Upload
gerard-bryan
View
216
Download
2
Embed Size (px)
Citation preview
DB2 I/O Performance: NCCMG
®
August 1, 2006
Speaker: Jeff Berger
Northern California CMG
Enhanced I/O Performance
On the System z9 with DB2, FICON and the DS8000
®
2
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
The following are trademarks of the International Business Machines Corporation in the United States and/or other countries.
Trademarks
The following are trademarks or registered trademarks of other companies.
* Registered trademarks of IBM Corporation
* All other products may be trademarks or registered trademarks of their respective companies.
Linux is a registered trademark of Linus Torvalds
Penguin (Tux) compliments of Larry Ewing
Java and all Java-related trademarks and logos are trademarks of Sun Microsystems, Inc., in the United States and other countries
UNIX is a registered trademark of The Open Group in the United States and other countries.
Microsoft, Windows and Windows NT are registered trademarks of Microsoft Corporation.
SET and Secure Electronic Transaction are trademarks owned by SET Secure Electronic Transaction LLC.
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.
390ACF/VTAM*AIX*APPN*CICS*DB2*e-business logo*ESCON*GDPS*Geographically Dispersed Parallel Sysplex*
FICONHiperSocketsHPRIBM *IBM logo *IMSMagstar *MVS/ESANet.Data*Netfinity
OS/2 *OS/390 *Parallel Sysplex *pSeriesRACF *RMFRS/6000 *S/390 *S/390 Parallel Enterprise ServerSysplex Timer *
Virtual Image Facility *VM/ESA *VSE/ESAVTAM *WebSphere *xSeriesz/OSzSeriesz/VM
3
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
Disclaimers
This document contains performance informationPerformance is based on measurements done in a controlled environment. The actual throughput or performance 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 or performance improvements equivalent to the numbers stated here.
4
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
Agenda
Introduction to sequential I/O performance
Review CKD architecture, track level CCWs, and MIDAWs
Introduce the z9, FICON Express 2 & 4, and the DS8000
Introduce some DB2 V8 and V9 features
The anatomy of a DB2 prefetch operation
Review of Media Manager, EF data sets, striping
DB2 logging and other DB2 functions
Performance of backup/restore, parallel queries
Using new hardware to solve your performance problems
5
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
Sequential I/O performance
This presentation is about sequential I/O, not random I/O
DB2 sequential I/O is important in the following applications:
daily full image copy cycles
processing of LOBs (large objects)
daily/weekly table unloads to feed data warehouse/marts
database recovery
Measurements of DB2 prefetch I/Os are sufficient to understand how hardware affects all of the above
6
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
0
20
40
60
80
100
120
Throughput (MB/sec)
3990-6 RVA ESS E20 ESS F20 ESS 800 DS8000 MIDAWs
EF Non-EF
DB2 Sequential Prefetch (4K pages)Exponential Growth
1993 1997 2000 2001 2002 2005 2005 ESCON FICON EXP1 FICON EXP2 200Mbit/sec 1Gbit/sec 2GBit/sec
Performance is based on measurements and projections using standard IBM benchmarks in a controlled environment. The actual throughput or performance 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 or performance improvements equivalent to the numbers stated here.
7
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
0
20
40
60
80
100
120
140
Th
rou
gh
pu
t (
MB
/sec
)
FICON Express 2 FICON Express 4 DB2 V9
DB2 Sequential Prefetch (4K pages)Continuing to improve
Performance is based on measurements and projections using standard IBM benchmarks in a controlled environment. The actual throughput or performance 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 or performance improvements equivalent to the numbers stated here.
8
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
Sequential “scenarios”
1. Single stream, data in cache Zero disconnect time
Sensitive to the speed of the channel, link and Host Adapters
2. Single stream, prestage data from disk Also may have zero disconnect time
Non-zero disconnect time indicates that the disk or Device Adapter(s) is slower than the channel
3. Highly parallel streams spread across Raid ranks Zero disconnect time
Sensitive to CU bus capacity as well as the number of channels and the channel speeds
9
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
DB2 Sequential Prefetch I/O Response Timesz9 EC, FICON Express 2, DS8000, 10K RPM 300GB drives
0
1
2
3
4
32x4K 64x4K
Prefetch quantity
I/O R
esp
on
se T
ime
(ms)
CPU Pend
Connect DisconnectIf data is cache resident,
disconnect goes away, but other components are unaffected
Disconnect time is not caused by latency
As the prefetch quantity increases…..
both connect time and disconnect time increase
the ratio of connect to disconnect decreases
10
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
DB2 Sequential Prefetch Throughputz9 EC, FICON Express 2, DS8000
0
20
40
60
80
100
120
140
160
Th
rou
gh
pu
t (M
B/s
ec)
From cache From 10K RPM 300GB disk
64x4K
8x32K
32x4K
4x32K
11
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
DB2 V8 Sequential vs Dynamic Prefetch (4K pages)z9 EC, FICON Express 2, DS8000Assuming that the data set is cache resident
0
50
100
150
200
250
Th
rou
gh
pu
t (M
B/s
ec)
Sequential Prefetch Dynamic Prefetch
12
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
Measurements
2002/2003
DB2 Version 8, z990, FICON Express 1, ESS 800, 2 Gbit/sec link, z/OS 1.3
2005
DB2 Version 8, z9, FICON Express 1 and 2, DS8000, 2 Gbit/sec link, z/OS 1.6 (pre-MIDAWs) and z/OS 1.7 (MIDAWs)
All z9 measurements done using z9 Enterprise Class (EC)
2006
DB2 Version 9
FICON Express 4, DS8000, 4 Gbit/sec link, z/OS 1.7
13
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
Review of CKD Architecture
14
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
FICON EF Performance PenaltyDB2 Table Scans
0
50
100
Thro
ughp
ut
(MB/
sec)
Non-EF EF
FICON Exp. 2, DS8000
0
10
20
30
40
50
Thro
ughp
ut
(MB
/sec
)
Non-EF EF
FICON Exp. 1, ESS 800
15
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
Scatter Reads and Gather WritesChannel programs use real addresses
Must be able to scatter reads or gather writes to/from discontiguous real storage
CKD architecture has always had two methods
Indirect Address Words (IDAW)
• An IDAW used to be 4 bytes. (Hence, a “word”)
• Now IDAWs are usually 8 bytes to accommodate 64-bit addresses
Data Chaining
A MIDAW is a Modified Indirect Address Word
A MIDAW is 16 bytes, including a two-byte length field
16
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
Channel Programs
A Channel Program consists of a series of 8-byte CCWs (Channel Command Word)
Each CCW contains opcode, flags, length field, and address pointer
CCWs are executed sequentially unless the opcode is a TIC which acts like a branch or “go to” statement
17
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
CKD Channel Program Architecture
4K page on 4K boundary
4K page on 4K boundary
CCW = Channel Command Word 0 8 16 32 63 Real AddressLengthFlagsOpcode
IDAW List (Indirect Address Words)
0 63
Real Address
Real Address
Real Address
Real Address
Must end on 4K boundary
Must start on 4K boundary
The operation continues until the Length is satisfied (or there is no more data to read)
18
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
CCW Chaining 0 8 16 32 63 Real AddressLengthFlagsOpcode
Chain command (CC)Chain data (CD)
Channel Program
0 63
If CD is set, continue the current CCW operation to the next CCW.
The channel will fetch the next CCW if CC is set.
If neither CD nor CC are set, the channel program terminates.
InDirect Address Word List (IDAW)
19
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
EF suffix
Write Count Key Data using data chaining
count field
user’s data
1DDC 8 real address
CCWs
4096 real address
32 real address
DC
real address
real address
real address
Write an 8 byte count field followed by a 4128 byte data field.
20
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
Chaining prior to Shark and FICON
Command chaining was the only way to transition from one record to the next record in the same channel program.
i.e. To read/write two records required two CCWs
IDAWs were used for “scatter reads” and “scatter writes”, when the record required more than one 4K frame.
Data chaining was rarely used. There were two cases where data chaining was used:
Format writes, to chain the 8-byte Count field to the Data field
EF datasets, to chain the user’s data to the 32-byte suffix
Data chaining performed well on ESCON.
21
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
Track-level CCW op-codesSupported by Shark (and other compatible control units)
Purpose: To reduce the number of CCWs (but actually it didn’t because we didn’t have MIDAWs)
Read Track Data and Write Track Data
Alternative to traditional Read Data and Update Write op-codes
One CCW can stream data across records within a track
CKD count fields are not read or written
Traditional Write Count-Key-Data still used for most formatting
These op-codes were introduced in 2000 and implemented by Media Manager.
This was good for ESCON, but did not help FICON.
22
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
Data chaining with Shark and FICON
FICON data chaining hurts performance when the data chunks are small
i.e. Data chaining hurts the performance of EF datasets, as well as the performance of format writes
With ESCON, data chaining was not a performance problem, and EF I/O was as efficient as non-EF I/O
Since the overhead of data chaining was “per block”, large blocks were largely unaffected by data chaining
Data chaining causes higher channel utilization than IDAWs.
Especially so with 32 byte chunks
Command chaining is also inefficient. i.e. The best way to optimize FICON performance is to minimize the number of CCWs.
23
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
MIDAWs
24
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
What is the MIDAW Facility?Modified Indirect Data Address Word
Improves the performance of sequential I/Os using 4K datasets, especially with Extended Format
Eliminates the EF performance penalty and shrinks the small DB2 page size performance penalty
MIDAWs implemented by Media Manager, and supported by EXCPVR
EXCP does not support MIDAWs
Lower utilization of the FICON channel, link, and CU host adapter
25
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
MIDAW Requirements
Requires the System z9 processor and z/OS® 1.7 or 1.6 + APARs
APARs OA10984, OA13324, OA13384
Currently not supported by non-IBM storage vendor products
If you want to be allow a multi-volume data set to span a mixture of device types that span CUs where some support MIDAWs and some don’t, you need….
UA25186 (HDZ11J0) or UA25187 (HDZ11K0)
26
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
MIDAW Concepts0 40 48 127
reserved flags
64
count data addresses (64 bits)
Flags: Bit 40 - last MIDAW in listBit 41 - skip transfer to main storage (like Skip in CCW)Bit 42 - data transfer interruption control (like PCI in CCW)
CCWs with data chaining can direct arbitrary portions of a data block to separate buffer areas, that is, scatter-read or gather-write
27
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
Write Count Key Data using MIDAWs
1D 4136 real addressCCW
8
4096
32L
real address
real address
real address
The sum of all MDAW counts must equal the
CCW count and the “Last” flag must be set.
count field
user’s data
EF suffix
IDA
28
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
MIDAWs reduce the number of CCWsExample: Read or write a full 3390 track
Type of dataset,With or without MIDAWs
Number of CCWs per track
Read or Update
Format Write
4K, Non-EF, Pre-MIDAWs 12 24
4K, Non-EF, MIDAWs 1 12
4K, EF, Pre-MIDAWs 24 36
4K, EF, MIDAWs 1 12
16K, Non-EF, Pre-MIDAWs 3 6
16K, Non-EF, MIDAWs 1 3
16K, EF, Pre-MIDAWs 6 9
16K, EF, MIDAWs 1 3
29
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
Media manager channel program limitations
If DB2 provides too many buffers to MM, the I/O request is split into two I/Os, which can effect throughput by 10 to 15%
Prior to today, MIDAWs and CCWs in same 4K real frame.
Impact: 256K format writes of 4K pages
UA25366 (HDZ11J0) and UA25367 (HDZ11K0) move MIDAWs to a separate frame, which is still not enough for 512K
Sufficient for DB2 V8, but not always sufficient for DB2 V9
Impact: 512K reads and writes of 4K pages, and 512K format writes of 8K pages.
Future: Allow MIDAWs to use more than frame
30
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
MIDAWs improves the performance in two ways
Simplex effects (i.e. single stream):
This benefit is exclusive to EF datasets
Multiplexing effects (i.e. parallel I/O):
This benefit affects both EF and non-EF datasets
Single stream and multiplexing benefits are additive
EF benefits are much greater than non-EF benefits
There is no non-EF benefit unless multiple streams share the channel
31
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
Channel Efficiency
0
40
80
120
160
200
0 50 100 150
Channel Utilization (%)
Thro
ughp
ut (M
B/s
ec)
LessEfficient
MoreEfficient
Channel efficiency is the relationship of throughput to channel utilization
32
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
DB2 V8 Table Scans using 4K pages, EF DatasetsFICON Express 2, DS8000
Channel Efficiency
0
50
100
150
200
0 20 40 60 80 100
Channel Utilization
Th
rou
gh
pu
t (M
B/s
ec)
Pre-MIDAWs MIDAWs
I/O Response Time
0
1
2
3
4
5
0 1 2 3 4
Number of streams
I/O
Res
po
nse
Tim
e (m
s)
Pre-MIDAWs MIDAWs
33
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
DB2 V8 Table Scans using 4K pages, Non-EF DatasetsFICON Express 2, DS8000
Channel Efficiency
0
50
100
150
200
0 20 40 60 80 100
Channel Utilization
Th
rou
gh
pu
t (M
B/s
ec)
Pre-MIDAWs MIDAWs
I/O Response Time
00.5
11.5
22.5
33.5
0 1 2 3 4
Number of streams
I/O
Res
po
nse
Tim
e (m
s)
Pre-MIDAWs MIDAWs
34
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
Benefits of MIDAWs for DB2 table scan & 4k pagesFICON Express 2, DS8000
MIDAWs increases the throughput of EF datasets by 67%
MIDAWs eliminates the performance gap between EF and non-EF datasets 0
20
40
60
80
100
120
Th
rou
gh
pu
t (M
B/s
ec)
Non-EF EF
Type of data set
Pre-MIDAWs MIDAWsConfiguration:
MIDAW : z/OS 1.7
Pre-MIDAW: z/OS 1.6
DB2 for z/OS Version 8
4000 byte row size
System z9 109
FICON® Express2
2 Gbit/sec link
DS8000 control unit
*This performance data was measured in a controlled environment. The actual throughput or performance 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.
35
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
DB2 V8 Table Scans using 4K pages, EF DatasetsFICON Express 2 versus 4
Channel Efficiency
050
100150200250300
0 20 40 60 80 100
Channel Utilization
Th
rou
gh
pu
t (M
B/s
ec)
FICON Express 2
FICON Express 4
I/O Response Time
02468
1012
0 5 10 15
Number of streams
I/O
Res
po
nse
T
ime
(ms)
FICON Express 2
FICON Express 4
36
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
Small Block Performance Penalty
When prefetching 128K per I/O, 16K page size performs worse than 4K page size, even though the number of bytes per track is the same, and the I/O is not disk bound
Why? Because the channel program for 4K required more CCWs than for 16K.
If the number of CCWs were the same, the channel efficiency would be the same. In fact, with MIDAWs and non-EF datasets, the channel programs look identical. Therefore, the channel efficiency is the same. With EF, they are almost the same.
Yet, even if the channel efficiency is the same, the control unit may or may not perform better with larger blocks
37
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
Benefits of MIDAWs: 4K vs 16K DB2 page sizeFICON Express 2, DS800
MIDAWs shrinks the performance gap between 4K and 16K CIs
0
0.5
1
1.5
2
I/O
Res
po
nse
T
ime
(ms)
4K page 16K page
EF Data SetPrefetch I/Os
Pre-MIDAWs MIDAWs
Configuration:
MIDAW : z/OS 1.7
Pre-MIDAW: z/OS 1.6
DB2 for z/OS Version 8
4000 byte row size
System z9 109
FICON® Express2
2 Gbit/sec link
DS8000 control unit
*This performance data was measured in a controlled environment. The actual throughput or performance 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.
38
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
When is MIDAWs not beneficial?
MIDAWs do not help large block sizes
MIDAWs do not help if the FICON channel utilization is low
If you FICON utilizations are always low, then maybe you could get better performance if your batch workload used more parallelism.
MIDAWs mitigates the effect of using parallelism (for DB2 queries and online utilities) on OLTP
39
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
New hardware
40
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
z9 I/O highlights
Improved System I/O capabilities
40% increase in the maximum number of FICON Express2 channels
• Up to 336 on z9-109 vs. 240 on z990
80% increase in peak I/O bandwidth per book
Up to 30% increase in effective Sap capacity*
• based on OLTP-T workload measurements (55K io/sec per sap @70% sap utilization)
Enterprise Class (EC) and Business Class (BC)
BC machines have lower MIPS than EC (except for zIIPs)
MIDAWs
Improved channel efficiency *This performance data was measured in a controlled environment on a z9-109 running the LSPR OLTP-T workload under z/OS 1.6. The actual throughput or performance 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.
41
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
New hardware components in 2005
FICON Express2 channels (based on 2gbit links)
DS8000 storage subsystem
System z9
New hardware components in 2006FICON Express4 channels (4gbit links)
DB2 Version 9 to become available in 2007
42
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
Introducing FICON Express 2
0
50100
150200
250300
MB
/sec
FICONExpress 1
FICONExpress 2
Bi-directional Throughput
Capacity
Available in 1Q/05
FICON Express 2 increased the channel bandwidth by 58%
FICON Express 2 exceeds the uni-directional bandwidth of a 2Gbit/sec link
Given a 2Gbit channel, a mixture of reads and writes is necessary to exceed 200 MB/sec
These measurements were made using DB2. Other sources of performance data are not based on DB2 workloads.
43
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
Introducing FICON Express 4
050
100
150200250300
MB
/sec
FICONExpress 2
FICONExpress 4
Uni-directional Throughput
Capacity
Availability?
Utilizing a 4gbit/sec link, FICON Express 4 increases the uni-direction channel bandwidth by 50%
These measurements were made using DB2. Other sources of performance data are not based on DB2 workloads.
44
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
DS8000 highlightsNew POWER5 SMP processors Up to 256 GB cache FC-AL to link DA pairs to DDMsScalability up to 192 TBUp to 128 2 Gb/sec FICON host ports (max 16 ports recommended for an
individual host) Performance
A DS8100 has the performance capacity of two ESS 800s with half the footprint
Combined with FICON Express 2, the performance capacity for DB2 queries and utilities can be as much as three times higher than the ESS 800
The DS8300 is recommended for LPAR configuration, or for workloads making extensive use of advanced copy services (i.e. PPRC and Flashcopy)
45
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
DB2 page sizes
DB2 index compression
Dasd space management
46
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
DB2 Page Size …..
DB2 supports page sizes of 4K, 8K, 16K and 32K Large pages accommodate larger rows
A row cannot space pages and a page is limited to 255 rows
Larger objects are stored in LOB table spaces, and a row may contain a row id that identifies a large object in a corresponding LOB table space.
Small pages are best for OLTP to maximize buffer hit ratio
Beginning with DB2 V8, the CI size can be equal to page size A large CI size may achieves better sequential performance, depending
on the I/O configuration
For data integrity reasons, DB2 does not allow striping unless the CI Size matches the page size
47
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
….. DB2 Page Size
Starting in V9, DB2 supports up to 32K page size for indexes
A small page size will maximize the buffer hit ratio
A large page size will reduce the number of index levels
48
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
DB2 Table Scan using an EF Dataset
0102030405060
Th
rou
gh
pu
t
(MB
/se
c)
4K 16K
Page Size
ESS 800, Pre-MIDAWs, FE1
Configuration:
z/OS 1.3
DB2 for z/OS V8
4000 byte row size
System z900
FICON® Express1
2 Gbit/sec link
ESS 800 control unit
Configuration:
z/OS 1.7
DB2 for z/OS V8
4000 byte row size
System z9-109
FICON® Express2
2 Gbit/sec link
DS8000 control unit
020406080
100120
Th
rou
gh
pu
t (M
B/S
ec)
4K 16K
Page Size
DS8000, MIDAWs, FE2
Large page sizes were good with FICON Express 1, but less beneficial with FICON Express 2
49
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
DB2 Index Compression…..
Index compression is new to V9
Page level compression
Unlike data row compression:
Buffers contain expanded pages
Pages are decompressed when read from dasd
Prefetch performs the decompression asynchronously
A buffer hit does not need to decompress
Pages are compressed by the deferred write engine
Like data row compression:
An I/O bound scan will run faster
DSN1COMP utility can be used to predict space savings
50
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
Index Compression Example: Suppose an index could compress 3-to-1
Compressed 4K
Decompressed 16K buffer
12K used
4K unused
Compressed 4K
Decompressed 8K buffer
8K Bufferpool50% dasd space reductionNo increase in virtual storage cost
16K Bufferpool67% dasd space reduction33% increase in virtual storage cost
51
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
…..DB2 Index Compression
The CI Size on dasd of a compressed index is always 4K
A 4K expands into a 8K or 16K buffer, which is the DBA’s choice. This choice determines the maximum compression ratio.
Compression of key prefix and Rid Lists
A Rid List decribes all of the rows for a particular index key
An index with a high level of non-uniqueness, producing long Rid Lists, achieves about 1.4-to-1 compression
Compression of unique keys depends on prefix commonality
52
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
Index Compression: Performance
CPU cost is mostly inconsequential. Most of the cost is asynchronous, the exception being a synchronous read. The worst case is an index with a poor buffer hit ratio.
Example: Suppose the index would compress 3-to-1. You have three options…..
1. Use 8K buffer pool. Save 50% of dasd. No change in buffer hit ratio or real storage usage.
2. Use 16K buffer pool and increase the buffer pool size by 33%. Save 67% of dasd, increase real storage usage by 33%.
3. Use 16K buffer pool, with no change in buffer pool size. Save 67% of dasd, no change in real storage used, decrease in buffer hit ratio, with a corresponding increase in synchronous CPU time.
53
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
>4K Page Size for Indexes
V9 supports 4K, 8K, 16K and 32K page sizes for indexes
A large page size is very good for reducing the frequency of CI splits, which is costly in data sharing environment.
The downside: As with large pages for table spaces, the buffer hit ratio could degrade.
54
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
Non-padded indexes
Non-padded indexes were introduced in DB2 V8 Useful when an index contains one or more VARCHAR columns
Facilitates index only access
Saves DASD space
The CPU cost is significant. It grows as a function of the number of VARCHAR columns.
Index compression and non-padded indexes are complementary
55
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
DB2-managed “sliding scale” Extent sizes when the secondary space quantity is not specified
0
50
100
150
200
250
1 21 41 61
Extent numberE
xte
nt
siz
e (
cy
l)
DSSIZE=4G, PRIQTY=200 Cyl
0
20
40
60
80
100
120
1 21 41 61 81 101
Extent number
Ex
ten
t s
ize
(c
yl)
DSSIZE=4G, Default PRIQTY
Extent size = Extent number Extent size = max (extent number, 10% of PRIQTY)
56
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
Sliding scale when DSSIZE = 64GB
0
100
200
300
400
500
6001
32
61
92
12
2
15
3
18
3
21
4
24
5
Extent number
Ext
ent
size
(cy
l)
DSSIZE=64G, Default PRIQTY
More details can be found in the DB2V8 SQL Reference Guide
57
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
SMS Classes
DB2 V9 supports the specification of SMS constructs on the DB2 Stogroup definition
58
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
Long term page fixing
DB2 V8 moved all buffers into 64-bit storage, which increases the cost of page fix/free (unless you were using data space buffer pools)
To compensate, DB2 introduced the PGFIX(YES|NO) long term page fix option. It is a buffer pool option.
Long term page fixing can save 6% of the CPU for an I/O intensive OLTP DB2 workload.
Actual savings may be higher as the amount of real storage increases
Some throughput increase associated with DB2 queries and utilities
PGFIX can be altered dynamically, but the buffer pool must be flushed to alter it
All measurements shown in this presentation used PGFIX(YES)
59
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
The anatomy of a DB2 prefetch operation
60
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
DB2 Index Scan z9 EC, FICON Express 2, DS8000 Assuming that the index is cache resident
0
50
100
150
200
250
Th
rou
gh
pu
t (M
B/s
ec)
DB2 V8 DB2 V9
61
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
DB2 Dynamic Prefetch vs Sequential Prefetch
At Bind time, the DB2 optimizer may select Sequential Prefetch
At execution time, Data/Index Manager may choose Dynamic Prefetch
Sequential Prefetch schedule one SRB at a time (except at the start of a query), but Dynamic Prefetch schedules two concurrent SRBs
This can double the throughput if the data is cache resident, if the channels are not busy, and if the scan has low CPU time
Table space scans always use sequential prefetch.
In Version 8, for indexes the DB2 optimizer made the choice based on index statistics
In Version 9, the optimizer always relies on dynamic prefetch for everything other than table space scans.
62
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
DB2 Prefetch
DB2 employs prefetch for table scans, index scans, and LOB reads
Prefetch runs under an “asynchronous SRB” while the user’s thread evaluates the predicate.
While the predicate evaluation and prefetch engines are being done, the control unit may be “prestaging” data into its cache
Suspension due to prefetch delay is reported in DB2PE/Omegamon as “Other Read I/O”
“Other Read I/O” is characteristic of queries that have low CPU time:
LOBs and large rows
Multi-row fetch (e.g. DSNTEP4 Selects)
Simple SQL predicates
Copy and Recover utilities
63
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
Prefetch Quantity
When migrating from V8 to V9, you should expect fewer I/Os and higher I/O response times
Prefetch quantity for compressed index is based on uncompressed page size. This may result in more prefetch I/Os than an uncompressed index. For example, a scan of a compressed 16K index with a 3-to-1 compression ratio will do 33% more I/Os than its equivalent uncompressed index.
Prefetch type DB2 V8 DB2 V9 BPSIZE *
VPSEQT < 160MB
DB2 V9 BPSIZE *
VPSEQT >= 160MB
DB2 V9 BPSIZE *
VPSEQT >= 320MB
Dynamic 128K 128K 128K 128K
Sequential 128K 128K 256K 256K
Utility 256K 256K 256K 512K
64
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
Prestaging effects for a table scan
Suppose that the control unit prestage rate is 150 MB/sec and the channel transfer rate is 100 MB/sec
If the CPU predicate evaluation is slower than 150 MB/sec, there will be no “Other Read I/O” delays, and no I/O disconnect time, and the elapsed time will equal the CPU time of the query
As the channel speeds up, disconnect time increases
A larger prefetch quantity causes higher connect time and higher disconnect time
Disks with 15K RPMs will achieve less disconnect time than 10K RPMs and increase throughput if the data is not cache resident
Suppose the data is cache resident
No I/O disconnect time
If the CPU predicate evaluation is slower than 100 MB/sec, there will be no “Other Read I/O” delays, and the elapsed time will be equal to the CPU time of the query
65
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
DB2 Sequential Prefetch Throughputz9 EC, FICON Express 2, DS8000
0
20
40
60
80
100
120
140
160
Th
rou
gh
pu
t (M
B/s
ec)
From cache From 10K RPM 300GB disk
64x4K
8x32K
32x4K
4x32K
66
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
0
0.5
1
1.5
I/O
Resp
on
se
Tim
e (
ms)
FICON Express 2 FICON Express 4 FICON Express 4
DB2 Sequential Prefetch (4K pages)Using MIDAWs and the DS8000 with 4 gbit/sec FICONCache resident table
Performance is based on measurements and projections using standard IBM benchmarks in a controlled environment. The actual throughput or performance 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 or performance improvements equivalent to the numbers stated here.
DB2 Version 8: 32x4K prefetch
DB2 Version 9: 64x4K prefetch
67
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
DB2 V9 vs V8 Table Scan using FICON Express 4Cache resident data
Configuration:
z/OS 1.7
4000 byte row size
System z9 EC
FICON® Express4
4 Gbit/sec link
DS8000 control unit
Using V9 with FICON Express 4, the channel throughput is 193 MB/sec, but the effective throughput is only 139 MB/sec.
0
25
50
75
100
125
150
Th
rou
gh
pu
t(M
B/s
ec)
V8 V9
68
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
Parallel queries
69
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
DB2 Prefetch Throughput Capacity16 partitions, 4K CI Size2002 measurementsESS 800, 8 FEx1 channels, 2 Gbit/sec link, z990 host
The model 800 bus is capable of 550 MB/sec with half-track blocks, but DB2 table spaces are more limited
0
50
100
150
200
250
300
350
Th
rou
gh
pu
t (M
B/s
ec)
Non-EF EF
Type of dataset
Performance is based on measurements in a controlled environment. The actual throughput or performance 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 or performance improvements equivalent to the numbers stated here.
70
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
DS8000: DB2 Prefetch Throughput Capacity as a function of the number of FICON Express 2 Channels 64 degrees of parallelism
0
400
800
1200
1600
Th
rou
gh
pu
t (M
B/s
ec)
2 4 8 16
Number of FEx2 Channels
The DS8000 bus is capable of 1800 MB/sec with 16 FICON, but DB2 table spaces are more limited.
16 channels achieves 35% higher throughput than 8 channels, but this is probably not true with FICON Express 4
Each LCU is limited to 8 channels, but the LCUs may be divided into two channel sets
*This performance data was measured in a controlled environment. The actual throughput or performance 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.
Configuration:
Pre-MIDAWs: z/OS 1.6
System z9 109
DB2 for z/OS Version 8
4000 byte row size
System z9 EC
FICON® Express2
2 Gbit/sec link
DS8000 control unit
71
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
DS8000: DB2 Prefetch Throughput Capacity 8 FICON Express 2 Channels, 64 degrees of parallelism
0
200
400
600
800
1000
1200
Th
rou
gh
pu
t (M
B/s
ec)
Pre-MIDAWs MIDAWs
EF 4K Non-EF 4K Non-EF 16K
MIDAWs eliminates the EF performance penalty and shrinks the small CI penalty
Performance is still sensitive to the number of bytes per track
*This performance data was measured in a controlled environment. The actual throughput or performance 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.
Configuration:
MIDAW : z/OS 1.7
Pre-MIDAW: z/OS 1.6
DB2 for z/OS Version 8
4000 byte row size
System z9 EC
FICON® Express2
2 Gbit/sec link
DS8000 control unit
72
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
Review of Extended Format Data Sets
DB2 Striping Considerations
73
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
What is an Extended Format Data Set?
Both VSAM and QSAM/BSAM data sets may be EF
Supported only on SMS-managed volumes
DSNTYPE=Extended
Uses a fixed length records (even if RECFM=VB)
A 32-byte suffix is appended to every physical record
The suffix contains a length field to logically describe “short blocks”
The suffix is also used to increase reliability
EF datasets may only be accessed via Media Manager
Media Manager manages the suffix so that the suffix is completely transparent to the caller
74
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
What is Media Manager?
Why do we have it?
To limit dependencies on changes to the I/O architecture
Implements the latest and greatest I/O channel program architecture, such as track level CCWs and MIDAWs
What is it?
It is an I/O driver that knows how to map a relative byte address to a CCHHR, using a fixed block architecture to translate the address. It also supports striping.
What isn’t it?
It is not a buffer manager. The caller supplies the buffers.
Who can use it?
Any supervisor state program accessing a fixed block data set.
75
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
Why use Extended Format?
Benefit from the performance advantages of Media Manager
VSAM Extended Addressability (>4GB data sets)
Applies to linear datasets, but there are dependencies on the caller of Media Manager (e.g. DB2) to support EA
Striping
BSAM, QSAM and KSDS Compression
QSAM
Instead of default BUFNO=5. the default BUFNO is 2 times the number of tracks times the number of stripes
Double Buffering (to facilitate CPU and I/O overlap)
76
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
DB2 striping restrictions
With DB2 Version 9, all DB2 data sets can be striped provided that the CI size matches the page size. (Version 8 was the first version to support large CIs.)
Version 8 restriction: Archive logs could not be striped unless you have the Archive Log Accelerator
77
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
DB2 striping versus partitioning
CURRENT DEGREE = ‘ANY’ causes a query to use parallelism if the table space is partitioned.
Partition parallelism uses more buffers than does striping
Partitioning is always a better method of achieving parallelism than striping, unless you are storage constrained, because…..
Partitioning uses more storage
Partitioning is unaffected by I/O variance
Partitioning uses CPU parallelism
Striping be combined with partitioning to improve the performance of partition-level operations (e.g. Backup or reorg one partition)
Striping can also be used for non-partitioned indexes
78
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
DB2 Table Space Striping versus DB2 Partitioning
0
50
100
150
200
Th
rou
gh
pu
t(M
B/s
ec)
1 2
Number of stripes
Configuration:
z/OS 1.7
System z9 EC
DB2 for z/OS Version 8
4000 byte row size
FICON® Express2
2 Gbit/sec link
DS8000 control unit
*This performance data was measured in a controlled environment. The actual throughput or performance 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.
0
50
100
150
200
Thro
ughp
ut
(MB
/sec
)1 2
Number of partitions
79
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
When is striping advantageous? Striping is a good strategy for taking high I/O connect time and
spreading it across multiple channels.
If some control units are busier than others, then spreading stripes across control units is good. If this helps increase the cache hit ratio (for random access), disconnect time will go down. If this helps reduce control unit bus contention (for sequential access) , connect time will go down.
Striping isn’t necessary to avoid IOSQ time, because PAV avoids IOS queue time.
Striping does not help if your channels are saturated.
As your channel technology improves, you should reduce the number of stripes or stop using striping.
80
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
Recommended number of stripes, examples:
Configuration
Connect time(ms) Recommended
number of stripes32x4K prefetch
64x4K prefetch
ESS 800FICON Express 1
2.7 5.0 4 to 8
DS8000FICON Express 2
0.9 1.5 2 to 4
DS8000FICON Express 4
0.7 1.1 1 to 2
Rule of thumb: Try to reduce connect time to 1.0ms. For selected data sets, you may try to reduce to 0.5ms if this won’t cause high channel utilizations.
81
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
Why can too many stripes be bad?
If pend time is high relative to connect time, then too much of your channel time is consumed by I/O overhead
High channel utilizations or CU utilizations cause I/O response time variance which hurts striping
An imbalance in the number of blocks per stripe causes I/O response time variance which limits parallelism.
Example: CI Size is 32K. Prefetch quantity is 4x32K. A 5th stripes is useless. If you use 3 stripes, two stripes will be idle while the system is reading the 4th CI. This limits parallelism.
If the CI size is large, choose the number of stripes for a DB2 data set to be a power of 2.
Adding unnecessary stripes for DB2 data sets causes extra I/Os, and adding unnecessary stripes for non-DB2 data sets wastes real storage.
82
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
Logging, preformatting, Reorg, etc.
83
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
Striping the DB2 log datasets
Two stripes increases the maximum throughput of DB2 inserts by 30% to 50%, but on faster devices most insert streams are CPU bound.
Use the same connect time rule of thumb for logs and table spaces. For logs, consider mass inserts and Fast Log Apply (e.g online Reorg of a very large table space).
Version 9 increases the effectiveness of striping during Fast Log Apply, because the number of input I/O buffers was increased from 15 to 120.
Version 9 permits the archive log datasets on Dasd to be striped too. This prevents log offloading from becoming a bottleneck when the active logs are striped. Version 9 also permits the archive logs to use BSAM compression.
84
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
DB2 Inserts: Log throughput Capacity
0
10
20
30
40
Th
rou
gh
pu
t (M
B/s
ec)
Non-EF 2 stripes
Type of Log Dataset
ESS 800, Pre-MIDAWs, FE1
Configuration:
z/OS 1.3
DB2 for z/OS V8
4000 byte row size
System z900
FICON® Express1
2 Gbit/sec link
ESS 800 control unit
Configuration:
z/OS 1.7
DB2 for z/OS V8
4000 byte row size
System z9-109
FICON® Express2
2 Gbit/sec link
DS8000 control unit
0
20
40
60
80
100
120
Thro
ughp
ut
(MB
/Sec
)Non-EF 2 stripes
Type of Log Dataset
DS8000, MIDAWs, FE2
85
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
Preformatting (versus formatting)
Preformatting accomplishes two things:
Allow subsequent update writes to be done in parallel without regard to track boundaries.
Update HI-U-RBA (High Used RBA) in the catalog
OLTP uses preformatting, because datasets don’t get closed
DB2 Utilities can use normal format writes, because the data doesn’t need to be secure until the dataset is closed. Close will update the catalog.
Preformatting is asynchronous, except after allocating a new extent
The “preformat quantity” is the frequency that we update the catalog
If the quantity is too little, we update the catalog too often
If the quantity is too much, then synchronous preformatting will cause user’s to wait.
86
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
Mass Inserts
Mass inserts are affected by preformatting performance, especially with LOBs. LOBs don’t require logging and are not CPU bound.
DB2 Version 9 increased the preformat quantity from 2 cylinders to 16
This can yield >20% improvement in elapsed time for LOBs.
If the datasets are striped, the preformat quantity is in terms of Control Areas
E.g. The control area with 2 stripes is 8 tracks per stripes
The benefit of updating the catalog less often is higher for striped datasets and for faster devices
Cross Loader utility (used to copy one table space to another) also benefits
87
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
DB2 Load Utilityz9 EC, FICON Express 2, DS8000Cache resident input file
0
20
40
60
80
100
120
140
Th
rou
gh
pu
t (M
B/s
ec)
16x16K 32x16K
Table Space: #pages/IO x page size
I/O reduction produces up to 10% increase in throughput. Higher benefits expected with striping and with FICON Express 4.
88
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
Online LOB Reorg
Version 8 Reorg LOB
Online Reorg not supported
Used only to defragment each LOB, but could not reclaim unused space
Version 9 Reorg LOB
Online Reorg copies LOBs to a shadow dataset, just like Reorg for tables and indexes
Unused space is reclaimed
89
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
Online Reorg
Version 8 PTF: PK25742
At one shop this PTF reduced Online Reorg time from 2 hours to 45 minutes due to a significant reduction in the Sortbuild phase
90
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
Fast Log Apply
Fast Log Apply is the last phase of both Online Reorg and Recover
FLA consists of log reads and database updates.
DB2 Version 9 improves log read throughput, especially with the faster channels.
When applying a small percentage of log records, FLA is gated by log reads. When applying a large percentage of log records, FLA is gated by database I/Os. Ergo, V9 improves the case where the percentage is small.
DB2 Version 9 also permits the archive logs on dasd to be striped and compressed, yielding yet further FLA performance improvements as well as storage space savings.
91
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
Backup, Restore and Recover
92
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
Backup and Recovery
Version 8 introduced Backup/Restore utilities Uses SMS Copypools, new to z/OS 1.5, and requires HSM
BACKUP creates a “Point In Time” backup of databases and logs
• Uses volume-level Flashcopy
RESTORE restores all volumes using volume-level Flashcopy
Version 9 Recover utility will use dataset level Flashcopy using the volume-level backups created by BACKUP utility Requires z/OS 1.8
93
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
Online Check
DB2 Version 8
Check Index uses dataset level Flashcopy
DB2 Version 9
Check Data and LOB use dataset level Flashcopy
94
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
DB2 Utilities using eight FICON Express2 channelChannel utilizations ~10%
020406080
100120
Thro
ughp
ut
(MB
/sec
)
Non-EF EF
Type of data set
DB2 Copy Utility
Pre-MIDAWs MIDAWs
Performance is based on measurements in a controlled environment. The actual throughput or performance 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 or performance improvements equivalent to the numbers stated here.
020406080
100120
Th
rou
gh
pu
t (M
B/s
ec)
Non-EF EF
Type of data set
DB2 Recover Utility *
Pre-MIDAWs MIDAWs
Configuration:
MIDAW : z/OS 1.7
Pre-MIDAW: z/OS 1.6
DB2 for z/OS Version 8
4000 byte row size
4K CI Size
System z9 EC
FICON® Express2
2 Gbit/sec link
DS8000 control unit
* Restore phase only
Recover impacted by Media Manager limitation which has since been lifted by apar
95
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
Conclusion
96
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
Creating a balanced configuration
A balanced I/O configuration is one where all components perform about equally.
FICON Express 1 and the ESS 800 are a good match
FICON Express 1 and the DS8000 are a good match
FICON Express 2 and the ESS 800 is a poor match
However, FICON Express 2 channels can be used to support several ESS 800s
FICON Express 2 or 4 and the DS8000 are an excellent match
z9-109 FEx2 FICON Link DS8000HA
97
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
Using hardware to address I/O response time components
High pend time: More channels or faster channels, MIDAWs
High connect time:
Faster FICON links/directors, more/faster channels and CU host adapters, more CU bus capacity
MIDAWs
Striping
High disconnect time
More real storage, more control unit cache, more disks, more control units
Faster disk RPMs (tradeoff between performance and capacity)
High IOSQ time is always a secondary symptom. It is better to attack the primary symptom first, but to reduce IOSQ time:
More UCBs, new PAV
98
IBM Software Group | DB2 Information Management Software
DB2 I/O Performance: NCCMG
DB2 Logging Issues (unique to synchronous writes)
If you are using remote copy, assume you have a network problem, unless you can see that you have high log connect times
If you aren’t using remote copy, and you have any disconnect time on your logs, then you need more NVS storage, or faster RPMs
If you have high connect time:
Upgrade your hardware as you would for table spaces
Use striping or add more stripes