VSP MAINFRAME DYNAMIC TIERING PERFORMANCE CONSIDERATIONS STEVE RICE, MASTER PERFORMANCE CONSULTANT, MAINFRAME, HITACHI DATA SYSTEMS
SEPTEMBER 12, 2012
WEBTECH EDUCATIONAL SERIES
VSP Mainframe Dynamic Tiering Performance Considerations
A key reason for using dynamic tiering for mainframe storage is
performance. This session will focus on dynamic tiering in mainframe
environments and how to configure and control tiering. The session
ends with a detailed discussion of performance considerations when
using Hitachi Dynamic Tiering.
By attending this webcast, you will
• Understand Hitachi Dynamic Tiering and the options for configuring
and controlling tiering
• Understand the performance considerations and the type of
performance improvements you might experience when you
implement Hitachi Dynamic Tiering
UPCOMING WEBTECHS
Mainframe Series
Mainframe Replication, Sept 19, 9 a.m. PT, 12 p.m. ET
Why Networked FICON Storage Is Better than Direct-attached
Storage, Oct 3, 9 a.m. PT, 12 p.m. ET
Other
Storage Analytics, Sept 20, 9 a.m. PT, 12 p.m. ET
Maximize Availability and Uptime by Clustering your Physical
Datacenters within Metro Distances, Oct 24, 9 a.m. PT, 12 p.m. ET
Check www.hds.com/webtech for
Links to the recording, the presentation and Q&A (available next week)
Schedule and registration for upcoming WebTech sessions
AGENDA
Brief description of tiering
‒ MK-90RD7021 Hitachi Virtual Storage Platform Provisioning Guide for Mainframe
Review RAID terminology
Hitachi Dynamic Provisioning (HDP) terminology and methodology
Discuss Hitachi Dynamic Tiering (HDT) dynamic parameter
Experiments 1 and 2
First steps with HDT and hierarchical storage management
LET TIERING LEARN YOUR WORKLOAD
HDD 3
Continues
HDD 2
Continues
HDD 1
Continues
HDD 0
Continues
RAID TERMINOLOGY: RAID 5 – 3D + 1P LDEV
Parity
Next
8
Tracks
Next
8
Tracks
Track
0-7
Next
8
Tracks
Parity
Next
8
Tracks
Track
8-15
Next
8
Tracks
Next
8
Tracks
Parity
Track
16-23
Next
8
Tracks
Next
8
Tracks
Next
8
Tracks
Parity
RAID Chunk 0
of RAID Stripe 0
RAID Chunk 1
of RAID Stripe 0
RAID Chunk 2
of RAID Stripe 0
RAID Chunk 3
of RAID Stripe 0
RAID Stripe 0
RAID Stripe 1
RAID Stripe 2
RAID Stripe 3
RAID Stripes
Continue
PHYSICAL TO LOGICAL: HDD TO 3390-X HOST-ADDRESSABLE DEVICE LET HDP SHARE THE LOAD
Hard Disk Drive
(HDD)
Parity Group
(PG)
3390-3/9/27/54/A(EAV)
Logical Device
(LDEV)
3390-V
MF-HDP Pool Volume
(Pool-VOL)
3390 –A(EAV)
HDP Volume
(DP-VOL)
3390 –A(EAV) Track Space-Efficient Volume
(TSE-VOL)
Physical Layer
HDS Storage
Subsystem
Abstraction Layer
Host Abstraction Layer or HCD / IOCDS Addressable
HDP PAGE ALLOCATION: 38MB (672 TRACKS)
1 3 2
5
Pool-VOL2
3390-V
(Parity Group 2)
Pool-VOL3
3390-V
(Parity Group 3)
Host
3 2
DP-VOL 3390-A EAV
R/W
4
4
Pool-VOL4
3390-V
(Parity Group 4)
Pool-VOL1
3390-V
(Parity Group 1)
1 5
HDP Pool 7
HDP wide stripes across parity groups
Current implementation of 3390-A(EAV) can have from 1 to 262,668 cylinders
7.5K RPM
SAS or SATA
62%
10K RPM
SAS
36%
HDT PYRAMID: ENGINEERING RECOMMENDATIONS
HDT pyramid within an HDP pool
SSD
2%
LET TIERING LEARN YOUR WORKLOAD
Lo
MULTIPLE TIERS WITHIN A SINGLE HDP POOL
Tier 1 - SSD
Tier 2 – SAS 10K or SAS 15K
Tier 3 – SAS 7.5K or External
A SINGLE HDP POOL
EVERY HDP POOL HAS AT LEAST 1 TIER
DP volumes
live within the
HDP pool
DP volume pages
live within 1 or
more tiers
DEFAULT TIER BUFFER SPACE
SAVE A LITTLE SPACE FOR PAGE RELOCATION
Hard Disk Type Buffer area for
Tier Relocation
Buffer Area for New
Page Assignment
Total
SSD 2% 0% 2%
Non-SSD
- SAS 15K HDD
- SAS 10K HDD
- SAS 7.5K HDD
- External
2% 8% 10%
Tier buffer space set at the HDP pool level
HDT CYCLE
Cycle time set at the HDP pool level
Manual mode
‒ User can start and stop performance monitoring using any interval up to 7 days
Automatic mode
‒ Continuous monitoring followed by relocation cycles
‒ Monitor interval from 30 minutes to 1, 2, 4, 8 or 24 hours (default)
LET TIERING LEARN YOUR WORKLOAD
Monitor 1 Calc
1
Relocate
1
Monitor 2 Calc
2
Relocate
2
Monitor 3 Calc
3
Relocate
3
Monitor 4 Calc
4
Relocate
4 TIME
T0 T1 T2 T3 T4
Tier 2
Tier 1
HDT MONITORING MODES
LET TIERING LEARN YOUR WORKLOAD
Monitoring modes set at the
HDP pool level
Period
‒ The value used in the calculation cycle is the actual IO load on DP volume page from previous monitoring cycle
Continuous
‒ The value used in the calculation cycle is the weighted average of multiple previous monitoring cycles for DP volume page
‒ Reduces page trashing
‒ May slow migration to upper tiers
0
10
20
30
40
50
60
70
80
90
100
110
120
1,005 1,006 1,007 1,008 1,009 1,010
DP
-Vo
lum
e P
age
IOP
H
Tiering Cycle Number
Period vs. Continuous
Period Mode Continuous Mode
RELATIONSHIP BETWEEN NUMBER OF TIERS AND TIERING POLICY
DP VOLUME TIERING POLICIES
Tiering Number of Tiers in Pool Note
Policy 1 Tier 2 Tier 3 Tiers
All All tiers All tiers All tiers Default value; data is
stored in all tiers.
Level 1 All tiers* Tier 1 Tier 1 Data is always stored in
the highest-speed tier
Level 2 All tiers* All tiers* Tier 1 and
Tier 2
Level 3 All tiers* All tiers* Tier 2
Level 4 All tiers* All tiers* Tier 2 and
Tier 3
Level 5 All tiers* Tier 2 Tier 3 Data is always stored in
the lowest-speed tier
* Data is stored in all tiers as in the case of All specified in the tiering policy
NEW PAGE ASSIGNMENT POLICY
TWO TIER TABLE FROM MANUAL
Tiering Level Policy Description
High The new page is assigned from the higher tier of tiers set in the tiering policy
Middle The new page is assigned from the middle tier of tiers set in the tiering policy
Low The new page is assigned from the lower tier of tiers set in the tiering policy
Tiering Level
Policy
When
Specifying
High
When
Specifying
Middle
When
Specifying
Low
Notes
All From Tier 1 to
2
From Tier 1 to
2
From Tier 2 to
1
In the Low setting, tier 2 is given
priority over tier 1
Level 1 From Tier 1 to
2
From Tier 1 to
2
From Tier 1 to
2
Every assignment sequence is the
same as when All is specified as the
tiering level
Level 2
Level 3
Level 4
From Tier 1 to 2 From Tier 1 to 2 From Tier 2 to 1 Every assignment sequence is the same
as when All is specified as the tiering
level
Level 5 From Tier 2 to
1
From Tier 2 to
1
From Tier 2 to
1
Assignment sequences when
High, Middle, and Low are the same
RELOCATION PRIORITY
Use the relocation priority function to set the selection
priority of a DP-VOL when performing relocation – a
prioritized DP-VOL can be relocated earlier during a
relocation cycle
For most effectiveness, use sparingly
“Level 1”?
‒ (NO)
“Level 5”?
‒ (NO)
“ALL”
‒ (Yes, sparingly)
DP VOLUME RELOCATION PRIORITY
HDT EXAMPLE 1
This quick storyboard shows preliminary results from HDT testing
‒ It attempts to show how HDT learns your workload
Scenario: Customer reluctant to upgrade from 300GB to 600GB HDD
Same capacity of HDD (not Including SSD)
‒ (128) 300GB SAS
‒ (64) 600GB SAS + (8) 400GB SSD
IMPORTANT NOTE: SSD drives are added to the pool after all data sets are created
BASIC CONFIGURATION
Config.
Name
RAID
Type
LCU DP-
VOL
per
Pool
PAIO
Data
set
Base /
Alias
Dev.
Num
.
Desc.
HDT3HF RAID-
6(6D+2P)
00 - 03 256 1024 64/192 70xx (128) 300GB
SAS HDP pool
HDT6HF RAID-
6(6D+2P)
08 – 0B 256 1024 64/192 72xx (64) 600GB
SAS HDP pool
HDT6HF
Run 1
through
Run 4
RAID-
6(6D+2P)
08 – 0B 256 1024 64/192 72xx HDT pool
(8) 400GB
SSD
(64) 600GB
SAS
300GB AND 600GB HDP BASELINES HAVE BEEN RUN (NO SSD DRIVES)
HDP (128)
300GB HDD
HDP (64)
600GB HDD
FIRST RUN: 600GB TIER 2 + SSD TIER 1 – 0 MINUTES – NO LEARNING
LET TIERING LEARN YOUR WORKLOAD
HDP (128)
300GB HDD
HDP (64)
600GB HDD
Tiering
(64) 600GB:
No learning,
Same as HDP
SECOND RUN: 600GB TIER 2 + SSD TIER 1 – 30 MINUTES OF REST AFTER RUN 1
LET TIERING LEARN YOUR WORKLOAD
HDP (128)
300GB HDD
HDP (64)
600GB HDD
Tiering – Run 1:
No learning,
Same as HDP
Tiering – Run 2:
After 30 minutes
of migration
FOURTH RUN: 600GB TIER 2 + SSD TIER 1 – 30 MINUTES OF REST AFTER RUN 3
LET TIERING LEARN YOUR WORKLOAD
HDP (128)
300GB HDD
HDP (64)
600GB HDD
Tiering – Run 1:
No learning,
Same as HDP
Tiering – Run 2:
After 30 minutes
of migration
Tiering – Run 3:
After 30 minutes
of migration
Tiering – Run 4:
After 30 minutes
of migration
SUMMARY EXPERIMENT 1
After HDT had a chance to “learn” the workload, it achieved a better response time and more throughput
Other interesting results
‒ After several cycles, HDT migrated 90% of the active PAIO data sets to tier 1, utilizing ONLY 10% of tier 1 – HDT did NOT migrate 100% of the active data sets In other words, not all of the PAIO data sets deserved to be in
the SSD tier, even though PAIO was the only thing active on the system for several hours
‒ If a customer has a performance issue in an HDT environment, more SSD capacity could be installed, increasing residency of active pages in tier 1
HDT EXAMPLE 2
This quick storyboard is another attempt to show how
HDT learns your workload
Same 600GB tier as previous experiment except at a
steady state of 24K IOPS
‒ (64) 600GB SAS drives + (8) 400GB SSD
IMPORTANT NOTE: SSDs are added to the pool after all
data sets are created.
24K IOP STEADY STATE
LET TIERING LEARN YOUR WORKLOAD
FIRST STEPS FOR HDT AND HIERARCHICAL STORAGE MANAGEMENT (HSM)
“Hierarchical storage management (HSM) is a data
storage technique which automatically moves data
between high-cost and low-cost storage media”
–Wikipedia
The following slide is an simple example of introducing
HDT in a mainframe environment that utilizes IBM
DFSMShsm
LET TIERING LEARN YOUR WORKLOAD
FIRST STEPS FOR HDT AND HSM
HDP (Pool 1)
Primary Space
Level 0
HDP (Pool 2) HSM ML1 and Low-I/O Density DP Volumes
HDP (Pool 3) HSM ML2 and Very Low-I/O Density DP Volumes
Tier 1 - SSD
Tier 2 – SAS 10K or SAS 15K
Near-Line SAS 7.5K
External Volumes
“Level 1”
Adored
“All”
Loved
“Level 5”
Liked
Unliked
Unloved
“All” with “Relocation
Priority” Most Loved
HDT FINAL EXAM
Question
‒ If you utilize all HDT dynamic tiering policies, how many levels of service can be defined in a 3-level tier?
‒ Hint: I did not talk about all of the possible levels, but you can figure it out
Lo
Tier 2 – SAS 10K or SAS 15K
WITH A 3-LEVEL TIER, 7 LEVELS OF SERVICE
Tier 1 - SSD
Tier 3 – SAS 7.5K or External
A SINGLE HDP POOL
AN EXAMPLE OF UTILIZING DYNAMIC TIERING POLICIES
“Level 1”
“Level 4”
“Level 3”
“Level 2”
“Level 5”
“ALL”
“All” with
“Relocation Priority”
SUMMARY
LET TIERING LEARN YOUR WORKLOAD
Brief description of tiering
Discuss HDT dynamic parameters
‒ HDT cycle
‒ HDT tiering policy
‒ All
‒ Level 1
‒ Level 5
‒ HDT default tier buffer space
‒ HDT new page assignment policy
‒ ALL (Low) for 2 tier, ALL (Middle) for 3 tier
Experiments 1 and 2
First steps with HDT
QUESTIONS AND DISCUSSION
UPCOMING WEBTECHS
Mainframe Series
Mainframe Replication, Sept 19, 9 a.m. PT, 12 p.m. ET
Why Networked FICON Storage Is Better than Direct-attached
Storage, Oct 3, 9 a.m. PT, 12 p.m. ET
Other
Storage Analytics, Sept 20, 9 a.m. PT, 12 p.m. ET
Maximize Availability and Uptime by Clustering your Physical
Datacenters within Metro Distances, Oct 24, 9 a.m. PT, 12 p.m. ET
Check www.hds.com/webtech for
Links to the recording, the presentation and Q&A (available next week)
Schedule and registration for upcoming WebTech sessions
THANK YOU