SSD)
• Basic Principles for Planning Logical Configurations
• Performance Data Collection and Analysis
© 2010 IBM Corporation
A practical Introduction to Disk Storage System Performance
4 IBM Power Systems and Storage Symposium, Wiesbaden, Germany – May 10-12, 20102010-09-13
IBM System Storage Disk Subsystems – Making a Choice
Selecting a storage subsystem:
entry-level, midrange or enterprise class
support for host systems and interfaces
overall capacity & growth considerations
overall box performance
advanced features and copy services
price, costs / TCO, footprint, etc.
needs to meet client & application requirements
Subsystem performance:
overall I/O processing capability
overall bandwidth
choosing the right
number and type of
Disk Drives
Enterprise
DS3000 DS5000 DS6000 DS8000
MidrangeEntry-level
XIV
© 2010 IBM Corporation
A practical Introduction to Disk Storage System Performance
5 IBM Power Systems and Storage Symposium, Wiesbaden, Germany – May 10-12, 20102010-09-13
Storage Subsystem Specs – DS4000 Data Rate
max. throughput may be achieved with a relatively low no. of disk drives
subsystem architecture: frontend / backend bandwidth capabilities are key
SATA may be considered for applications requiring throughput
Note: Results as of 6-26-2006. Source of information from Engenio and not confirmed by IBM. Performance results achieved under ideal circumstances in a benchmark
test environment. Actual customer results will vary based on configuration and infrastructure components.
The number of drives used for MB/s performance does not reflect an optimized test config. The number of drives required could be lower/higher.
© 2010 IBM Corporation
A practical Introduction to Disk Storage System Performance
6 IBM Power Systems and Storage Symposium, Wiesbaden, Germany – May 10-12, 20102010-09-13
Storage Subsystem Specs – DS4000 I/O Rate
Note: Results as of 6-26-2006. Source of information from Engenio and not confirmed by IBM. Performance results achieved under ideal circumstances in a benchmark
test environment. Actual customer results will vary based on configuration and infrastructure components.
Drives were “short-stroked” to optimize for IOPs performance. Real-life may take more drives to achieve the numbers listed..
max. IOps performance requires a high no. of fast FC/SAS disk drives
subsystem architecture: I/O processing capability >> disk drives IOps capability
SATA is not a good fit for enterprise class applications requiring transaction performance
© 2010 IBM Corporation
A practical Introduction to Disk Storage System Performance
7 IBM Power Systems and Storage Symposium, Wiesbaden, Germany – May 10-12, 20102010-09-13
Storage Performance Council (SPC) - Benchmarks
The Storage Performance Council (SPC) is a vendor-neutral standards body focused on the
storage industry. It has created the first industry-standard performance benchmark targeted at
the needs and concerns of the storage industry. From component-level evaluation to the
measurement of complete distributed storage systems, SPC benchmarks will provide a
rigorous, audited and reliable measure of performance.
http://www.storageperformance.org
© 2010 IBM Corporation
A practical Introduction to Disk Storage System Performance
8 IBM Power Systems and Storage Symposium, Wiesbaden, Germany – May 10-12, 20102010-09-13
Agenda
• Disk Storage System Selection & Specs
• Application I/O & Workload Characteristics
• Hard Disk Drive (HDD) Basics – It‟s all mechanic
• HDD Performance & Capacity Aspects (SATA vs FC)
• RAID Level Considerations (RAID-5 / RAID-6 / RAID-10)
• New Trends & Directions: Solid State Drive (SSD)
• Basic Principles for Planning Logical Configurations
• Performance Data Collection and Analysis
© 2010 IBM Corporation
A practical Introduction to Disk Storage System Performance
9 IBM Power Systems and Storage Symposium, Wiesbaden, Germany – May 10-12, 20102010-09-13
Avg. Access Time for an I/O operation:
CPU cycle < 0.000001ms
MEMORY < 0.001ms
DISK (HDD) < 10ms
Application I/O – An Overview
HardwareHardwareSetup
Software
System Software
ApplicationSoftware
↔ ↔
Application
File Systems
SANFC
iSCSI
IB
SAS
SATA
SCSI
S
E
R
V
E
R
M
E
M
O
R
y
S
T
O
R
A
G
E
C
A
C
H
EDevice Drivers
Volume Manager
Storage subsystem
Cache hit: ~ 1 ms
Physical HDD: ~ 5...15 ms
Storage I/O performance:
Proper data placement is key!
Application I/O performance: Efficient memory usage is key!
Access to memory is >10000 times faster than disk access!
Disk access is ‚SLOW‘
compared to
CPU and MEMORY
© 2010 IBM Corporation
A practical Introduction to Disk Storage System Performance
10 IBM Power Systems and Storage Symposium, Wiesbaden, Germany – May 10-12, 20102010-09-13
Application I/O – On a typical System Time Scale
CPU 1ns (1GHz) = 0.000000001s
MEMORY 100ns = 0.000000100s
DISK 10ms = 0.010000000s
CPU 1 cycle := 1 second
MEMORY 1:40 minutes
DISK 115.74 days‚SLOW‘
© 2010 IBM Corporation
A practical Introduction to Disk Storage System Performance
13 IBM Power Systems and Storage Symposium, Wiesbaden, Germany – May 10-12, 20102010-09-13
Application I/O – Where does it come from?
Transaction Processing
A single end-user is capable of initiating only a moderate number of transactions with a limited amount of data changes per minute
Thousands of end-users can already initiate thousands of transactions and generate high I/O rates with only low data rates
End-users are directly affectedby the application response time
People„s work time is expensive
Excellent overall response time of the application is business criticaland requires low I/O response times at high I/O rates
Batch Jobs
A single batch job can alreadygenerate a considerable amount of disk I/O operations in terms ofI/O rate and data rate
Multiple batchjobs can createa huge amount of disk activity
Batch jobs should not interact with end-user transactionsand are typically run outsideend-user business hours
Time frames for batch jobseven during nights / weekendsare limited
Overall job runtime is critical andmostly dependent on achievedoverall data rate
I/O I/O
© 2010 IBM Corporation
A practical Introduction to Disk Storage System Performance
14 IBM Power Systems and Storage Symposium, Wiesbaden, Germany – May 10-12, 20102010-09-13
Throughput dependent workloads
Application I/O – Workload Characteristics
I/O rate in IO/s (IOps) Data rate in MB/s (MBps)
! Time to data is critical
! Dependent on number and type of
disk drives
! Data transfer rate enables
performance
! Dependent on internal
controller bandwidth
typical for transaction processing workloads
with random, small-block I/O requests, e.g.
OLTP – online transaction processing,
databases, mail servers – the majority of
enterprise applications
avg. I/O response time is most important here
(RT < 10ms is a good initial choice)
number and speed of of disk drives is essential
(e.g. 73GB15k FC drives as best choice)
SATA disk drives not generally recommended,
high speed FC/SAS/SCSI disk drives preferred
balanced system configuration and volume
layout is key to utilize all disk spindles
typical for throughput dependent workloads
with sequential, large-block I/O requests, e.g.
HPC, seismic processing, data mining,
streaming video applications, large file
access, backup/restore, batch jobs
avg. I/O response time is less important
(high overall throughput required)
bandwidth requirements (no. of adapters and
host ports, link speed) must be met
not necessarily a high number of disk drives
required
SATA disk drives may be a suitable choice
balanced system configuration and volume
layout is important to utilize full system
bandwidth
Transaction processing workloads
© 2010 IBM Corporation
A practical Introduction to Disk Storage System Performance
15 IBM Power Systems and Storage Symposium, Wiesbaden, Germany – May 10-12, 20102010-09-13
Application I/O – Workload Performance Characteristics
Basic workload performance characteristics:
I/O rate [IOps] (transactions) or data rate [MBps] (throughput)
Random access or sequential access workload pattern
Read:write ratio (percentage of read:write I/O requests, e.g. 70:30)
average I/O request size (average I/O transfer size or block size, e.g.
8kB for Oracle DB, 64kB or larger for streaming applications, 256kB for TSM)
Additional workload performance characteristics:
Read cache hit ratio (percentage of read cache hits)
average response time (RT) requirements (e.g. RT < 10ms)
© 2010 IBM Corporation
A practical Introduction to Disk Storage System Performance
16 IBM Power Systems and Storage Symposium, Wiesbaden, Germany – May 10-12, 20102010-09-13
Agenda
• Disk Storage System Selection & Specs
• Application I/O & Workload Characteristics
• Hard Disk Drive (HDD) Basics – It‟s all mechanic
• HDD Performance & Capacity Aspects (SATA vs FC)
• RAID Level Considerations (RAID-5 / RAID-6 / RAID-10)
• New Trends & Directions: Solid State Drive (SSD)
• Basic Principles for Planning Logical Configurations
• Performance Data Collection and Analysis
© 2010 IBM Corporation
A practical Introduction to Disk Storage System Performance
17 IBM Power Systems and Storage Symposium, Wiesbaden, Germany – May 10-12, 20102010-09-13
Hard Disk Drive (HDD) Basics – It‟s all mechanic...
Read / write cache hits are in the range ~ 1ms
Physical disk I/O operations are in the range of > 5ms because mechanicalcomponents such as head movements and spinning disks are involved
Each hard disk drive (HDD) can only process a limited no. of I/O operationsper second, mainly determined by :
– Average Seek Time [ms] (head movement to required track)
– Rotational Latency [ms] (disk platter spinning until the first sector addressed passes under the r/w heads; avg. time = half a rotation)
– Transfer Time [ms] (read/write data sectors, 1 sector = 512 Byte)
Start Seek
Time
Rotational
Latency
Transfer
Time
© 2010 IBM Corporation
A practical Introduction to Disk Storage System Performance
18 IBM Power Systems and Storage Symposium, Wiesbaden, Germany – May 10-12, 20102010-09-13
Simple IOps Calculation per Hard Disk Drive (HDD)
Avg. Seek Time = see manufacturer specs (typical: 4-10ms)
Rotational Latency = ½ (60000/RPM) [ms] (typical: 2-4ms)
Transfer Time = 1000 sectors sector size / avg. Transfer Rate [ms]
(typically << 1ms for small I/O request sizes < 16kB)
© 2010 IBM Corporation
A practical Introduction to Disk Storage System Performance
19 IBM Power Systems and Storage Symposium, Wiesbaden, Germany – May 10-12, 20102010-09-13
Manufacturer Specs for Hard Disk Drives
Source: www.seagate.com (2008)
This is just an example for getting a view on typical disk drive characteristics. The chosen disk types above do not necessarily represent
the characteristics of the disk drive modules used in IBM System Storage systems.
© 2010 IBM Corporation
A practical Introduction to Disk Storage System Performance
20 IBM Power Systems and Storage Symposium, Wiesbaden, Germany – May 10-12, 20102010-09-13
A single disk driveis only capable of processing a
limited number of I/O operations per second!
Example Random IOps Calculation per Hard Disk Drive
Disk Drive SpeedRotational
Latency
Avg. Seek
TimeIOps
FC
146GB15k15000 rpm 2 ms 4 ms 167
FC
146GB10k10000 rpm 3 ms 5 ms 125
SATA2
500GB7.2k7200 rpm 4.2 ms 9 ms 76
Rules of Thumb - Random IOps/HDD (conservative estimate to start with):
• FC 15k DDM : ~160 IOps
• FC 10k DDM : ~120 IOps
• SATA2 7.2k DDM: ~75 IOps
© 2010 IBM Corporation
A practical Introduction to Disk Storage System Performance
21 IBM Power Systems and Storage Symposium, Wiesbaden, Germany – May 10-12, 20102010-09-13
Efforts to improve HDD Performance
Efforts to reduce HDD access times (mechanical delays)
Disk Drive: Introduce Command Queuing and Re-Ordering of I/Os
SATA: NCQ (Native Command Queuing)
SCSI: TCQ (Tagged Command Queuing)
Disk Drive Usage: 'Short Stroking' of HDDs
Disk Subsystem: Subsystem Cache
Intelligent Cache Page Replacement & Prefetching Algorithms
Standard: LRU (least recently used) / LFU (least frequently used)
IBM System Storage DS8000 - Advanced Caching Algorithms
2004 – ARC (Adaptive Replacement Cache)
2007 – AMP (Adaptive Multi-stream Prefetching)
2009 – IWC (Intelligent Write Caching)
IBM Almaden Research Center - Storage Systems Caching Technologieshttp://www.almaden.ibm.com/storagesystems/projects/arc/technologies/
Seek latency optimization
Caching / Cache Hits
© 2010 IBM Corporation
A practical Introduction to Disk Storage System Performance
22 IBM Power Systems and Storage Symposium, Wiesbaden, Germany – May 10-12, 20102010-09-13
Increase HDD Performance - Command Queuing
Tagged Command Queuing (TCQ, SCSI-2) & Native Command Queuing (NCQ, SATA2)
further improves disk drive random access performance by re-ordering the I/O commands so that workloads can experience seek times which are considerably less than the nominal seek times
Queue Depth: SATA2 (NCQ): 32 in-flight commands, SCSI (TCQ): 2^64 in-flight commands
© 2010 IBM Corporation
A practical Introduction to Disk Storage System Performance
23 IBM Power Systems and Storage Symposium, Wiesbaden, Germany – May 10-12, 20102010-09-13
Increase HDD Performance - Short Stroking
Short Stroking:
Approach to achieve maximum possible
performance from an HDD by
limiting the overall head movement and
thus minimizing the average seek time.
Implementation:
- Use only a small portion of overall capacity
- Use tracks on outer edge with higher data density
Disadvantage:
- Typically large number of HDDs involved
- Only small portion of storage capacity used
Typical usage:
Applications with high access densities (IOps/GB) that require high random I/O rates at
low response times but with only a comparatively small amount of data.
© 2010 IBM Corporation
A practical Introduction to Disk Storage System Performance
24 IBM Power Systems and Storage Symposium, Wiesbaden, Germany – May 10-12, 20102010-09-13
Increase HDD Performance - Subsystem Cache
Disk Subsystem Cache
–Read Cache Hits
–Write Cache Hits / Write behind
–Sequential Prefetch Algorithms
Intelligent Cache Page Replacement & Prefetch Algorithms
–What data should be stored in cache based upon the
recent access and frequency needs of the hosts (LRU/LFU)?
–Determine what data in cache can be removed to
accommodate newer data.
–Predictive algorithms to anticipate data prior to a host request and
loading it into cache.
© 2010 IBM Corporation
A practical Introduction to Disk Storage System Performance
25 IBM Power Systems and Storage Symposium, Wiesbaden, Germany – May 10-12, 20102010-09-13
Sample Random IOps Calculation with reduced ⅓ Seek Times
Disk Drive SpeedRotational
Latency
Avg. Seek
Time
⅓ Reduced
Seek Time
IOps
Red. Seek
FC
146GB15k15000 rpm 2 ms 4 ms 4/3 ms 300
FC
146GB10k10000 rpm 3 ms 5 ms 5/3 ms 214
SATA2
500GB7.2k7200 rpm 4.2 ms 9 ms 9/3 ms 138
Even with reduced average seek times you cannot expect more than
a few hundred random I/O operations per second from a single HDD.
So a single HDD can only process a limited number of random IOps with
average access times in the typical range of 5...15ms due to the mechanical
delays associated with spinning disks (HDDs).
© 2010 IBM Corporation
A practical Introduction to Disk Storage System Performance
26 IBM Power Systems and Storage Symposium, Wiesbaden, Germany – May 10-12, 20102010-09-13
Storage Disk Subsystem – Typical I/O Rate & Response Time Relation
Response Time versus I/O Rate
0
5
10
15
20
25
30
0
1000
2000
3000
4000
5000
6000
7000
8000
9000
10000
11000
Total I/O [IO/s]
Resp
on
se T
ime [
ms]
+/- 10% change in I/O rate
© 2010 IBM Corporation
A practical Introduction to Disk Storage System Performance
27 IBM Power Systems and Storage Symposium, Wiesbaden, Germany – May 10-12, 20102010-09-13
Agenda
• Disk Storage System Selection & Specs
• Application I/O & Workload Characteristics
• Hard Disk Drive (HDD) Basics – It‟s all mechanic
• HDD Performance & Capacity Aspects (SATA vs FC)
• RAID Level Considerations (RAID-5 / RAID-6 / RAID-10)
• New Trends & Directions: Solid State Drive (SSD)
• Basic Principles for Planning Logical Configurations
• Performance Data Collection and Analysis
© 2010 IBM Corporation
A practical Introduction to Disk Storage System Performance
28 IBM Power Systems and Storage Symposium, Wiesbaden, Germany – May 10-12, 20102010-09-13
Subsystem Sizing – Meeting Performance and Capacity Requirements
Capacity:
– number of disk drives to meet capacity requirements
– only low no. of large capacity disks required to meet capacity needs
Performance:
– number and speed of disk drives (spindles) to meet IOps requirements
– high no. of fast, low capacity drives required to meet performance needs
Cost:
COST
CapacityPerformance
higher lower 146GB15k drives are
an excellent trade off
between performance
and capacity needs
IOps
GB
no. of drives drive capacity
© 2010 IBM Corporation
A practical Introduction to Disk Storage System Performance
29 IBM Power Systems and Storage Symposium, Wiesbaden, Germany – May 10-12, 20102010-09-13
1x 1TB 7.2k SATA
(75 IOps/HDD; 9.8W)
75 IOps
1000 GB
9.8 W
7x 146GB15k FC
(160 IOps/HDD; 15W)
1120 IOps
1022 GB
105 W
Subsystem Sizing – Meeting Performance and Capacity Requirements
Application: Capacity 1000GB; Performance 1000 IOps (1.0 IOps/GB)
FC
SATA
14x 1TB 7.2k SATA
(75 IOps/HDD; 9.8W)
1050 IOps
14000 GB (!)
137.2 W
SATA SATA
© 2010 IBM Corporation
A practical Introduction to Disk Storage System Performance
30 IBM Power Systems and Storage Symposium, Wiesbaden, Germany – May 10-12, 20102010-09-13
2005
0.7
Average Access Density over recent Years
Access Density is a measure of I/O throughput per unit of usable storage capacity (backstore). The primary use of access
density is to identify a range on a response time curves to give the typical response time expected by the average customer,
based on the amount of total usable storage in their environment. The average industry value for access density in the year
2005 is thought to be approximately 0.7 I/Os per second per GB. Year-to-year industry data is incomplete, but the value has
been decreasing as companies acquire usable storage faster than they access it.
Source: IBM data, other consultants
IOps
GB
= Access Density
[IOps/GB]
cold
hot
data
© 2010 IBM Corporation
A practical Introduction to Disk Storage System Performance
31 IBM Power Systems and Storage Symposium, Wiesbaden, Germany – May 10-12, 20102010-09-13
Average Access Density – Customer Distribution
Cumulative Customer Percentages vs Access Density
0%
20%
40%
60%
80%
100%
120%
0.01
0.10
0.20
0.30
0.40
0.50
0.60
0.70
0.80
0.90
1.00
1.10
1.20
1.30
1.40
1.50
1.60
1.70
1.80
1.90
2.00
Average access density ~ 0.7 IO/sec/GB (2005)
Note: Chart is based on a survey of 58 customers in 2005.
FC
14
6G
B1
5k
FC
73
GB
15
k
FC
30
0G
B1
5k
SA
TA
1T
B7.2
k
SA
TA
50
0G
B7
.2k
Access Density [IOps/GB]
datacold hot
© 2010 IBM Corporation
A practical Introduction to Disk Storage System Performance
32 IBM Power Systems and Storage Symposium, Wiesbaden, Germany – May 10-12, 20102010-09-13
SATA vs FC - HDD Performance Positioning
up to 80%
around 45% of FC 15kSATA 7.2k
Fibre Channel (FC) disk drives / Serial Attached SCSI (SAS)
Offer highest enterprise-class performance, reliability, and availability for business-critical applications requiring high I/O transaction performance
Serial Advanced Technology Attachment (SATA) disk drives
Price-attractive alternative to the enterprise class FC drives for near-line applications with lower production costs, larger capacities but also lower specifications (e.g.rotational speeds, data rates, seek times)
SATA vs. FC Drive Positioning & Considerations
sequential workloads:
SATA drives perform quite well with only about 20% reduction inthroughput compared to FC drives.
random workloads:
SATA drive transaction performance is considerably below FCdrives and their use in environments with critical online transactionworkloads and lowest response times is not generally recommended!
SATA drives typically are very well suited for various fixed content, data archival, reference data, and near-line applications that require large amounts of data at low cost, e.g. bandwidth / streaming applications, audio/video streaming, surveillance data, seismic data, medical imaging or secondary storage. They also can be a reasonable choice for business critical applications in selected environments with less critical IOPS performance requirements (e.g. low access densities).
© 2010 IBM Corporation
A practical Introduction to Disk Storage System Performance
33 IBM Power Systems and Storage Symposium, Wiesbaden, Germany – May 10-12, 20102010-09-13
Agenda
• Disk Storage System Selection & Specs
• Application I/O & Workload Characteristics
• Hard Disk Drive (HDD) Basics – It‟s all mechanic
• HDD Performance & Capacity Aspects (SATA vs FC)
• RAID Level Considerations (RAID-5 / RAID-6 / RAID-10)
• New Trends & Directions: Solid State Drive (SSD)
• Basic Principles for Planning Logical Configurations
• Performance Data Collection and Analysis
© 2010 IBM Corporation
A practical Introduction to Disk Storage System Performance
34 IBM Power Systems and Storage Symposium, Wiesbaden, Germany – May 10-12, 20102010-09-13
RAID Level Comparison - RAID5 vs RAID10
RAID5
• cost-effective with regard to performance and usable capacity (87.5% usable capacity for 7+P)
• provides fault tolerance for one disk drive failure
• data is striped across all drives in the array with the parity being distributed across all the drives
• A single random small block write operation typically causes a “RAID5 write penalty”, initiating four I/O operations to the disk back-end by reading the old data and the old parity block before finally writing the new data and the new parity block (this is kind of a worst-case scenario – it may take less operations when writing partial or even full stripes dependent on the I/Os in cache).
• On modern disk systems write operations are generally cached by the storage subsystem and thus handled asynchronously so that RAID5 write penalties are generally shielded from the users in terms of disk response time. However, with steady and heavy random write workloads, the cache destages to the back-end may still become a limiting factor so that either more disks or a RAID10 configuration might be required to provide sufficient disk back-end write performance.
RAID10
• best choice for fault-tolerant, write-sensitive environments at the cost of 50% usable capacity
• can tolerate at least one, and in most cases even multiple disk failures.
• data is striped across several disks and the first set of disk drives is mirrored to an identical set.
• each write operation initiates two write operations at the disk back-end
© 2010 IBM Corporation
A practical Introduction to Disk Storage System Performance
35 IBM Power Systems and Storage Symposium, Wiesbaden, Germany – May 10-12, 20102010-09-13
RAID5 – Writing a single data block
ParityRAID5 (7+P) ARRAY
RAID5 - Read-Modify-Write: RAID5 Write Penalty
Worst case scenario with one write operation requiring four disk operations to array
(1) read old data (2) read old parity [ MODIFY ] (3) write new data (4) write new parity
= data being read from disk = data being written to disk
(1) (2)(3) (4)Cache
performing XOR calculation
© 2010 IBM Corporation
A practical Introduction to Disk Storage System Performance
36 IBM Power Systems and Storage Symposium, Wiesbaden, Germany – May 10-12, 20102010-09-13
Cache
RAID5 – Writing a full stripe
ParityRAID5 (7+P) ARRAY
RAID5 - Full Stripe Write
Especially with large I/O transfer sizes or sequential workoads full stripe writes
can be accomplished with RAID5 where the parity can be calculated on the fly
without the need to read any old data from the array prior to the write operation
= data being read from disk = data being written to disk
© 2010 IBM Corporation
A practical Introduction to Disk Storage System Performance
37 IBM Power Systems and Storage Symposium, Wiesbaden, Germany – May 10-12, 20102010-09-13
RAID5 vs RAID10 – Backend I/O rate calculation example
Sustained front-end I/O rate: 1000 IOps (70:30:50)
Example for a typical 70:30:50 random, small-block application workload
(Read:write ratio = 70:30; Read cache hit ratio = 50%)
Sustained back-end I/O rate: 1550 IOps RAID5 vs 950 IOps RAID10
RAID5 : 1000 logical random IOps
– 700 reads 50% cache hits = 350 reads
– 300 writes 4 (write penalty: read old data/parity, write new data/parity)
= 1200 reads & writes
– a total of 1550 physical IOps on the disks at the physical backend
RAID10 : 1000 logical random IOps
– 700 Reads 50% Cache Hits = 350 Reads
– 300 Writes 2 (two mirrored writes) = 600 Writes
– a total of 950 physical IOps on the disks at the physical backend
RAID10 already outperforms RAID5 in a typical 70-30-50 workload.!!! Consider using RAID10 if random write percentage is higher than 35% !!!
© 2010 IBM Corporation
A practical Introduction to Disk Storage System Performance
38 IBM Power Systems and Storage Symposium, Wiesbaden, Germany – May 10-12, 20102010-09-13
RAID5 vs RAID10 – Performance summary
RAID5 vs RAID10 - Performance
RAID5 and RAID10 basically deliver a comparable performance for read operations.
RAID5 typically performs better than RAID10 for large block sequential writes.
RAID10 always performs better than RAID5 for small block random writes.
RAID5 vs RAID10 - Selection
RAID5 is a good choice for most environments requiring high availability and fewer writes than reads (e.g. multi-user environments with transaction database applications and a high read activity).
RAID10 should be considered for fault-tolerant and performance-critical, write-sensitive transaction processing environments with a high random write percentage above 35%.
RAID levelRandom
Read
Random
Write
Sequential
Read
Sequential
Write
Capacity
8-DDMs
RAID5 + o + + 87.5%
RAID10 + + + o 50.0%
© 2010 IBM Corporation
A practical Introduction to Disk Storage System Performance
39 IBM Power Systems and Storage Symposium, Wiesbaden, Germany – May 10-12, 20102010-09-13
RAID6 - Overview
RAID LevelReliability
(#Erasures)Space efficiency
Write penalty
(Disk ops)
RAID-5, 7+P 1 87.5% 4
RAID-10, 4+4 at least 1 50% 2
RAID-6, 6+P+Q 2 75% 6
RAID6: Dual parity RAID– DS8000: 5+P+Q+S or 6+P+Q arrays (using modified EVENODD code)
– Survives 2 “erasures”
• 2 drive failures
• 1 drive failure plus a medium error, such as during rebuild
(especially with large capacity drives)
– Like RAID5, parity is distributed in stripes, with the parity blocks in a different place in
each stripe
– RAID6 does have a higher performance penalty on write operations than RAID5 due to
the additional parity calculations.
RAID Level Comparison:
© 2010 IBM Corporation
A practical Introduction to Disk Storage System Performance
40 IBM Power Systems and Storage Symposium, Wiesbaden, Germany – May 10-12, 20102010-09-13
DS8000 - Single Rank RAID Performance (1/2)
DS8000 R4.0
(no IWC)
full stroke
© 2010 IBM Corporation
A practical Introduction to Disk Storage System Performance
41 IBM Power Systems and Storage Symposium, Wiesbaden, Germany – May 10-12, 20102010-09-13
DS8000 - Single Rank RAID Performance (2/2)
RAID6 RAID10RAID5
DS8000 R4.0
(no IWC)
full stroke
© 2010 IBM Corporation
A practical Introduction to Disk Storage System Performance
42 IBM Power Systems and Storage Symposium, Wiesbaden, Germany – May 10-12, 20102010-09-13
Agenda
• Disk Storage System Selection & Specs
• Application I/O & Workload Characteristics
• Hard Disk Drive (HDD) Basics – It‟s all mechanic
• HDD Performance & Capacity Aspects (SATA vs FC)
• RAID Level Considerations (RAID-5 / RAID-6 / RAID-10)
• New Trends & Directions: Solid State Drive (SSD)
• Basic Principles for Planning Logical Configurations
• Performance Data Collection and Analysis
© 2010 IBM Corporation
A practical Introduction to Disk Storage System Performance
43 IBM Power Systems and Storage Symposium, Wiesbaden, Germany – May 10-12, 20102010-09-13
Processing Capabilities and Disk Performance over 50 years
1956 IBM RAMAC (1st disk drive) 5 MB storage, 1200 RPMdata transfer rate 8800 characters per second
2010 Enterprise FC Hard Disk Drive (HDD)600GB storage capacity, 15000 RPMdata transfer rate 122 to 204 MB/s
Last 50 years of HDD technology:
HDD RPM: 12.5 x
HDD Capacity 120 000 x
Op
era
tio
ns p
er
se
co
nd
Time
Performance
Gap
New: SSD drives (STEC-inc):
0.1 MHz
4 GHz
© 2010 IBM Corporation
A practical Introduction to Disk Storage System Performance
44 IBM Power Systems and Storage Symposium, Wiesbaden, Germany – May 10-12, 20102010-09-13
What are solid-state drives?
• Semiconductor (NAND flash, non-volatile)
• No mechanical read/write interface, no rotating parts:
i.e. no seek time or rotational delays
• Electronically erasable medium
• Random access storage
• Capable of driving tens of thousands of IOps
with response times less than 1ms
• Absence of mechanical moving parts makes SSDs
significantly more reliable than HDDs
• Wear issues are overcome through over-provisioning
and intelligent controller algorithms (Wear-Levelling)
Application benefits
Increased performance for transactional applications with high random IO rates (IOps):
Online Banking / ATM / Currency Trading, Point-of-Sale Transactions / Processing, Real-time data mining
Solid state disks in DS8000 offer a new higher performance option for enterprise applications.
Best suited for cache-unfriendly data with high access densities (IOps/GB) requiring low response times
Additional benefit of lower energy consumption, cooling and space requirements (data center footprint)
New Trends & Directions - Solid State Drives (SSD)
© 2010 IBM Corporation
A practical Introduction to Disk Storage System Performance
45 IBM Power Systems and Storage Symposium, Wiesbaden, Germany – May 10-12, 20102010-09-13
Solid State Drive (SSD) - DS8000 R4.2 Single Rank Performance
Single RAID5 Rank - Random Read Single RAID5 Rank - Random I/O
Single RAID5 Rank - Random Read
Single RAID5 Rank - Sequential I/O
Random I/O: SSDs >> HDDs
Sequential I/O: SSDs ~ HDDs
RAID5 Write-Penalty
(1:4 Backend Ops)
SSDs show exceptionally low response times
Source: IBM Whitepaper, IBM System Storage DS8000 with SSDs - An In-Depth Look at SSD Performance in the DS8000,
http://www.ibm.com/support/techdocs/atsmastr.nsf/WebIndex/WP101466
© 2010 IBM Corporation
A practical Introduction to Disk Storage System Performance
46 IBM Power Systems and Storage Symposium, Wiesbaden, Germany – May 10-12, 20102010-09-13
Solid State Drive (SSD) - Tiered Storage Concepts
Solid State Drive technology remains more expensive than traditional spinning disks, so the two
technologies will coexist in hybrid configurations for several years.
Tiered storage is an approach of utilizing different types of storage throughout the storage infrastructure.
Using the right mix of tier 0, 1, and 2 drives will provide optimal performance at the minimum cost, power,
cooling and space usage.
Data Placement is key! To maximize the benefit of SSDs it is important to analyze application workloads and
only place data which requires high access densities (IOps/GB) and low response times on them.
IBM System Storage DS8000 with SSDs - An In-Depth Look at SSD Performance in the DS8000
http://www.ibm.com/support/techdocs/atsmastr.nsf/WebIndex/WP101466
Driving Business Value on Power Systems with Solid State Drives
ftp://ftp.software.ibm.com/common/ssi/sa/wh/n/pow03025usen/POW03025USEN.PDF
Tier 0Solid State Drives (SSD):
Highest performance and cost/GB
Tier 115k RPM HDDs (FC/SAS):
High performance lower cost/GB
Tier 27200 RPM HDDs (SATA):
Lowest performance and cost/GB
SSD
whitepapers
cold
hot
data
© 2010 IBM Corporation
A practical Introduction to Disk Storage System Performance
47 IBM Power Systems and Storage Symposium, Wiesbaden, Germany – May 10-12, 20102010-09-13
The challenges with SSDs
SSDs are considerably more expensive than
traditional disks
Without optimization tools, clients have been over-
provisioning them
And administrators spend too much time
monitoring, reporting, and tuning tiers
Inefficient use of a very expensive asset is difficult to justify
Result: Many clients feel they can‟t afford solid-state storage yet
© 2010 IBM Corporation
A practical Introduction to Disk Storage System Performance
48 IBM Power Systems and Storage Symposium, Wiesbaden, Germany – May 10-12, 20102010-09-13
Solid-state drives (SSDs) offer significantly improved performance compared to
mechanical disk drives... but it takes more than just supporting SSDs in a disk
subsystem for clients to achieve the full benefit:
Task: Optimizing data placement across tiers of drives with different price and
performance attributes can help clients operate at peak price/performance.
Implementing this type of optimization is a three-step process:
(1) Data performance information must be collected.
(2) Information must be analyzed to determine optimal data placement.
(3) Data must be relocated to the optimal tier.
Solution: With DS8700 R5.1 IBM introduced IBM System Storage Easy Tier
which automates data placement throughout the DS8700 disk pool (including
multiple drive tiers) to intelligently align the system with current workload
requirements. This includes the ability for the system to automatically and
nondisruptively relocate sub-volume data (at the extent level) across drive tiers,
and the ability to manually relocate full volumes or merge extent pools. Easy
Tier enables smart data placement and optimizes SSD deployments with
minimal costs. The additional Storage Tier Advisor Tool provides guidance
for SSD capacity planning based on existing client workloads on the DS8700.
IBM System Storage DS8700 R5.1 Announcement Letter (Easy Tier)
http://www.ibm.com/common/ssi/rep_ca/5/877/ENUSZG10-0125/ENUSZG10-0125.PDF
IBM Redpaper: IBM System Storage DS8700 Easy Tier
http://www.redbooks.ibm.com/abstracts/redp4667.html?Open
IBM DS8700 R5.1 Solid-State Storage Optimization with Easy Tier
© 2010 IBM Corporation
A practical Introduction to Disk Storage System Performance
49 IBM Power Systems and Storage Symposium, Wiesbaden, Germany – May 10-12, 20102010-09-13
Easy Tier optimizes SSD deployments by balancing performance AND cost requirements
Easy Tier delivers the full promise of SSD performance while balancing the costs associated
with over provisioning this expensive resource
“Just Right”“Slower, inexpensive” “Fast, expensive”
IBM Easy Tier
LUN Heatmap
© 2010 IBM Corporation
A practical Introduction to Disk Storage System Performance
50 IBM Power Systems and Storage Symposium, Wiesbaden, Germany – May 10-12, 20102010-09-13
Smart data placement with Easy Tier: SPC-1 (SATA/SDD)
First ever Storage Performance Council (SPC-1) benchmark
submission with SATA and SSD technology
Source:
Storage Performance Council, April 2010: http://www.storageperformance.org/results/benchmark_results_spc1#a00092
IBM Whitepaper, May 2010: IBM® System Storage™ DS8700™ Performance with Easy Tier®, http://www.ibm.com/support/techdocs/atsmastr.nsf/WebIndex/WP101675
Easy Tier
0:00 2:00 4:00 6:00 8:00 10:00 12:00 14:00 16:00 18:00
Time
Th
ro
ug
hp
ut (IO
/s)
Over 3x IOPSImprovement
Increase of
over 3X!
System configuration:
16x SSD + 96x 1TB SATA
© 2010 IBM Corporation
A practical Introduction to Disk Storage System Performance
51 IBM Power Systems and Storage Symposium, Wiesbaden, Germany – May 10-12, 20102010-09-13
Smart data placement with Easy Tier: SPC-1 (SATA/SDD) SSD + SATA + Easy Tier Config vs. FC 15K HDDs Config
0.00
5.00
10.00
15.00
0 10000 20000 30000 40000 50000 60000
Throughput (IO/s)
Re
sp
on
se
Tim
e (
ms
)
192 FC HDD 96 SATA + 16 SSD
Improves RT in rangeof ordinary use
Dual frames Single Frame
© 2010 IBM Corporation
A practical Introduction to Disk Storage System Performance
52 IBM Power Systems and Storage Symposium, Wiesbaden, Germany – May 10-12, 20102010-09-13
Smart data placement with Easy Tier: SPC-1 Backend I/O Migration
0
1
2
3
4
5
6
0 10 20 30 40 50 60 70 80 90
% Backend IO migrated
% C
ap
acit
y m
igra
ted
© 2010 IBM Corporation
A practical Introduction to Disk Storage System Performance
53 IBM Power Systems and Storage Symposium, Wiesbaden, Germany – May 10-12, 20102010-09-13
Agenda
• Disk Storage System Selection & Specs
• Application I/O & Workload Characteristics
• Hard Disk Drive (HDD) Basics – It‟s all mechanic
• HDD Performance & Capacity Aspects (SATA vs FC)
• RAID Level Considerations (RAID-5 / RAID-6 / RAID-10)
• New Trends & Directions: Solid State Drive (SSD)
• Basic Principles for Planning Logical Configurations
• Performance Data Collection and Analysis
© 2010 IBM Corporation
A practical Introduction to Disk Storage System Performance
54 IBM Power Systems and Storage Symposium, Wiesbaden, Germany – May 10-12, 20102010-09-13
Logical Configuration - Basic Principles
Three major principles for the logical configuration to optimize storage subsystem performance:
Workload isolation (e.g. on extent pool and array level)– dedicate a subset of hardware resources to a high priority workload in order to reduce impacts of
less important workloads (protect the loved ones) and meet given service level agreements (SLAs)– limit low priority workloads which tend to fully utilize given resources to only a subset of hardware
resources in order to avoid impacting other more important workloads (isolate the badly behaving ones)
– provides guaranteed availability of the dedicated hardware resources but also limits the isolated workload to only a subset of the total subsystem resources and overall subsystem performance
Workload resource sharing– multiple workloads share a common set of subsystem hardware resources, such as arrays,
adapters, ports– single workloads now can utilize more subsystem resources and experience a higher performance
than with only a smaller subset of dedicated resources if the workloads do not show contention with each other
– good approach when workload information is not available, with workloads that do not try to consume all the hardware resources available, or with workloads that show workload peaks at different times
Workload spreading– most important principle of performance optimization, applies to both isolated workloads and
resource-sharing workloads– simply means using all available resources of the storage subsystem in a balanced manner by
spreading the workload evenly across all available resources that are dedicated to that workload, e.g. arrays, controllers, disk adapters, host adapters, host ports
– host-level striping and multi-pathing software may further help to spread workloads evenly
(1) Workload isolation (2) Workload resource-sharing (3) Workload spreading
© 2010 IBM Corporation
A practical Introduction to Disk Storage System Performance
55 IBM Power Systems and Storage Symposium, Wiesbaden, Germany – May 10-12, 20102010-09-13
Logical Configuration – DS8000 Examples
© 2010 IBM Corporation
A practical Introduction to Disk Storage System Performance
56 IBM Power Systems and Storage Symposium, Wiesbaden, Germany – May 10-12, 20102010-09-13
Logical Configuration – DS8000 Examples
© 2010 IBM Corporation
A practical Introduction to Disk Storage System Performance
57 IBM Power Systems and Storage Symposium, Wiesbaden, Germany – May 10-12, 20102010-09-13
Agenda
• Disk Storage System Selection & Specs
• Application I/O & Workload Characteristics
• Hard Disk Drive (HDD) Basics – It‟s all mechanic
• HDD Performance & Capacity Aspects (SATA vs FC)
• RAID Level Considerations (RAID-5 / RAID-6 / RAID-10)
• New Trends & Directions: Solid State Drive (SSD)
• Basic Principles for Planning Logical Configurations
• Performance Data Collection and Analysis
© 2010 IBM Corporation
A practical Introduction to Disk Storage System Performance
58 IBM Power Systems and Storage Symposium, Wiesbaden, Germany – May 10-12, 20102010-09-13
Analyzing Disk Subsystem I/O Performance
Questions to ask when a performance problem occurs: What exactly is considered to perform poorly? Which application, server, volumes?
Is there a detailed description of the performance problem and environment available?
What is the actual business impact of the performance problem?
What was the first occurrance of the problem and were there any changes in the environment?
When does the problem typically occur, e.g. during daily business hours or nightly batch runs?
What facts indicate that the performance problem is related to the storage subsystem?
What would be the criteria for the problem to be considered as solved? Any expectations?
Data to collect and analyze:
description & config of the architecture (application – server – SAN – storage)
application characteristics, logical and physical volume layout (usage, mapping server/storage)
I/O performance data collection during problem occurrance on server and storage subsystem:
(a) Server Performance Data Collection:
AIX # iostat (–D) [interval] [no. of intervals]# filemon –o fmon.log –O lv,pv; sleep 60; trcstop
Linux # iostat –x [interval] [no. of intervals]
Windows # perfmon → GUI, then select Physical Disk Counters
(b) Storage Subsystem Performance Data Collection:
DS3k/DS4k/DS5k (SMcli), XIV (XCLI), DS6k/DS8k and other (TPC for Disk)
© 2010 IBM Corporation
A practical Introduction to Disk Storage System Performance
59 IBM Power Systems and Storage Symposium, Wiesbaden, Germany – May 10-12, 20102010-09-13
DS3000/4000/5000 Performance Monitor
• only counters for quantity of processed I/Os up to current point in time
• no counters for quality of processed I/Os as, for example, I/O service times
• additional host system performance statistics required for I/O response times
© 2010 IBM Corporation
A practical Introduction to Disk Storage System Performance
60 IBM Power Systems and Storage Symposium, Wiesbaden, Germany – May 10-12, 20102010-09-13
DS3000/4000/5000 Performance Data Collection
SMcli script for continuous performance data collection over given time frame:
on error stop; set performanceMonitor interval=60 iterations=1440; upload storageSubsystem file="c:\perf01.txt" content=performanceStats;
>smcli [IP-Addr. Ctr.A] [IP-Addr. Ctr.B] -f perfmon.scrPerforming syntax check...Syntax check complete.Executing script...Script execution complete.SMcli completed successfully.
Always collect the
Performance Statistics
together with latest
Subsystem Profile
to document the actual
subsystem configuration
used during data collection
perfmon.scr
© 2010 IBM Corporation
A practical Introduction to Disk Storage System Performance
61 IBM Power Systems and Storage Symposium, Wiesbaden, Germany – May 10-12, 20102010-09-13
DS3000/4000/5000 Performance Data Collection Example
"Performance Monitor Statistics for Storage Subsystem: DS4700_PFE1 -Date/Time: 12.02.08 10:29:13 - Polling interval in seconds: 20"
"Storage Subsystems ","Total IOs ","Read Percentage ","Cache Hit Percentage ","Current KB/second ","Maximum KB/second ","Current IO/second ","Maximum IO/second"
"Capture Iteration: 1","","","","","","","""Date/Time: 12.02.08 10:29:14","","","","","","","""CONTROLLER IN SLOT A","0.0","0.0","0.0","0.0","0.0","0.0","0.0""Logical Drive Data_1","0.0","0.0","0.0","0.0","0.0","0.0","0.0""Logical Drive Data_3","0.0","0.0","0.0","0.0","0.0","0.0","0.0"[...]"CONTROLLER IN SLOT B","0.0","0.0","0.0","0.0","0.0","0.0","0.0""Logical Drive Data_2","0.0","0.0","0.0","0.0","0.0","0.0","0.0""Logical Drive Data_4","0.0","0.0","0.0","0.0","0.0","0.0","0.0"[...]"STORAGE SUBSYSTEM TOTALS","0.0","0.0","0.0","0.0","0.0","0.0","0.0"[...]
Example of performance statistics file collected on DS4000 with v7.xx firmware
For more information about how to collect and process these DS4000 performance statistics please see:
How to collect performance statistics on IBM DS3000 and DS4000 subsystems (on IBM Techdocs)
IBMers http://w3.ibm.com/support/techdocs/atsmastr.nsf/WebIndex/TD103963
IBM BPs http://partners.boulder.ibm.com/src/atsmastr.nsf/WebIndex/TD103963
(same format as DS3000/DS5000 performance statistics)
© 2010 IBM Corporation
A practical Introduction to Disk Storage System Performance
62 IBM Power Systems and Storage Symposium, Wiesbaden, Germany – May 10-12, 20102010-09-13
DS3000/4000/5000 – Performance Data Analysis
Subsystem total IOps / MBps (average / peak)
Controller A and B total IOps / MBps
Identify busiest volumes
Identify busiest arrays
Verify if
- Array/volume configuration
- RAID level
- Disk type
is appropriate for the workload
Verify if workload distribution
is balanced across all arrays
and both controllers
Evaluate response times with
appropriate Disk Magic models
© 2010 IBM Corporation
A practical Introduction to Disk Storage System Performance
63 IBM Power Systems and Storage Symposium, Wiesbaden, Germany – May 10-12, 20102010-09-13
XIV – XIVGUI Performance Data Collection
© 2010 IBM Corporation
A practical Introduction to Disk Storage System Performance
64 IBM Power Systems and Storage Symposium, Wiesbaden, Germany – May 10-12, 20102010-09-13
XIV – XCLI Performance Data Collection
XCLI (one command line):
>xcli -m IPADDR -u USER -p PASSWD -s -y
statistics_get start=2009-10-07.11:00 count=300
interval=1 resolution_unit=minute > C:\xiv_20091007.csv
© 2010 IBM Corporation
A practical Introduction to Disk Storage System Performance
65 IBM Power Systems and Storage Symposium, Wiesbaden, Germany – May 10-12, 20102010-09-13
DS6000/DS8000 – DSCLI Performance Metrics Examplesdscli> showfbvol -metrics 2000
Date/Time: 24. April 2007 14:32:15 CEST IBM DSCLI Version: 5.2.2.224 DS: IBM.2107-7503461
ID 2000
Date 04/24/2007 14:30:25 CEST
normrdrqts 17
normrdhits 5
normwritereq 121050
normwritehits 121050
seqreadreqs 0
seqreadhits 0
seqwritereq 151127
seqwritehits 151127
cachfwrreqs 0
cachfwrhits 0
cachefwreqs 0
cachfwhits 0
inbcachload 0
bypasscach 0
DASDtrans 29
seqDASDtrans 0
cachetrans 33315
NVSspadel 0
normwriteops 0
seqwriteops 0
reccachemis 2
qwriteprots 0
CKDirtrkac 0
CKDirtrkhits 0
cachspdelay 0
timelowifact 0
phread 25
phwrite 33420
phbyteread 5
phbytewrite 2082
recmoreads 2
sfiletrkreads 0
contamwrts 0
PPRCtrks 0
NVSspallo 272177
timephread 28
timephwrite 40138
byteread 0
bytewrit 8508
timeread 4
timewrite 4061
dscli> showrank -metrics r2
Date/Time: 24. April 2007 14:37:43 CEST IBM DSCLI Version: 5.2.2.224 DS: IBM.2107-7503461
ID R2
Date 04/24/2007 14:35:53 CEST
byteread 587183
bytewrit 287002
Reads 1176760
Writes 315629
timeread 2509716
timewrite 392892
dscli> showioport -metrics I001
Date/Time: 24. April 2007 14:41:47 CEST IBM DSCLI Version: 5.2.2.224 DS: IBM.2107-7503461
ID I0001
Date 04/24/2007 14:39:56 CEST
byteread (FICON/ESCON) 0
bytewrit (FICON/ESCON) 0
Reads (FICON/ESCON) 0
Writes (FICON/ESCON) 0
timeread (FICON/ESCON) 0
timewrite (FICON/ESCON) 0
bytewrit (PPRC) 0
byteread (PPRC) 0
Writes (PPRC) 0
Reads (PPRC) 0
timewrite (PPRC) 0
timeread (PPRC) 0
byteread (SCSI) 56586
bytewrit (SCSI) 454426
Reads (SCSI) 414404
Writes (SCSI) 4906333
timeread (SCSI) 2849
timewrite (SCSI) 111272
© 2010 IBM Corporation
A practical Introduction to Disk Storage System Performance
66 IBM Power Systems and Storage Symposium, Wiesbaden, Germany – May 10-12, 20102010-09-13
TPC for Disk – Subsystem Performance Monitoring
The IBM Tivoli Storage Productivity Center (TPC) is a suite of storage infrastructure management tools for storage environments by centralizing, simplifying and automating storage tasks associated with storage systems, Storage Area Networks (SAN), replication services and capacity management.
IBM Tivoli Storage Productivity Center for Disk (TPC for Disk) is an optional component of TPC, that is designed to manage multiple SAN storage devices and to monitor the performance of SMI-S compliant storage subsystems from a single user interface.
IBM Tivoli Storage Productivity Center Standard Edition includes three components of the TPC suite as one bundle at a single price: TPC for Data, Fabric and Disk.
New customers with IBM System Storage Productivity Center (SSPC) which includes the pre-installed (but separately purchased) IBM Tivoli Storage Productivity Center Basic Edition only need to purchase the additional „TPC for Disk“ component to be able to collect performance statistics from their supported IBM storage subsystems.
TPC for Disk is the official IBM product for clients requiring performance monitoring of their IBM storage subsystems (e.g. DS4k, DS5k, DS6k, DS8k, SVC, ESS, 3584 Tape, ...)
TPC V4.1 introduces Tivoli Common Reporting (TCR) & BIRT (Business Intelligence Reporting Tools) for creating customized reports from TPC database
© 2010 IBM Corporation
A practical Introduction to Disk Storage System Performance
67 IBM Power Systems and Storage Symposium, Wiesbaden, Germany – May 10-12, 20102010-09-13
TPC for Disk – Subsystem Performance Reports
Select to initiate the report creation
3
1
2
4
© 2010 IBM Corporation
A practical Introduction to Disk Storage System Performance
68 IBM Power Systems and Storage Symposium, Wiesbaden, Germany – May 10-12, 20102010-09-13
TPC for Disk – Subsystem Performance Reports
Select for creating a chart
© 2010 IBM Corporation
A practical Introduction to Disk Storage System Performance
69 IBM Power Systems and Storage Symposium, Wiesbaden, Germany – May 10-12, 20102010-09-13
TPC for Disk – Subsystem Performance Reports
© 2010 IBM Corporation
A practical Introduction to Disk Storage System Performance
70 IBM Power Systems and Storage Symposium, Wiesbaden, Germany – May 10-12, 20102010-09-13
TPC for Disk – Export Subsystem Performance Reports
Select to export performance data as CSV output fileusing ‚File > Export Data„ dialog
© 2010 IBM Corporation
A practical Introduction to Disk Storage System Performance
71 IBM Power Systems and Storage Symposium, Wiesbaden, Germany – May 10-12, 20102010-09-13
TPC for Disk – Analyzing Reports in a Spreadsheet
© 2010 IBM Corporation
A practical Introduction to Disk Storage System Performance
72 IBM Power Systems and Storage Symposium, Wiesbaden, Germany – May 10-12, 20102010-09-13
TPC for Disk – Reports of Interest by Subsystem
ESS, DS6000 and DS8000:
By Storage Subsystem
By Controller
By Array
By Volume
By Port
SAN Volume Controller:
By Storage Subsystem
By IO Group
By Node
By Managed Disk Group
By Volume
By Managed Disk
By Port
DS4000 and other supported SMI-S compliant storage subsystems:
By Storage Subsystem
By Volume
By Port
Some reports may give more
or less data, depending on
the exact level of SMI-S
compliance by the vendor
supplied CIM agents.
Don‟t forget to export a complete set of reports for the subsystem of interest, e.g. for a DS8000:
20080131-75APNK1-subsystem.csv,
20080131-75APNK1-controller.csv,
20080131-75APNK1-ports.csv,
20080131-75APNK1-arrays.csv,
20080131-75APNK1-volumes.csv
Limit the reports to a representative time frame as the amount of data especially for the volume report can be
extremly large!
© 2010 IBM Corporation
A practical Introduction to Disk Storage System Performance
73 IBM Power Systems and Storage Symposium, Wiesbaden, Germany – May 10-12, 20102010-09-13
TPC for Disk – How to start with Performance Monitoring Simply start monitoring and thus understanding the current workload patterns (workload range
and workload profile) developing over the day/week/month for normal operation conditions where no end-user complaints are present. Develop an understanding of the expected behaviour. I/O rates and response times may vary considerably from hour to hour or day to day simply due to various application loads, business times and changes in the workload profile. You may even experience times with high I/O rates and extremly low response times (e.g. high cache hit ratios) as well as times with only moderate I/O rates but higher response times (e.g. lower cache hit ratios) still not being of any concern. Appropriate thresholds for I/O rates and response times can be derived from these statistics based on particular application and business requirements.
Regularly collect selected data sets for historical reference and do projections of workload trends. Evaluate trends in I/O rate and response time and plan for growth accordingly. Typically response times increase with increasing I/O rates. Historical performance data is the best source for performance and capacity planning.
Watch for any imbalance of the overall workload distribution across the subsystem resources. Avoid single resources from becoming overloaded (hot spots). Redistribute workload if needed.
When end-user performance complaints arise simply compare current and historical data and look for appropriate changes in the workload that may lead to performance impacts.
Additional performance metrics may help to better understand the workload profile behind the changes in I/O rates and response times:
Read:Write ratio
Read Cache Hit Percentage [%]
avg. Read/Write/Overall Transfer Size [kB] per I/O operation
Required for appropriate
Disk Magic models and
performance evaluations
© 2010 IBM Corporation
A practical Introduction to Disk Storage System Performance
74 IBM Power Systems and Storage Symposium, Wiesbaden, Germany – May 10-12, 20102010-09-13
TPC for Disk – Basic Performance Metrics There are lots of performance metrics available. Which ones are best to start with?
Most important metrics for a storage subsystem are:
I/O Rate: number of I/O operations per second [IOps or IO/s]
Response Time (RT): average service time per I/O operation in milliseconds [ms]
These metrics are typically available
for read operations, write operations and the total number of processed I/O operations
on subsystem, controller, port, array, volume, I/O group, node, mdisk & mdisk group level
Basic performance statistics to look at for storage subsystems are in principle:
front-end I/O statistics on subsystem level for overview of system overall workload
front-end I/O statistics on volume level for selected critical applications / host systems
backend I/O statistics on array level (i.e. on the physical disk level / spindles)
General thresholds for front-end statistics are difficult to provide, because
I/O rate thresholds depend on workload profile and subsystem capabilities
RT thresholds depend on application, customer requirements, business hours
Additional metric is Data Rate: throughput in megabytes per second [MBps]
on subsystem level for overview of overall throughput
on port level together with Port RT for overview of port and I/O adapter utilization
© 2010 IBM Corporation
A practical Introduction to Disk Storage System Performance
75 IBM Power Systems and Storage Symposium, Wiesbaden, Germany – May 10-12, 20102010-09-13
In general, there do not exist typical values or fixed thresholds for all performance metrics as they typically strongly depend on the nature of the workload:
Online Transaction Processing (OLTP) workloads (e.g. database)
- small transfer sizes (4kB...16kB) with high I/O rates
- low front-end response times around 5ms commonly expected
Backup, batch or sequential-like workloads
- large transfer sizes (32kB...256kB) with low I/O rates but high data rates
- high front-end response times even up to 30ms still can be acceptable
Subsystem level front-end metrics (subsystem total average):
- Overall Response Time < 10ms
Array level back-end metrics (physical disk access):
- Back-end Read Response Time < 25ms
- Disk Utilization Percentage << 80%
- I/O rate: depends on RAID level, workload profile, number and speed of DDMsconsidered very busy with I/O rates near or above 1000 I/Os (DS8000/DS6000)
Volume level front-end metrics (I/O performance as experienced by the host systems):
- Overall Response Time < 15ms (depends on application requirements and workload)
- Write-cache Delay Percentage < 3% (typically should be 0%)
TPC for Disk – Basic Guidelines for DS8000
© 2010 IBM Corporation
A practical Introduction to Disk Storage System Performance
76 IBM Power Systems and Storage Symposium, Wiesbaden, Germany – May 10-12, 20102010-09-13
TPC – Customized reports with BIRT
Two BIRT components: Report Designer and Report Runtime Engine
Report Designer Report Engine
Presentation
ServicesReport Design Engine
HTML
CSV
Eclipse
Report
Designer
Generation
Services
Data
Transform.
Services
Charting
Engine
Data
Data
Chart
Designer
Eclipse
DTP
ODA
Custom
Designer
XML
Report
Design
Report
Document
TPC V4.1
Redbook: IBM Tivoli Storage Productivity Center V4.1 Release Guide
http://www.redbooks.ibm.com/redpieces/abstracts/sg247725.html
Chapter 10, Customized Reporting through Tivoli Common Reporting (TCR) / BIRT
© 2010 IBM Corporation
A practical Introduction to Disk Storage System Performance
77 IBM Power Systems and Storage Symposium, Wiesbaden, Germany – May 10-12, 20102010-09-13
© 2010 IBM Corporation
A practical Introduction to Disk Storage System Performance
78 IBM Power Systems and Storage Symposium, Wiesbaden, Germany – May 10-12, 20102010-09-13
IBM Dynamic Infra-
structure Leadership
Center for Information
Infrastructure
Business, Channel & Skill
Enablement & Training
DI Education & Briefings
Demos & Showcases
IT Transformation Road-
maps & Workshops
BP Certification
IBM European Storage
Competence Center
& Systems Lab Europe
IBM Executive Briefing
Center & TMCC
Business, Channel & Skill
Enablement & Training
Customer and Group
Briefings
Product & SW Demos
Integrated Solution Demos
Exhibition Support &
Organization
IBM STG Europe Storage
Software Development
Software Development
Storage & Tape
Linux
Mainframe
File Systems
Storage Competence at the Mainz Location
IBM Germany„s fourth
largest location offers
you a broad portfolio of
IBM System Storage
Services Business, Channel & Skill
Enablement & Training
End-to-end client support
Workshops
Solution Design
Lab Services
Customer Relationship
Management
© 2010 IBM Corporation
A practical Introduction to Disk Storage System Performance
79 IBM Power Systems and Storage Symposium, Wiesbaden, Germany – May 10-12, 20102010-09-13
Our Services
Client Briefings &
Education
Systems Lab Services &
Training
Customized Workshops
System Storage Demos
Advanced Technical
Support
Solution Design
Proof of Concepts
Benchmarks
Product Field Engineering
Our Expertise
Skilled technical storage
experts covering the whole
IBM System Storage
Portfolio
Information Infrastructure:
- Compliance
- Availability
- Retention
- Security
HW / SW & Performance
IBM System Storage Solutions Center of Excellence
We offer technical
support from the
planning phase through
well after installation
Our Systems Lab Europe
1500 sqm lab space
IBM & heterogenous hardware
© 2010 IBM Corporation
A practical Introduction to Disk Storage System Performance
80 IBM Power Systems and Storage Symposium, Wiesbaden, Germany – May 10-12, 20102010-09-13
DisclaimerCopyright © 2010 by International Business Machines Corporation.
No part of this document may be reproduced or transmitted in any form without written permission from IBM Corporation. Product data has been
reviewed for accuracy as of the date of initial publication. Product data is subject to change without notice. This information could include technical
inaccuracies or typographical errors. IBM may make improvements and/or changes in the product(s) and/or programs(s) at any time without notice.
Any statements regarding IBM's future direction and intent are subject to change or withdrawal without notice, and represent goals and objectives only.
References in this document to IBM products, programs, or services does not imply that IBM intends to make such products, programs or services
available in all countries in which IBM operates or does business. Any reference to an IBM Program Product in this document is not intended to state or
imply that only that program product may be used. Any functionally equivalent program, that does not infringe IBM's intellectually property rights, may
be used instead. It is the user's responsibility to evaluate and verify the operation of any on-IBM product, program or service.
The performance information contained in this document was derived under specific operating and environmental conditions. The results obtained by
any party implementing the products and/or services described in this document will depend on a number of factors specific to such party‟s operating
environment and may vary significantly. IBM makes no representation that these results can be expected in any implementation of such products
and/or services. Accordingly, IBM does not provide any representations, assurances, guarantees, or warranties regarding performance.
THE INFORMATION PROVIDED IN THIS DOCUMENT IS DISTRIBUTED "AS IS" WITHOUT ANY WARRANTY, EITHER EXPRESS OR IMPLIED. IBM
EXPRESSLY DISCLAIMS ANY WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE OR NONINFRINGEMENT.
IBM shall have no responsibility to update this information. IBM products are warranted according to the terms and conditions of the agreements (e.g.,
IBM Customer Agreement, Statement of Limited Warranty, International Program License Agreement, etc.) under which they are provided. IBM is not
responsible for the performance or interoperability of any non-IBM products discussed herein.
The provision of the information contained herein is not intended to, and does not, grant any right or license under any IBM patents or
copyrights. Inquiries regarding patent or copyright licenses should be made, in writing, to:
IBM Director of Licensing, IBM Corporation, North Castle Drive, Armonk, NY, 10504-1785, U.S.A.
© 2010 IBM Corporation
A practical Introduction to Disk Storage System Performance
81 IBM Power Systems and Storage Symposium, Wiesbaden, Germany – May 10-12, 20102010-09-13
Trademarks
The following are trademarks of the International Business Machines Corporation in the United States, other countries, or both.
* All other products may be trademarks or registered trademarks of their respective companies.
Notes:
Performance is in Internal Throughput Rate (ITR) ratio based on measurements and projections using standard IBM benchmarks in a controlled environment. The actual throughput that any user will experience will vary depending upon considerations such as the amount of multiprogramming in the user's job stream, the I/O configuration, the storage configuration, and the workload processed. Therefore, no assurance can be given that an individual user will achieve throughput improvements equivalent to the performance ratios stated here.
IBM hardware products are manufactured from new parts, or new and serviceable used parts. Regardless, our warranty terms apply.
All customer examples cited or described in this presentation are presented as illustrations of the manner in which some customers have used IBM products and the results they may have achieved. Actual environmental costs and performance characteristics will vary depending on individual customer configurations and conditions.
This publication was produced in the United States. IBM may not offer the products, services or features discussed in this document in other countries, and the information may be subject to change without notice. Consult your local IBM business contact for information on the product or services available in your area.
All statements regarding IBM's future direction and intent are subject to change or withdrawal without notice, and represent goals and objectives only.
Information about non-IBM products is obtained from the manufacturers of those products or their published announcements. IBM has not tested those products and cannot confirm the performance, compatibility, or any other claims related to non-IBM products. Questions on the capabilities of non-IBM products should be addressed to the suppliers of those products.
Prices subject to change without notice. Contact your IBM representative or Business Partner for the most current pricing in your geography.
The following are trademarks or registered trademarks of other companies.
Adobe, the Adobe logo, PostScript, and the PostScript logo are either registered trademarks or trademarks of Adobe Systems Incorporated in the United States, and/or other countries.
Cell Broadband Engine is a trademark of Sony Computer Entertainment, Inc. in the United States, other countries, or both and is used under license therefrom.
Java and all Java-based trademarks are trademarks of Sun Microsystems, Inc. in the United States, other countries, or both.
Microsoft, Windows, Windows NT, and the Windows logo are trademarks of Microsoft Corporation in the United States, other countries, or both.
Intel, Intel logo, Intel Inside, Intel Inside logo, Intel Centrino, Intel Centrino logo, Celeron, Intel Xeon, Intel SpeedStep, Itanium, and Pentium are trademarks or registered trademarks of Intel
Corporation or its subsidiaries in the United States and other countries.
UNIX is a registered trademark of The Open Group in the United States and other countries.
Linux is a registered trademark of Linus Torvalds in the United States, other countries, or both.
ITIL is a registered trademark, and a registered community trademark of the Office of Government Commerce, and is registered in the U.S. Patent and Trademark Office.
IT Infrastructure Library is a registered trademark of the Central Computer and Telecommunications Agency, which is now part of the Office of Government Commerce.
LSI is a trademark or registered trademark of LSI Corporation.
AS/400®, e business(logo)®, eServer, FICON, IBM®, IBM (logo)®, iSeries®, OS/390®, pSeries®, RS/6000®, S/30, VM/ESA®, VSE/ESA, WebSphere®, xSeries®, z/OS®, zSeries®, z/VM®, System i, System i5, System p, System p5, System x, System z, System z9®, BladeCenter®, System Storage, System Storage DS®, TotalStorage®
For a complete list of IBM Trademarks, see www.ibm.com/legal/copytrade.shtml
Not all common law marks used by IBM are listed on this page. Failure of a mark to appear does not mean that IBM does not use the mark nor does it mean that the product is not
actively marketed or is not significant within its relevant market.
Those trademarks followed by ® are registered trademarks of IBM in the United States; all others are trademarks or common law marks of IBM in the United States.