Download - How fast works in symmetrix vmax
1
FAST VP for EMC® Symmetrix
® VMAX
®
Theory and Best Practices
for Planning and Performance
Technical Notes P/ N 300-012-014
REV A04
June 2012
This technical notes document contains information on these top ics:
Executive summary ................................................................................... 2
Introduction and overview ....................................................................... 2 Fully Automated Storage Tiering ............................................................ 3 FAST and FAST VP comp arison .............................................................. 5 Theory of operation ................................................................................... 6 Performance considerations ................................................................... 11 Product and feature interop erability ..................................................... 14
Planning and design consid erations ..................................................... 17 Summary and conclusion ....................................................................... 34 Appendix: Best practices quick reference ............................................. 37
2
Executive summary
FAST VP for EMC Symmetrix VMAX Theory and Best Prac tices for Planning and Performanc e
Executive summary
EMC® Symmetrix
® VMAX
® Family with Enginu ity™ incorporates a
scalable fabric-interconnect design that allows a storage array to
seamlessly grow from an entry-level configuration to a 4 PB system.
Symmetrix VMAX Series arrays provide pred ictable, self-optimizing
performance and enables organ izations to scale out on demand in
private cloud environments.
VMAX Series arrays automate storage operations to exceed business
requirements in virtualized environments, with management tools that
integrate with virtualized servers and reduce ad ministra tion time in
private cloud infrastructures. Customers are able to achieve always-on
availability with maximum security, fully nondisruptive operations , and
multi-site migration, recovery, and restart to prevent application
downtime.
Information infrastructure must continuously adapt to changing
business requirements. EMC Symmetrix Fully Automated Storage
Tiering for Virtual Pools (FAST VP) automates tiered storage strategies,
in Virtual Provisioning™ environments, by easily moving workloads
between Symmetrix tiers as performance characteristics change over
time. FAST VP performs data movements, improving performance, and
reducing costs, all while maintaining vital service levels.
Introduction and overview
EMC Symmetrix VMAX FAST VP for Virtual Provisioning environments
automates the identification of data volumes for the purposes of
relocating application data across d ifferent performance/ capacity tiers
within an array. FAST VP proactively monitors workloads at both the
LUN and sub-LUN level in order to identify busy d ata that would
benefit from being moved to higher-performing drives. FAST VP also
identifies less-busy data that cou ld be moved to higher -capacity d rives
without existing performance being affected . This promotion/ demotion
activity is based on policies that associate a storage group to multip le
d rive technologies, or RAID protection schemes, by way of thin storage
pools, as well as the performance requ irements of the application
contained within the storage group. Data movement executed during
this activity is performed nondisruptively, without affecting business
continuity and data availability.
3
Fully Automated Storage Tiering
FAST VP Theory for EMC Symmetrix VMAX and Best Prac tic es for Planning and Performanc e
Audience
This technical notes document is intended for anyone who needs to
understand FAST VP theory, best practices, and associated
recommend ations as necessary to achieve the best performance for FAST
VP configurations. This document is specifically targeted to EMC
customers, sales, and field technical staff who are either running FAST
VP or are considering FAST VP for future implementation.
Significant portions of this document assume a base knowledge
regard ing the implementation and management of FAST VP. For
information regard ing the implementation and management of FAST VP
in Virtual Provisioning environments, refer to the Implementing Fully
Automated Storage Tiering for V irtual Pools (FAST VP) for EMC Symmetrix
VMAX Series Arrays Technical Note (P/ N 300-012-015).
Fully Automated Storage Tiering
Fully Automated Storage Tiering (FAST) automates the identification of
data volumes for the purposes of relocating application data across
d ifferent performance/ capacity tiers within an array , or to an external
array using Federated Tiered Storage (FTS).
The primary benefits of FAST include:
Elimination of manually tiering applications when performance
objectives change over time.
Automating the process of identifying data that can benefit from
Enterprise Flash Drives or that can be kep t on higher -capacity,
less-expensive SATA drives without impacting performance.
Improving application p erformance at the same cost, or
provid ing the same application performance at lower cost. Cost
is defined as acquisition (both hard ware and software),
space/ energy, and management expense.
Optimizing and prioritizing business applications, allowing
customers to dynamically allocate resources within a single
array.
Delivering greater flexibility in meeting d ifferent
price/ performance ratios throughout the lifecycle of the
information stored .
Due to advances in d rive technology, and the need for storage
consolid ation, the number of d rive types supported by Symmetrix arrays
has grown significantly. These drives span a range of storage service
specializations and cost characteristics that d iffer greatly.
4
Fully Automated Storage Tiering
FAST VP for EMC Symmetrix VMAX Theory and Best Prac tices for Planning and Performanc e
Several d ifferences exist between the four d rive technologies supported
by the Symmetrix VMAX Series arrays: Enterprise Flash Drive (EFD), Fibre
Channel (FC), Serial Attach SCSI (SAS), and SATA. The primary
d ifferences are:
Response time
Cost per unit of storage capacity
Cost per unit of storage request processing
At one extreme are EFDs, which have a very low response
time, but w ith a high cost per unit of storage capacity
At the other extreme are SATA drives, which have a low
cost per unit of storage capacity, bu t high response times
and high cost per u nit of storage request processing
Between these two extremes lie Fibre Channel and SAS
drives
Based on the nature of the d ifferences that exist between these four d rive
types, the following observations can be made regard ing the most suited
workload type for each drive.
Enterprise Flash Drives: EFDs are more su ited for
workloads that have a high back-end random read storage
request density. Such workload s take advantage of both
the low response time provided by the d rive, and the low
cost per unit of storage request processing without
requiring a log of storage capacity.
SATA drives: SATA drives are suited toward workloads
that have a low back-end storage request density.
Fibre Channel/ SAS drives: Fibre Channel and SAS drives
are the best d rive type for workloads with a back-end
storage request density that is not consistently high or low.
This d isparity in suitable workloads presents both an opportunity and a
challenge for storage administrators.
To the degree it can be arranged for storage workloads to be served by
the best su ited drive technology, the opportunity exists to improve
application performance, reduce hardware acquisition expenses, and
reduce operating expenses (includ ing energy costs and space
consumption).
The challenge, however, lies in how to realize these benefits w ithout
introd ucing additional ad ministrative overhead and complexity.
The approach taken with FAST is to au tomate the process of identifying
which regions of storage should reside on a given drive technology, and
5
FAST and FAST VP comparison
FAST VP Theory for EMC Symmetrix VMAX and Best Prac tic es for Planning and Performanc e
to au tomatically and nondisruptively move storage between tiers to
optimize storage resource usage accord ingly. This also needs to be done
while taking into account optional constraints on tier capacity usage that
may be imposed on specific groups of storage devices.
FAST and FAST VP comparison
EMC Symmetrix VMAX FAST and FAST VP automate the identification
of data volumes for the purposes of relocating app lication d ata across
d ifferent performance/ capacity tiers within an array, or to an external
array using Federated Tiered Storage (FTS). While the administration
procedures used with FAST VP are very similar to those available with
FAST, there is a major d ifference: The storage pools used by FAST VP are
thin storage pools.
FAST operates on non-thin, or d isk-group-provisioned , Symmetrix
volumes. Data movements executed between tiers are performed at the
full-volume level. FAST VP operates on Virtual Provisioning thin
devices. As such, data movements execu ted can be performed at the sub-
LUN level, and a single thin d evice may have extents allocated across
multiple thin pools within the array, or on an external array using FTS.
Note: For more information on Virtual Provisioning, refer to the Best Practices
for Fast, Simple Capacity Allocation with EMC Symmetrix Virtual Provisioning
Technical Note available at http :/ / powerlink.emc.com.
Note: For more information on Federated Tiered Storage, refer to the Design
and Implementation Best Practices for EMC Symmetrix Federated Tiered Storage
(FTS) technical note available at http:/ / powerlink.emc.com.
Because FAST and FAST VP support d ifferent device types, non-thin and
thin, respectively, they both can operate simultaneously within a single
array. Aside from some shared configuration parameters, the
management and operation of each can be considered separately.
Note: For more information on FAST refer to the Implementing Fully
Automated Storage Tiering (FAST) for EMC Symmetrix VMAX Series Arrays
technical note available at http :/ / powerlink.emc.com.
The goal of FAST and FAST VP is to optimize the performance and cost-
efficiency of configurations containing mixed drive technologies. While
FAST monitors and moves storage in units of entire logical devices,
FAST VP monitors data access with much finer granularity. This allow s
FAST VP to determine the most appropriate tier (based on optimizing
performance and cost efficiency) for each 120 track region of storage. In
6
Theory of operation
FAST VP for EMC Symmetrix VMAX Theory and Best Prac tices for Planning and Performanc e
this way, the ability of FAST VP to monitor and move data with much
finer granularity greatly enhances the value proposition of au tomated
tiering.
FAST VP builds upon and extends the existing capabilities of Virtual
Provisioning and FAST to provide the user with enhanced Symmetrix
tiering options. The Virtual Provisioning underpinnings of FAST VP
allow FAST VP to combine the core benefits of Virtual Provisioning
(wide striping and thin provisioning) with the benefits of automated
tiering.
FAST VP more closely aligns storage access workload s with the best-
suited drive technology than is possible if all regions of a given dev ice
must be mapped to the same tier. At any given time, the hot regions of a
thin device managed by FAST VP may be mapped to an EFD tier, and
the warm parts may be mapped to an FC tier , while the cold parts may
be mapped to a SATA tier. By more effectively exploiting drive
technology specializations, FAST VP delivers better performance and
greater cost efficiency that FAST.
FAST VP also better ad apts to shifting workload locality of reference,
changes in tier-allocation limits, or storage-group priority. This is due to
the fact that the workload may be ad justed by moving less data. This
further contributes to making FAST VP more effective at exploiting drive
specializations and also enhances some of the operational advantages of
FAST. This includes the ability to nondisruptively ad just the quality of
storage service (response time and throughput) provid ed to a storage
group.
Theory of operation
There are two components of FAST VP: Symmetrix Enginu ity and the
FAST controller. Symmetrix Enginuity storage is the operating
environment that controls components w ithin the array. The FAST
controller is a service that runs on the service processor.
7
Theory of operation
FAST VP Theory for EMC Symmetrix VMAX and Best Prac tic es for Planning and Performanc e
Figure 1 . FAST VP components
When FAST VP is active, both components participate in the execution of
two algorithms (the intelligent tiering algorithm and the allocation
compliance algorithm) to d etermine appropriate d ata placement.
The intelligent tiering algorithm uses performance data collected by
Enginuity, as well as supporting calculations performed by the FAST
controller, to issue d ata movement requests to the Virtual LUN (VLUN)
VP data movement engine.
The allocation compliance algorithm enforces the upper limits of storage
capacity that can be used in each tier by a given storage group by also
issuing d ata movement requests to the VLUN VP d ata movement
engine.
Performance time windows can be defined to specify when the FAST VP
controller should collect performance data, upon which analysis is
performed to determine the appropriate tier for devices. By default, this
8
Theory of operation
FAST VP for EMC Symmetrix VMAX Theory and Best Prac tices for Planning and Performanc e
occurs 24 hours a day.
Defined data movement windows determine when to execute the data
movements necessary to move data between tiers.
Data movements performed by Enginuity are achieved by moving
allocated extents between tiers. The size of data movement can be as
small as 12 tracks, representing a single allocated thin device extent, but
more typ ically it is a unit known as an extent group, which is 120 tracks.
The following sections further describe each of these algorithms.
Intelligent tiering algorithm
The goal of the intelligent tiering algorithm is to use the performance
metrics collected at the sub-LUN level to determine which tier each
extent group should reside in and to submit the needed data movements
to the VLUN VP d ata movement engine. The determination of which
extent groups need to be moved is performed by a task that runs within
the Symmetrix array.
The intelligent tiering algorithm is structured into two components : a
main component that execu tes within Enginuity and a second ary,
supporting component that executes within the FAST controller on the
service processor.
The main component assesses whether extent groups need to be moved
in order to optimize the use of the FAST VP storage tiers. If so, the
required data movement requests are issued to the VLUN VP data
movement engine.
When determining the appropriate tier for each extent group, the main
component makes use of both the FAST VP metrics, p reviously
d iscussed , and supporting calculations performed by the secondary
component on the service p rocessor.
The intelligent tiering algorithm runs continuously during open data
movement wind ows when FAST is enabled and the FAST VP operating
mode is Automatic. As such, performance-related data movements can
occur continuously during an open data movement window.
Allocation compliance algorithm
The goal of the allocation compliance algorithm is to detect and correct
situations where the allocated capacity for a particular storage group
within a thin storage tier exceeds the maximum capacity allowed by the
associated FAST policy.
9
Theory of operation
FAST VP Theory for EMC Symmetrix VMAX and Best Prac tic es for Planning and Performanc e
A storage group is considered to be in compliance with its associated
FAST policy when the configured capacity of the thin d evices in the
storage group is located on tiers defined in the policy , and when the
usage of each tier is w ithin the upper limits of the tier usage limits
specified in the policy.
Compliance violations may occur for multiple reasons, includ ing:
New extent allocations performed for thin devices
managed by FAST VP
Changes made to the upper usage limits for a VP tier in a
FAST policy
Adding thin devices to a storage group that are themselves
out of compliance
Manual VLUN VP migrations of thin devices
The compliance algorithm attempts to minimize the amount of
movements performed to correct compliance that may, in turn, generate
movements performed by the intelligent tiering algorithm . This is d one
by coord inating the movement requests w ith the analysis performed by
the intelligent tiering algorithm to determine the most appropriate
extents to be moved , and the most appropria te tier to be moved to, when
correcting compliance violations.
When FAST is enabled and the FAST VP operating mode is Automatic,
the compliance algorithm runs every 10 minutes during open d ata
movement wind ows.
Data movement
Data movements execu ted by FAST VP are performed by the VLUN VP
data movement engine, and involve moving thin device extents between
thin pools w ithin the array. Extents are moved by way of a move
process only; extents are not swapped between pools.
The movement of extents, or extent groups, does not change the thin
device bind ing information. The thin device still remain s bound to the
pool it was originally bound to. New allocations for the thin device, as
the result of host writes, will continue to come from the bound pool,
unless VP allocation by FAST VP is enabled .
To complete a move, the following must hold true:
The FAST VP operating mode must be Automatic.
The VP data movement window must be open.
The thin device affected must not be pinned .
There must be sufficient unallocated space in the thin
10
Theory of operation
FAST VP for EMC Symmetrix VMAX Theory and Best Prac tices for Planning and Performanc e
pools included in the destination tier to accommodate the
data being moved .
The destination tier must contain at least one thin pool that
has not exceeded the pool reserved capacity (PRC).
Note: If the selected destination tier contains only pools that have reached the
PRC limit, then an alternate tier may be considered by the movement task.
Other movement considerations include:
Only extents that are allocated will be moved .
No back-end configuration changes are performed during
a FAST VP data movement, and , as such, no configuration
locks are held during the process.
As swaps are not performed, there is no requirement for
any swap space, such as DRVs, to facilitate data
movement.
Data movement time wind ows are used to specify date and time ranges
when d ata movements are allowed , or not allowed , to be performed .
FAST VP data movements run as low -priority tasks on the Symmetrix
back end . Data movement wind ows can be planned so as to minimize
impact on the performance of other more critical workloads.
A default data movement time window excludes all performance data
samples 24 hours a d ay, 7 d ays a week, 365 days a year. There are two
types of data movement that can occur under FAST VP: Those generated
by the intelligent tiering algorithm and those generated by the allocation
compliance algorithm. Both types of data movement only occur during
user-defined data movement windows.
Related movements of the intelligent tiering algorithm are requested and
executed by Enginuity. These data movements are governed by the
workload on each extent group, but may only be execu ted within the
constraints of the associated FAST policy. Therefore, a performance
movement cannot cause a storage group to become non-compliant with
its FAST policy.
Allocation compliance related movements are generated by the FAST
controller and executed by Enginuity. These movements bring the
capacity of the storage group back within the boundaries specified by the
associated policy. Performance information from the intelligent tiering
algorithm is used to determine more appropriate sub-extents to move
when restoring compliance.
When a compliance violation exists, the algorithm generates a d ata
movement request to return the allocations within the required limits.
11
Performance considerations
FAST VP Theory for EMC Symmetrix VMAX and Best Prac tic es for Planning and Performanc e
This request explicitly ind icates which thin device extents should be
moved , and the specific thin pools they should be moved to.
Performance considerations
The Enginuity operating environment collects and maintains
performance data that is used by FAST VP. The FAST controller analyzes
the performance d ata to determine the best location to place the thin
device data on the defined VP tiers w ithin the Symmetrix array.
Performance metrics
When collecting performance data at the LUN and sub-LUN level for use
by FAST VP, Enginuity only collects statistics related to Symmetrix back-
end activity that is the result of host I/ O. The metrics collected are:
Read miss
Write
Prefetch (sequential read )
The read miss metric accounts for each DA read operation that is
performed. Read s to areas of a thin device that have not had space
allocated in a thin pool are not counted . Also, read hits, which are
serviced from cache, are not considered .
Write operations are counted in terms of the number of d istinct DA
operations that are performed. The metric accounts for when a write is
destaged – write hits, to cache, are not considered .
Writes related to specific RAID protection schemes are also not counted .
In the case of RAID 1 protected devices, the write I/ O is only counted for
one of the mirrors. In the case of RAID 5 and RAID 6 protected devices,
parity reads and writes are not counted .
Prefetch operations are accounted for in terms of the number of d istinct
DA operations performed to prefetch data spanning a FAST VP extent.
This metric considers each DA read operation performed as a front-end
prefetch operation.
Workload related to internal copy operations, such as d rive rebuilds,
clone operations, VLUN migrations, or even FAST VP data movements,
is not included in the FAST VP metrics.
These FAST VP performance metrics provide a measure of activity that
assigns greater weight to more recent I/ O requests, but are also
influenced by less recent activity. By default, based on a workload
analysis period of 24 hours, an I/ O that has just been received is
weighted two times more heavily than an I/ O received 24 hours
12
Performance considerations
FAST VP for EMC Symmetrix VMAX Theory and Best Prac tices for Planning and Performanc e
previously.
Note: Performance metrics are only collected during user -defined
performance time windows. The times during which metrics are not being
collected does not contribute to reducing the weight assigned to those metrics
already collected .
The metrics collected at the sub-LUN level for thin devices under FAST
VP control contain measurements to allow FAST VP to make separate
data movement requests for each 120 tracks unit of storage that make up
the thin device. This unit of storage consists of 10 contiguous thin device
extents and is known as an extent group.
In order to maintain the sub-LUN-level metrics collected by Enginuity,
the Symmetrix array allocates one cache slot for each thin device that is
under FAST VP control.
When managing metadevices, cache slots are allocated for both the
metahead and for each of the metamembers.
Note: Each cache slot on a Symmetrix VMAX Series array is one track in size.
FAST VP tuning
FAST VP provides a number of parameters that can be used to tune the
performance of FAST VP and to control the aggressiveness of the data
movements. These parameters can be used to nondisruptively ad just the
amount of tier storage that a given storage group is allowed to use, or to
ad just the manner in which storage groups using the same tier compete
with each other for space.
FAST VP relocation rate
The FAST VP relocation rate (FRR) is a quality of service (QoS) setting
for FAST VP and affects the aggressiveness of d ata movement requests
generated by FAST VP. This aggressiveness is measured as the amount
of data that is requested to be moved at any given time, and the priority
given to moving the d ata between pools.
Note: The rate at which data is moved between pools can also be controlled
by means of the Symmetrix Quality of Service VLUN setting.
Pool reserved capacity
The Pool reserved capacity (PRC) reserves a percentage of each pool
included in a VP tier for non-FAST VP activities. The purpose of this is to
ensure that FAST VP d ata movements do not fill a thin pool, and
13
Performance considerations
FAST VP Theory for EMC Symmetrix VMAX and Best Prac tic es for Planning and Performanc e
subsequently cause a new extent allocation (a result of a host write) to
fail.
The PRC can be set both system -wide and for each ind ividual pool. By
default, the system-wide setting is applied to all thin pools that have
been included in VP tier definitions. However, this can be overridden for
each ind ividual pool by using the pool-level setting.
When the percentage of unallocated space in a thin pool is equal to the
PRC, FAST VP no longer perform s data movements into that pool.
However, data movements may continue to occur out of the pool to
other pools. When the percentage of unallocated space becomes greater
than the PRC, FAST VP can begin performing data movements into that
pool again.
FAST VP time windows
FAST VP utilizes time windows to define certain behaviors regard ing
performance data collection and data movement. There are two possible
wind ow types:
Performance time window
Data movement time wind ow
The performance time windows are used to specify when performance
metrics should be collected by Enginuity.
The data movement time windows define when to perform the data
relocations necessary to move data between tiers. Separate data
movement wind ows can be defined for full LUN movement, performed
by FAST and Optimizer, and sub-LUN data movement performed by
FAST VP.
Both performance time windows and d ata movement wind ows may be
defined as inclusion or exclusion windows. An inclusion time window
ind icates that the action should be performed during the defined time
wind ow. An exclusion time wind ow ind icates that the action should be
performed outside the defined time wind ow.
Performance time window
The performance time windows are used to identify the business cycle
for the Symmetrix array. They specify d ate and time ranges (past or
future) when performance samples should be collected , or not collected ,
for the purposes of FAST VP performance analysis. The intent of
defining performance time wind ows is to d istingu ish periods of time
when the Symmetrix array is id le from periods when it is active, and to
only include performance d ata collected during the active periods.
By default, performance metrics are collected 24 hours a day, 7 days a
14
Product and feature interoperability
FAST VP for EMC Symmetrix VMAX Theory and Best Prac tices for Planning and Performanc e
week, 365 d ays a year.
Data movement time window
Data movement time wind ows are used to specify date and time ranges
when d ata movements are allowed , or not allowed , to be performed .
FAST VP data movements run as low-priority tasks on the Symmetrix
back end . By default, data movement is prevented 24 hours a day, 7 d ays
a week, 365 days a year.
Storage group priority
When a storage group is associated with a FAST policy, a priority value
must be assigned to the storage group. This priority value can be
between 1 and 3, with 1 being the highest priority . The default is 2.
When multiple storage groups share the same policy, the priority value
is used when the data contained in the storage groups is competing for
the same resources in one of the associated tiers. Storage groups with a
higher priority are given preference when decid ing which data needs to
be moved to another tier.
Product and feature interoperability
FAST VP is fully interoperable with all Symmetrix replication
technologies: EMC SRDF®, EMC TimeFinder
®/ Clone, TimeFinder/ Snap,
and Open Replicator. Any active replication on a Symmetrix device
remains intact while data from that device is being moved . Similarly, all
incremental relationships are maintained for the moved or swapped
devices. FAST VP also operates alongside Symmetrix features, such as
Symmetrix Optimizer, Dynamic Cache Partitioning, and Auto-
provisioning Groups.
SRDF
Thin SRDF devices, R1 or R2, can be associated with a FAST policy.
Extents of SRDF devices can be moved between tiers while the devices
are being actively replicated , in either synchronous or asynchronous
mode.
TimeFinder/ Clone
Both the source and target devices of a TimeFinder/ Clone session can be
managed by FAST VP. However, the source and target are managed
independently, and , as such, may end up with d ifferent extent
allocations across tiers.
15
Product and feature interoperability
FAST VP Theory for EMC Symmetrix VMAX and Best Prac tic es for Planning and Performanc e
TimeFinder/ Snap
The source device in a TimeFinder/ Snap session can be managed by
FAST VP. However, target device VDEVs are not managed by FAST VP.
TimeFinder VP Snap
The source device in a TimeFinder VP Snap session can be managed by
FAST VP. Target devices may also be managed by FAST VP, however,
extent allocations that are shared by multiple target devices will not be
moved .
Open Replicator for Symmetrix
The control device in an Open Replicator session, push or pull, can have
extents moved by FAST VP.
Virtual Provisioning
All thin devices, whether under FAST VP control or not, may only be
bound to a single thin pool. All host-write-generated allocations, or user-
requested pre-allocations, are performed to this pool. FAST VP data
movements do not change the bind ing information for a thin device.
It is possible to change the bind ing information for a thin device without
changing any of the current extent allocations for the d evice. However,
when rebind ing a device that is under FAST VP control, the th in pool the
device is being re-bound to must belong to one of the VP tiers contained
in the policy that the device is associated with.
Virtual Provisioning space reclamation
Space reclamation may be run against a thin device under FAST VP
control. However, during the space reclamation process, sub-LUN
performance metrics are not updated , and data movements are not
performed.
Note: If FAST VP is actively moving extents of a device, a request to reclaim
space on that device will fail. Prior to issuing the space reclamation task , the
device should first be pinned. This suspend s any active FAST VP data
movements for the device and allow s the request to succeed .
Virtual Provisioning T10 unmap
Unmap command s can be issued to thin devices under FAST VP control.
The T10 SCSI unmap command for thin devices advises a target thin
device that a range of blocks are no longer in use. If this range covers a
full thin device extent, that extent can be deallocated and the free space
returned to the pool.
If the unmap command range covers only some tracks in an extent, those
tracks are marked Never Written by Host (NWBH). The extent is not
16
Product and feature interoperability
FAST VP for EMC Symmetrix VMAX Theory and Best Prac tices for Planning and Performanc e
deallocated ; however those tracks will not have to be retrieved from d isk
should a read request be performed. Instead , the Symmetrix array
immed iately returns all zeros.
Virtual Provisioning pool management
Data devices may be added to or removed from a thin pool that is
included in the FAST VP tier. Data movements related to FAST VP, into
or out of the thin pool, continue while the data devices are being
modified .
In the case of add ing d ata devices to a thin pool, automated pool
rebalancing may be run. Similarly, when d isabling and removing data
devices from the pool, they drain their allocated tracks to other enabled
data devices in the pool.
While both d ata device d raining and automated pool rebalancing may be
active in a thin pool that is included in a VP tier, both of these processes
may affect performance of FAST VP data movements.
Virtual LUN VP mobility
A thin device under FAST VP control may be migrated using VLUN VP.
Such a migration resu lts in all allocated extents of the device being
moved to a single thin pool.
While the migration is in progress, no FAST VP-related data movements
are performed. However, once the migration is complete, all allocated
extents of the thin device will be available to be retiered .
To prevent the migrated device from being retiered by FAST VP
immed iately following the migration, it is recommend ed that the device
first be pinned . To re-enable FAST VP-related data movements, the
device can later be unpinned .
FAST
Both FAST and FAST VP may coexist w ithin a single Symmetrix array.
FAST only performs full device movements of non-thin devices. As such,
there is no impact to FAST VP’s management of thin d evices.
Both FAST and FAST VP share some configuration parameters. These
are:
Workload Analysis Period (WAP)
Initial Analysis Period (IAP)
Performance time windows
Symmetrix Optimizer
Symmetrix Optimizer operates only on non-thin devices. As a result,
there is no impact on FAST VP’s management of thin d evices.
17
Planning and design considerations
FAST VP Theory for EMC Symmetrix VMAX and Best Prac tic es for Planning and Performanc e
Both Optimizer and FAST VP share some configuration parameters.
These are:
Workload Analysis Period (WAP)
Initial Analysis Period (IAP)
Performance time windows
Dynamic Cache Partitioning (DCP)
Dynamic Cache Partitioning can be used to isolate storage hand ling of
d ifferent applications. As data movements use the same cache partition
as the application, movements of d ata on behalf of one application d o
not affect the performance of applications that are not sharing the same
cache partition.
Auto-provisioning Groups
Storage groups created for the purposes of Auto-provisioning may also
be used for FAST VP. However, while a device may be contained in
multiple storage groups for the purposes of Auto-provisioning, it may
only be contained in one storage group that is associated with a FAST
policy (DP or VP).
Should a storage group contain a mix of device types, thin and non -thin,
only the devices matching the type of FAST policy it is associated with
are managed by FAST.
If it is intended that both d evice types in an Auto-provisioning storage
group are to be managed by FAST and FAST VP, respectively, then
separate storage groups need to be created . A storage group with the
non-thin devices may then be associated with a policy containing DP
tiers. A separate storage group containing the thin devices will be
associated with a policy containing VP tiers.
Planning and design considerations
The following sections detail best practice recommend ations for
planning the implementation of a FAST VP environment .
The best practices documented are based on features available in
Enginuity 5876.82.57, Solutions Enabler 7.4, Symmetrix Management
Console 7.3.3, and Unisphere® for VMAX 1.0.
FAST VP configuration parameters
FAST VP has multiple configuration parameters that control its behavior.
These include settings to determine the effect of past workloads on d ata
analysis, quality of service for data movements, and pool space to be
18
Planning and design considerations
FAST VP for EMC Symmetrix VMAX Theory and Best Prac tices for Planning and Performanc e
reserved for non-FAST VP activities. Also, performance collection and
data movement time wind ows can be defined .
The following sections describe best practice recommendations for each
of these configuration p arameters.
Note: For more information on each of these configuration parameters, refer
to the Implementing Fully Automated Storage Tiering for Virtual Pools (FAST VP)
for EMC Symmetrix VMAX Series Arrays technical note available at
http://powerlink.emc.com.
Performance time window
The performance time windows specify date and time ranges when
performance metrics should be collected , or not collected , for the
purposes of FAST VP performance analysis. By default, performance
metrics are collected 24 hours a day, every day. Time windows may be
defined , however, to includ e only certain days or days of the week, as
well as exclude other time periods.
As a best practice, the default performance wind ow sh ould be left
unchanged . However, if there are extended periods of time when the
workloads managed by FAST VP are not active, these time period s
should be excluded .
Note: The performance time window is applied system -wide. If multiple
applications are active on the array, but active at d ifferent times, then the
default performance time window behavior should be left unchanged .
Data movement time window
Data movement time wind ows are used to specify date and time ranges
when d ata movements are allowed , or not allowed , to be performed .
The best practice recommendation is that, at a minimum, the data
movement wind ow should allow d ata movements for the same period of
time that the performance time windows allow d ata collection. This
allow s FAST VP to react more quickly and more dynamically to any
changes in workload that occur on the array.
Unless there are specific time periods to avoid data movements, during a
backup window, for example, it may be appropriate to set the data
movement wind ow to allow FAST VP to perform movements 24 hours a
day, every day.
19
Planning and design considerations
FAST VP Theory for EMC Symmetrix VMAX and Best Prac tic es for Planning and Performanc e
Note: If there is a concern about possible impact of data movements
occurring during a production workload then the FAST VP Relocation Rate
can be used to minimize this impact.
Workload Analysis Period
The Workload Analysis Period (WAP) determines the degree to which
FAST VP metrics are influenced by recent host activity, and also less-
recent host activity, that takes place while the performance time wind ow
is considered open.
The longer the time defined in the WAP, the greater the amount of
weight is assigned to less recent host activity.
The best practice recommendation for the WAP is to use the default
value of 168 hours (1 week).
Initial Analysis Period
The Initial Analysis Period (IAP) defines the minimum amount of time a
thin device should be under FAST VP management before any
performance-related d ata movements should be applied .
This parameter should be set to a long enough value so as to allow
sufficient d ata samples for FAST VP to establish a good characterization
of the typical workload on that device.
At the initial deployment of FAST VP, it may make sense to set the IAP
to 168 hours (1 week) to ensure that a typical weekly workload cycle is
seen. However, once FAST VP data movement has begun, you may
reduce the IAP to 24 hours (1 day), which allow s the newly associated
devices to benefit from FAST VP movement recommendations more
quickly.
FAST VP Relocation Rate
The FAST VP Relocation Rate (FRR) is a quality of service (QoS) setting
for FAST VP. The FRR affects the aggressiveness of data movement
requests generated by FAST VP. This aggressiveness is measured as the
amount of data that is requested to be moved at any given time, and the
priority given to moving the data between pools.
Setting the FRR to the most aggressive value, 1, causes FAST VP to
attempt to move the most d ata it can, as quickly as it can. Dependent on
the amount of d ata to be moved , an FRR of 1 is more likely to cause
impact to host I/ O response times, due to the add itional back -end
overhead being generated by the FAST VP data movements. However,
the d istribution of d ata across tiers is completed in a shorter period of
time.
Setting the FRR to the least aggressive value, 10, causes FAST VP to
20
Planning and design considerations
FAST VP for EMC Symmetrix VMAX Theory and Best Prac tices for Planning and Performanc e
greatly reduce the amount of data that is moved , and the pace at which it
is moved . This setting causes no impact to host response time, but the
final d istribu tion of data takes longer.
Figure 2 shows the same workload , 1,500 IOPS of type OLTP2, being run
on an environment containing two FAST VP tiers, Fibre Channel (FC)
and Enterprise Flash (EFD). The same test was carried out w ith three
separate relocation rates: 1, 5, and 8.
With a FRR of 1, an initial increase in response time is seen at the two-
hour mark, when FAST VP data movement began, while no increase in
response time is seen when the relocation rate is set to 8. However, the
steady state for the response time is seen in a much shorter period of
time for the lower, more aggressive, setting.
Figure 2. Example workload with varying relocation rates
The default value for the FRR is 5. However, the best practice
recommend ation for the initial deployment of FAST VP is to start with a
more conservative value for the relocation rate, perhaps 7 or 8. The
reason for this is that when FAST VP is first enabled , the amount of data
to be moved is likely to be greater , compared to when FAST VP has been
running for some time.
At a later d ate, when it is seen that the amount of data movement s
between tiers is less, the FRR can be set to a more aggressive level,
21
Planning and design considerations
FAST VP Theory for EMC Symmetrix VMAX and Best Prac tic es for Planning and Performanc e
possibly 2 or 3. This allow s FAST VP to ad just to small changes in
workload more quickly.
Pool Reserved Capacity
The Pool Reserved Capacity (PRC) reserves a percentage of each pool
included in a VP tier for non-FAST VP activities. When the percentage of
unallocated space in a thin pool is equal to the PRC, FAST VP no longer
performs data movements into that pool.
The PRC can be set both as a system -wide setting and for each ind ivid ual
pool. If the PRC has not been set for a pool, or the PRC for the pool has
been set to NONE, then the system-wide setting is used .
For the system-wide setting, the best practice recommendation is to use
the default value of 10 percent.
For ind ividual pools, if thin devices are bound to the pool, the best
practice recommendation is to set the PRC based on the lowest allocation
warning level for that thin pool. For example, if a warning is triggered
when a thin pool has reached an allocation of 80 percent of its capacity,
then the PRC should be set to 20 percent. This ensures that the remaining
20 percent of the pool is only used for new host-generated allocations,
and not FAST VP d ata movements.
If no thin devices are bound to the pool, or are going to be boun d , then
the PRC should be set to the lowest possible value, 1 percent.
VP Allocation by FAST Policy
The VP allocation by FAST policy feature allows new allocations to come
from any of the thin pools included in the FAST VP policy that the thin
device is associated with.
With this feature enabled , FAST VP attempts to allocate new writes in
the most appropriate tier first, based on available performance metrics. If
no performance metrics are available, the allocation is attempted in the
pool the device is bound to.
If the pool initially chosen to allocate the d ata is full, FAST VP then look s
to other pools contained within the FAST VP policy and allocate from
there. As long as there is space available in at least one of the pools
within the policy, all new extent allocations will be successful.
The allocation by policy feature is enabled at the Symmetrix array level
and applies to all allocations for all devices managed by FAST VP. The
feature is either enabled or d isabled (the default setting is d isabled ).
When d isabled , new allocations only come from the pool the thin device
is bound to.
As a best practice, it is recommended that VP allocation by FAST policy
22
Planning and design considerations
FAST VP for EMC Symmetrix VMAX Theory and Best Prac tices for Planning and Performanc e
be enabled .
Note: For more information on the decision making process of the VP
allocation by FAST policy feature, refer to the Advance FAST VP features
section of the Implementing Fully Automated Storage Tiering for Virtual Pools
(FAST VP) for EMC Symmetrix VMAX Series Arrays technical note.
FAST VP tier configuration
A Symmetrix storage tier is a specification of a set of resources of the
same d isk technology type (EFD, FC/ SAS, SATA, or FTS storage)
combined with a given RAID protection type (RAID 1, RAID 5, RAID 6,
or Unprotected), and the same emulation (FBA or CKD).
Note: The unprotected RAID type may only be applied to a tier resid ing on
an FTS-connected storage array.
FAST VP tiers can contain between one and four thin storage pools . Each
thin pool must contain d ata devices configured on the same drive
technology and emulation (and the same rotational speed , in the case of
FC, SAS, and SATA drives). However, two or more thin pools containing
data devices configured on rotating drives of d ifferent speeds may be
combined in a single VP tier.
Drive-size considerations
Drive size is not a factor when add ing d ata devices to a thin pool. For
example, data devices configured on 300GB FC 15k drives can coexist in
a pool with d ata devices configured on 600GB FC 15k drives.
However, when planning to have data devices on d ifferent d rive s izes
exist in the same storage tier, it is recommended to create two separate
pools for each drive size, and then combine those two pools into a single
tier.
Note: For more information on best practices for configuring thin pools, refer
to the Best Practices for Fast, Simple Capacity Allocation with EMC Symmetrix
Virtual Provisioning Technical Note available at http:/ / powerlink.emc.com .
External tiers
When creating a tier on an external array using FTS, it is recommended
that the eDisks be configured on a single array. The eDisks should also
be configured on similar d rives within that external array.
External tiers can only be configured on storage configured for external
provisioning. Encapsulated storage is n ot supported by FAST VP.
23
Planning and design considerations
FAST VP Theory for EMC Symmetrix VMAX and Best Prac tic es for Planning and Performanc e
Note: For more information on best practices for configuring eDisks, refer to
the Design and Implementation Best Practices for EMC Symmetrix Federated Tiered
Storage (FTS) technical note available at http:/ / powerlink.emc.com.
FAST VP policy configuration
A FAST VP policy groups between one and three VP tiers and assigns an
upper-usage limit for each storage tier. The upper limit specifies the
maximum amount of capacity of a storage group associated with the
policy can reside on that particular tier. The upper capacity usage limit
for each storage tier is specified as a percentage of the configured , logical
capacity of the associated storage group.
Figure 3. Storage tier, policy, storage group association
The usage limit for each tier must be between 1 percent and 100 percent.
When combined , the upper usage limit for all thin storage tiers in the
policy must total at least 100 percent, but may be greater than 100
percent.
Tier ranking
FAST VP ranks tiers within a policy based on known performance
models. For internal tiers, FAST VP orders the three available
technologies in terms of performance capabilities in descending order :
EFD, Fibre Channel/ SAS, and SATA.
FAST VP has no mechanism for determining the potential performance
of an external, FTS tier. As a result, any external tier included in a policy
is considered to be the lowest tier in that policy.
Tier usage limits
Creating a policy with a total upper usage limit greater than 100 percent
allows flexibility with the configuration of a storage group , whereby data
may be moved between tiers without necessarily having to move a
corresponding amount of other data within the same storage group.
The ideal FAST VP policy would be to specify 100 percent for each of the
included tiers. Such a policy would provide the greatest amount of
flexibility to an associated storage group, as it would allow 100 percent
24
Planning and design considerations
FAST VP for EMC Symmetrix VMAX Theory and Best Prac tices for Planning and Performanc e
of the storage group’s capacity to be promoted or demoted to any tier
within the policy.
While ideal, operationally it may not be appropriate to deploy the
100/ 100/ 100 policy. There may be reasons to limit access to a particular
tier within the array.
As an example, it may be appropriate to limit the amount of a storage
group’s capacity that can be placed on EFD. This may be used to prevent
one single storage group, or application, from consuming all of the EFD
resources. In this case, a policy containing just a small percentage for
EFD would be recommend ed .
Similarly, it may be appropriate to restrict the amount of SATA capacity
a storage group may utilize. Some applications, which can become
inactive from time to time, may require a minimum level of performance
when they become active again. For such applications, a policy exclud ing
the SATA tier could be appropriate.
The best way to determine appropriate policies for a FAST VP
implementation is to examine the workload skew for the application
data to be managed by FAST VP. The workload skew defines an
asymmetry in d ata usage over time. This means a small percentage of the
data on the array may be servicing the majority of the workload on the
array. One tool that provid es insight into this workload skew is Tier
Advisor.
Tier Advisor
Tier Advisor is a utility available to EMC technical staff that estimates
the performance and cost of mixing drives of d ifferent technology types
(EFD, FC, and SATA) within Symmetrix VMAX Series storage arrays.
Tier Advisor can examine performance data collected from Symmetrix,
VNX®, or CLARiiON
® storage arrays and determine the workload skew
at the full LUN level. It can also estimate the workload skew at the sub -
LUN level.
With this information, Tier Advisor can model an optimal storage array
configuration by enabling the ability to interactively experiment with
d ifferent storage tiers and storage policies, until achieving the desired
cost and performance preferences.
Thin device binding
Each thin device, whether under FAST VP control or not, may only be
bound to a single thin pool. By default, all host write generated
allocations, or user-requested pre-allocations, are performed from this
pool.
25
Planning and design considerations
FAST VP Theory for EMC Symmetrix VMAX and Best Prac tic es for Planning and Performanc e
From an ease-of-management and reporting perspective, it is
recommended that all thin devices be bound to a single pool w ithin the
Symmetrix array.
Note: FAST VP data movements do not change the bind ing information for a
thin device.
In determining the appropriate pool, both performance requirements
and capacity management should be taken into consid eration , as well as
the use of thin device preallocation and the system write pending limit.
Note: Unless there is a very specific performance need , it is not
recommended to bind thin devices, under FAST VP control, to the EFD tier.
Performance consideration
With VP allocation by FAST policy enabled , FAST VP attempts to
allocate new extents in the most appropriate tier. This is based on
available performance metrics for the thin device being allocated . Should
performance metrics not be available, the allocation comes from the pool
the thin device is bound to.
As such, the best practice recommend ation is to bind all thin devices to a
pool within the second highest tier in the policy.
In a three-tier configuration (EFD, FC, and SATA), this would imply the
Fibre Channel tier. In a FC and SATA configuration, this would imply
the SATA tier. Once the data has been written, and space has been
allocated in the pool, FAST VP then makes the decision to promote or
demote the data as appropriate.
Capacity management consideration
Binding all thin devices to a single pool u ltimately causes that single
pool to be oversubscribed . If the allocation by policy feature is not
enabled , this could potentially lead to issues as the pool fills up. Host
writes to unallocated areas of a thin device will fail if there is insufficient
space in the bound pool.
In an EFD, FC, SATA configuration, if the FC tier is significantly smaller
than the SATA tier, bind ing all thin devices to the SATA tier reduces the
likelihood of the bound pool filling up.
If performance needs requ ire the thin devices to be bound to FC, the Pool
Reserved Capacity, the FAST VP policy configuration , or a combination
of both, can be used to alleviate a potential pool-full condition.
26
Planning and design considerations
FAST VP for EMC Symmetrix VMAX Theory and Best Prac tices for Planning and Performanc e
Note: Enabling VP allocation by FAST policy on the Symmetrix VMAX Series
array greatly alleviates the capacity-management consideration .
Preallocation
A way to avoid writes to a thin device failing due to a pool being fully
allocated is to preallocate the thin device when it is bound . However,
performance requ irements of newly written data should be considered
prior to using preallocation.
When FAST VP performs d ata movements, only allocated extents are
moved . This applies not only to extents allocated as the result of a host
write, but also to extents that have been preallocated . These preallocated
extents w ill be moved even if no data has yet been written to them.
Note: When moving preallocated , but unwritten, extents, no data is actually
moved . The pointer for the extent is simply red irected to the pool in the
target tier.
Preallocated , but unwritten, extents show as inactive and , as a result, are
demoted to the lowest tier included in the associated FAST VP policy.
When these extents are eventually written to, the write performance will
be that of the tier it has been demoted to.
A best practice recommend ation is to not preallocate th in devices
managed by FAST VP. Preallocation should only be used selectively for
those devices that can never tolerate a write failure due to a full pool.
System write pending limit
FAST VP is designed to throttle back data movements, both promotions
and demotions, as the write pending count approaches the system write
pending limit. This throttling gives an even higher priority to host I/ O to
ensure that tracks marked as write pending are destaged appropriately.
Note: By default, the system write pending limit on a Symmetrix VMAX
Series array is set to 75 percent of the available cache.
If the write pending count reaches 60 percent of the write pending limit,
FAST VP data movements stop. As the write pending count decreases
below this level, data movements automatically restart.
A very busy workload running on SATA disks, with SATA disks at, or
near, 100 percent utilization, causing a high write pending count,
impacts FAST VP from promoting active extents to FC and/ or EFD tiers .
In an environment with a high write workload , it is recommended to
bind thin devices to a pool in the FC tier.
27
Planning and design considerations
FAST VP Theory for EMC Symmetrix VMAX and Best Prac tic es for Planning and Performanc e
Migration
When performing a migration to thin devices, it is possible that the thin
devices may become fully allocated as a resu lt of the migration. As such,
the thin devices being migrated to should be bound to a pool that has
sufficient capacity to contain the full capacity of each of the devices.
Virtual Provisioning zero space reclamation can be used following the
migration to deallocate zero data copied during the migration.
Alternatively, the front-end zero detection capabilities of SRDF and
Open Replicator can be used during the migration.
Rebinding a thin device
It is possible to change the bind ing information for a thin device without
moving any of the current extent allocations for the device. This is done
by a process called rebind ing.
Rebind ing a thin device increases the subscription level of the pool the
device is bound to, and decrease the subscription level of the pool it was
previously bound to. However, the allocation levels in both pools remain
unchanged .
Note: When rebind ing a device that is under FAST VP control, the thin pool
the device is being re-bound to must belong to one of the VP tiers contained
in the policy the device is associated with.
Thin pool oversubscription
Thin pool oversubscription allows an organization to p resent larger than
needed devices to hosts and applications without having to purchase
enough physical d rives to fully allocate all of the space represented by
the thin devices.
Note: For more information on oversubscription, refer to the Best Practices for
Fast, Simple Capacity Allocation with EMC Symmetrix Virtual Provisioning
technical note available at http :/ / powerlink.emc.com.
In a FAST VP configuration, the capacity available for thin devices is
now spread across multiple pools. Even if there is no plan to
oversubscribe the available capacity in the Symmetrix array , there
typically is insufficient space in a single tier to accommodate the
configured thin device capacity.
By following the previous recommend ation, bind ing all thin devices
being managed by FAST VP to a single thin pool inherently causes that
pool to be oversubscribed . As a result, the over-subscription limit for that
pool needs to be set to a value greater than 100 percent.
28
Planning and design considerations
FAST VP for EMC Symmetrix VMAX Theory and Best Prac tices for Planning and Performanc e
Available capacity
In order to determine at what level to set the over -subscription limit, the
total capacity available to d evices under the control of FAST VP should
first be calcu lated .
If the VP allocation by FAST policy feature is enabled , then 100 percent
of the capacity of each pool within the policy is available.
If the allocation by policy feature is not enabled , the following points
should be considered :
For pools with thin devices bound , 100 percent of the
capacity is available.
For pools with no thin devices bound , 100 percent of the
pool, less the capacity reserved by the PRC, is available.
Note: The pool reserved capacity value only affects FAST VP movements into
the pool; the PRC does not affect the ability for the thin devices to allocate
extents within the pool.
Over-subscription limits
After determining the capacity available to FAST VP, an over -
subscription limit can then be calculated for the pool devices that are
going to be bound to it.
To ensure the configured capacity of the array is not oversubscribed , the
limit can be calcu lated by d ivid ing the capacity of the thin pool being
used for bind ing into the available capacity of all the pools. This value is
then multiplied by 100 to get a percent value.
For example, consider a configuration with a 1TB EFD pool, a 5 TB FC
pool, and a 10TB SATA pool. The total available capacity is 16TB. If all
thin devices are bound to FC, the oversubscription limit could be set to
320% —(16/ 5)*100.
This value can be set higher if the intention is to initially oversubscribe
the configured physical capacity of the array, and then add storage on an
as-needed basis.
For thin pools where devices will never be bound , the subscription limit
can be set to 0 percent. This prevents any accidental bind ing or migration
to that pool.
RAID protection considerations
When designing a Virtual Provisioning configuration, and particu larly
when choosing a corresponding RAID protection strategy, both device-
29
Planning and design considerations
FAST VP Theory for EMC Symmetrix VMAX and Best Prac tic es for Planning and Performanc e
level performance and availability implications should be carefully
considered .
Note: For more information on these considerations, refer to the Best Practices
for Fast, Simple Capacity Allocation with EMC Symmetrix Virtual Provisioning
technical note available at http :/ / powerlink.emc.com.
FAST VP does not change these considerations and recommend ations
from a performance perspective. What FAST VP does change, however,
is that a single thin device can now have its d ata spread across multiple
tiers of varying RAID protection and drive technology within the array.
Because of this, the availability of an ind ividual thin device is not based
just on the availability characteristics of the thin pool the device is bound
to. Instead , availability is based on the characteristics of the tier with the
lowest availability.
While performance and availability requirements u ltimately determine
the configuration of each tier within the Symmetrix array, it is
recommended , as a best practice, to choose RAID 1 or RAID 5 protection
on EFDs. The faster rebuild times of EFDs provide higher availability for
these protection schemes on that tier.
Also, it is recommended to use either RAID 1 or RAID 6 on the SATA
tier. This is due to the slower rebuild times of the SATA drives
(compared to EFD and FC), and the increased chance of a dual-drive
failure, lead ing to data unavailability with RAID 5 protection.
For the FC tier, RAID 1 is the recommended protection level. Mirrored
data devices on FC pools provide a higher level of performance than
both RAID 5 and RAID 6, particu larly for write workload . Availability of
RAID 1, in regard to a dual-drive failure, is also greater than RAID 5. To
obtain the best availability numbers for RAID 1 on FC, the use of lower
capacity d rives is recommended .
Note: For new environments being configured in anticipation of
implementing FAST VP, or for existing environments having additional tiers
added , it is highly recommended that EMC representatives be engaged to
assist in determining the appropriate RAID protection schemes for each tier.
Drive configuration
For most customer configurations, the best practice configuration
guidelines recommend an even, balanced d istribution of physical d isks
across the d isk ad apters (DAs) and DA CPUs.
This is of particular relevance for EFDs, which are each able to support
thousands of I/ Os per second , and , therefore, are able to create a load on
30
Planning and design considerations
FAST VP for EMC Symmetrix VMAX Theory and Best Prac tices for Planning and Performanc e
the DA equivalent to that of multiple hard d isk drives.
An even d istribution of I/ O across all DA CPUs is op timal to maximize
their capability, and the overall capability of the Symmetrix array.
However, there are scenarios where it is not appropriate to evenly
d istribute d isk resources. An example of this is when the customer
desires partitioning and isolation of d isk resources to separate customer
environments and workloads within the VMAX Series array.
Generally, it is appropriate to configure more, smaller d rives, than fewer,
larger d rives for both EFD and FC tiers. Doing so spreads I/ O load as
wide as possible across the Symmetrix back-end .
For example, on a 2-engine VMAX Series array with 16 DAs, if the EFD
raw capacity requirement is 3.2 TB, it would be more optimal to
configure 16 x 200 GB EFDs than 8 x 400 GB EFDs, as each DA could
have one EFD configured on it.
Storage group priority
When a storage group is associated with a FAST policy, a priority value
must be assigned to the storage group. This priority value can be
between 1 and 3, with 1 being the highest priority . The default is 2.
When multiple storage groups are associated with FAST VP policies, the
priority value is used when the data contained in the storage groups is
competing for the same resources on one of the FAST VP tiers. Storage
groups with a higher priority are given preference when decid ing which
data needs to be moved to another tier.
The best practice recommendation is to use the default priority of 2 for
all storage groups associated with FAST VP policies. These values may
then be modified if it is seen, for example, that a high-priority
application is not getting sufficient resources from a higher tier.
SRDF
FAST VP has no restrictions in its ability to manage SRDF devices,
however, what must be considered is that data movements are restricted
to the array upon which FAST VP is operating. By default, there is no
coord ination of data movements; FAST VP acts independently on both
the local and remote arrays. However, enabling FAST VP SRDF
coord ination allows R1 performance metrics to be used when making
promotion-and-demotion d ecisions for d ata belonging to an R2 device.
Note: For more information on the operation of FAST VP SRDF coord ination ,
refer to the Advance FAST VP features section of the Implementing Fully
31
Planning and design considerations
FAST VP Theory for EMC Symmetrix VMAX and Best Prac tic es for Planning and Performanc e
Automated Storage Tiering for V irtual Pools (FAST VP) for EMC Symmetrix
VMAX Series Arrays technical note available at http:/ / powerlink.emc.com.
As a general best practice, FAST VP should be employed for both R1 and
R2 devices, particularly if the remote R2 array has also been configured
with tiered storage capacity. Where possible, similar FAST VP tiers and
policies should be configured at each site.
When available, FAST VP SRDF coord ination should be enabled for each
storage group containing SRDF devices that is associated with a FAST
VP policy.
Each SRDF configuration, however, presents its own unique behaviors
and workloads. Prior to implementing FAST VP in an SRDF
environment, the information in the follow ing sections should be
considered .
Note: The following sections assume that SRDF is implemented with all
Symmetrix arrays configured for Virtual Provisioning (all SRDF devices are
thin devices) and installed with the minimum Enginuity version capable of
running FAST VP (5875.135.91).
FAST VP behavior
For SRDF R1 devices, FAST VP promotes and demotes extent groups
based on the read -and-write activity experienced on th e R1 devices.
Meanwhile, by default, the SRDF R2 devices typically only experience
write activity during normal operations. As a result, FAST VP is likely to
promote only R2 device extents that are experiencing writes.
If there are R1 device extents that only experience read activity, no
writes, then the correspond ing extents on the R2 devices will see no I/ O
activity. This will likely lead to these R2 device extents being demoted to
the SATA tier, assuming this was included in the FAST VP policy.
SRDF coord ination changes this defau lt behavior by period ically
transmitting FAST VP performance metrics collected for R1 devices
across the SRDF link to the corresponding R2 devices. On the R2 device,
the R1 performance metrics are merged with the actual R2 metrics.
Movement decisions made by FAST VP for these R2 devices are then
based on the merged performance metrics. In this way, read activity on
the R1 device can be reflected on the R2.
SRDF operating mode
EMC best practices, for both synchronous and asynchronous modes of
SRDF operation, recommend implementing a balanced configuration on
both the R1 and R2 Symmetrix arrays. Ideally, data on each array would
32
Planning and design considerations
FAST VP for EMC Symmetrix VMAX Theory and Best Prac tices for Planning and Performanc e
be located on devices configured with the same RAID protection type, on
the same drive type.
As FAST VP operates independently on each array, and also promotes
and demotes d ata at the sub-LUN level, there is no guarantee that such a
balance may be maintained .
In SRDF synchronous (SRDF/ S) mode, host writes are transferred
synchronously from R1 to R2. These writes are only acknowledged by
the host when the data has been received into cache on the remote R2
array. These writes to cache are then destaged asynchronously to d isk on
the R2 array.
Note: For more information on SRDF/ S, see the EMC Solutions Enabler
Symmetrix SRDF Family CLI Product Guide available at
http:/ / powerlink.emc.com.
In an unbalanced configuration, where the R2 d ata resides on a lower -
performing tier than on the R1, performance impact may be seen at the
host if the number of write pendings builds up and writes to cache are
delayed on the R2 array.
With FAST VP, this typically does not cause a problem, as the
promotions that occur on the R2 side are the result of write activity.
Areas of the thin devices under heavy write workload are likely to be
promoted and maintained on the higher-performing tiers on the R2
array.
To avoid potential problems with the configuration becoming
unbalanced , it is recommended that FAST VP SRDF coord ination be
enabled in SRDF/ S environments.
In SRDF asynchronous (SRDF/ A) mode, host writes are transferred
asynchronously in pre-defined time periods or delta sets. At any given
time, there are three delta sets in effect: The capture set, the
transmit/ receive set, and the apply set.
Note: For more information on SRDF/ A, see the EMC Solutions Enabler
Symmetrix SRDF Family CLI Product Guide available at
http:/ / powerlink.emc.com.
A balanced SRDF configuration is more important for SRDF/ A, as data
cannot transition from one delta set to the next until the apply set has
completed destaging to d isk. If the data resides on a lower -performing
tier on the R2 array, compared to the R1, then the SRDF/ A cycle time
may elongate and eventually cause the SRDF/ A session to d rop.
33
Planning and design considerations
FAST VP Theory for EMC Symmetrix VMAX and Best Prac tic es for Planning and Performanc e
This is similar to SRDF/ S mode in most environments, so this may not
be a large issue, as the data under write workload is promoted and
maintained on the higher-performing tiers.
To avoid potential problems with the configuration becoming
unbalanced , it is strongly recommended that FAST VP SRDF
coord ination be enabled in SRDF/ A environments.
SRDF/ A DSE (delta set extension) should be considered to prevent
SRDF/ A sessions from dropping in situations where writes propagated
to the R2 array are being destaged to a lower tier, potentially causing an
elongation of the SRDF/ A cycle time.
Note: For more information on SRDF/ DSE, see the Best Practices for EMC
SRDF/A Delta Set Extension technical note available at
http:/ / powerlink.emc.com.
SRDF failover
As FAST VP works independently on both the R1 and R2 arrays, it
should be expected that the data layout will be d ifferent on each side if
SRDF coord ination is not enabled . When an SRDF failover operation is
performed, and host app lications are brought up on the R2 devices, it
should then also be expected that the performance characteristics on the
R2 will be d ifferent from those on the R1.
In this situation it takes FAST VP some period of time to ad just to the
change in workload and start promotion and demotion activities based
on the mixed read-and-write workload .
To better prepare for a failover, and to shorten the time period to achieve
an ad justment to the new workload , FAST VP SRDF coord ination should
be enabled .
Note: FAST VP performance metrics are only transmitted from R1 to R2.
When failed over, and with applications are running on R2 devices,
performance metrics will not be sent from R2 to R1.
Personality swap
An SRDF personality swap follows a similar pattern of behavior to an
SRDF failover scenario. Applications are brought up on the devices that
were formerly R2 devices.
In order for performance metrics to be transmitted from the new R1
devices to the corresponding R2 devices, SRDF coord ination need s to be
enabled on the storage groups containing the R1 devices.
SRDF coord ination can be enabled in advance on a storage group
containing R2 devices. However, the setting only takes effect when a
34
Summary and conclusion
FAST VP for EMC Symmetrix VMAX Theory and Best Prac tices for Planning and Performanc e
personality swap is performed, converting the R2 devices to R1.
SRDF bi-directional
Best practice recommend ations change slightly in a bi-d irectional SRDF
environment. In this case, each Symmetrix array has both R1 and R2
devices configured .
With SRDF coord ination enabled , it is possible that d ata belonging to an
R2 device could d isp lace d ata belonging to an R1 device on a higher tier .
This may happen if the R2 device’s corresponding R1 device has a higher
workload than the local R1 device.
In this scenario, it is recommended to reserve the EFD tier for R1 device
usage. This can be done by exclud ing any EFD tiers from the policies
associated with the R2 devices.
Should a failover be performed , then the EFD tier can be added
dynamically to the policy associated with the R2 devices. Alternatively,
the storage group containing the R2 devices could be reassociated with a
policy containing EFD.
Note: This recommendation assumes that the lower tiers are of sufficient
capacity and I/ O capability to handle the expected SRDF write workload on
the R2 devices.
EFD considerations
If there is a d ifference in the configuration of the EFD tier on the remote
array, then it is recommend ed not to include the EFD tier in the FAST VP
policies on the R2 array. Examples of a configuration d ifference include
either fewer EFDs or no EFDs. Similarly, if the EFD configuration on the
R2 array does not follow the best practice guideline of being balanced
across DAs, do not include the EFD tier on the R2 side.
Note: This recommendation assumes that the lower tiers are of sufficient
capacity and I/ O capability to handle the expected SRDF write workload on
the R2 devices.
Summary and conclusion
EMC Symmetrix VMAX Family with Enginu ity incorporates a scalable
fabric-interconnect design that allows a storage array to seamlessly grow
from an entry-level configu ration to a 4 PB system. Symmetrix VMAX
arrays provide pred ictable, self-optimizing performance and enables
organizations to scale ou t on demand in private cloud environments.
35
Summary and conclusion
FAST VP Theory for EMC Symmetrix VMAX and Best Prac tic es for Planning and Performanc e
Information infrastructure must continuously adapt to changing
business requirements. FAST VP automates tiered storage strategies, in
Virtual Provisioning environments, by easily moving workloads
between Symmetrix tiers as performance characteristics change over
time. FAST VP performs data movements, improving performance, and
reducing costs, all while maintaining vital service levels.
EMC Symmetrix VMAX FAST VP for Virtual Provisioning environments
automates the identification of data volumes for the purposes relocating
application d ata across d ifferent performance/ capacity tiers within an
array. FAST VP proactively monitors workloads at both the LUN and
sub-LUN level, in order to identify busy data that would benefit from
being moved to higher-performing drives. FAST VP also identifies less-
busy data that could be moved to higher -capacity d rives, without
existing performance being affected .
Promotion/ demotion activity is based on policies that associate a storage
group to multiple d rive technologies, or RAID protection schemes, by
way of thin storage pools, as well as the performance requirements of the
application contained within the storage group. Data movement
executed during this activity is performed nond isruptively, without
affecting business continuity and d ata availability.
There are two components of FAST VP: Symmetrix Enginu ity and the
FAST controller. Enginuity is the storage operating environment that
controls components w ithin the array. The FAST controller is a service
that runs on the service processor. FAST VP uses two d istinct algorithms,
one performance-oriented and one capacity-allocation-oriented , in order
to determine the appropriate tier a device should belong to. The
intelligent tiering algorithm considers the performance metrics of all thin
devices under FAST VP control, and determines the appropriate tier for
each extent group. The allocation compliance algorithm is used to
enforce the per-tier storage capacity usage limits.
Data movements execu ted by FAST VP are performed by the VLUN VP
data movement engine, and involve moving thin device extents be tween
thin pools w ithin the array. Extents are moved by means of a move
process only; extents are not swapped between pools.
Performance data for use by FAST VP is collected and maintained by
Enginuity. This d ata is then analyzed by the FAST controller , and the
guidelines generated for the placement of thin device data on the defined
VP tiers w ithin the array. When collecting performance data at the LUN
and sub-LUN level for use by FAST VP, Enginuity only collects statistics
related to Symmetrix back-end activity that is the result of host I/ O.
36
Summary and conclusion
FAST VP for EMC Symmetrix VMAX Theory and Best Prac tices for Planning and Performanc e
FAST VP provides a number of parameters that can be used to tune the
performance of FAST VP and to control the aggressiveness of the data
movements. These parameters can be used to nondisruptively ad just the
amount of tier storage that a given storage group is allowed to use, or to
ad just the manner in which storage groups using the same tier compete
with each other for space.
FAST VP is fully interoperable with all Symmetrix replication
technologies: EMC SRDF, EMC TimeFinder/ Clone, TimeFinder/ Snap,
and Open Replicator. Any active replication on a Symmetrix device
remains intact while data from that device is being moved . Similarly, all
incremental relationships are maintained for the moved or swapped
devices. FAST VP also operates alongside Symmetrix features, such as
Symmetrix Optimizer, Dynamic Cache Partitioning, and Auto-
provisioning Groups.
37
Appendix: Best practices quick reference
FAST VP Theory for EMC Symmetrix VMAX and Best Prac tic es for Planning and Performanc e
Appendix: Best practices quick reference
The following provides a qu ick reference to the general best practice
recommend ations for planning the implementation of a FAST VP
environment.
The best practices documented are based on features available in
Enginuity 5876.82.57, Solutions Enabler 7.4, Symmetrix Management
Console 7.3.3, and Unisphere for VMAX 1.0.
For more details on these recommend ations, and other considerations,
see ―Planning and design considerations.‖
FAST VP configuration parameters
FAST VP includes multiple configuration parameters that control its
behavior. The following sections describe best practice recommend ations
for each of these configuration parameters.
Performance time window
Use the default performance time window to collect performance metrics
24 hours a d ay, every day.
Data movement time window
Create a d ata movement window to allow data movement for the same
period of time that the performance time windows allow data collection.
Workload Analysis Period (WAP)
Use the default workload analysis period of 168 hours (7 days).
Initial Analysis Period (IAP)
At the initial deployment of FAST VP, set the initial analysis period to
168 hours to ensure, at a minimum, that a typ ical weekly workload cycle
is seen. After this period has passed , the value can be dropped to 24
hours.
FAST VP Relocation Rate
For the initial deployment of FAST VP, start with a conservative value
for the relocation rate, perhaps 7 or 8. At a later d ate the FRR can be
gradually lowered to a more aggressive level, possibly 2 or 3.
Pool Reserved Capacity (PRC)
For ind ividual pools with bound thin devices, set the PRC based on the
lowest allocation warning level for that thin pool.
For pools with no bound thin devices, set the PRC to 1 percent.
VP Allocation by FAST Policy
VP allocation by FAST policy should be enabled .
38
Appendix: Best practices quick reference
FAST VP for EMC Symmetrix VMAX Theory and Best Prac tices for Planning and Performanc e
FAST VP policy configuration
The ideal FAST VP policy would be 100 percent EFD, 100 percent FC,
and 100 percent SATA. While ideal, operationally it may not be
appropriate to deploy the 100/ 100/ 100 policy. There may be reasons to
limit access to a particular tier within the array.
The best way to determine the appropriate policies for a FAST VP
implementation is to examine the workload skew for the application
data to be managed by FAST VP. This can be done by using a tool such
as Tier Advisor.
Thin device binding
If no advance knowledge of the expected workload on newly written
data is available, the best p ractice recommendation is to bind all thin
devices to a pool within the second highest tier. In a three-tier
configuration, this would imply the FC tier.
VP allocation by FAST policy should be enabled .
It is not recommended to bind thin devices, under FAST VP control, to
the EFD tier.
A best practice recommend ation is to not preallocate thin devices
managed by FAST VP.
RAID protection considerations
For EFD, choose RAID 5 protection .
For FC, choose RAID 1 protection.
For SATA, choose RAID 6 protection.
Drive configuration
Where possible, balance physical d rives evenly across DAs. Configure
more, smaller EFDs, rather than fewer, larger EFDs, to spread I/ O load
as wide as possible.
Storage group priority
The best practice recommendation is to use the default priority of 2 for
all storage groups associated with FAST VP policies.
SRDF
As a general best practice, FAST VP should be employed for both R1 and
R2 devices, particularly if the remote R2 array has also been configured
with tiered storage capacity. Sim ilar FAST VP tiers and policies should
be configured at each site.
FAST VP SRDF coord ination should be enabled for storage gr oups
containing SRDF R1 devices.
39
Appendix: Best practices quick reference
FAST VP Theory for EMC Symmetrix VMAX and Best Prac tic es for Planning and Performanc e
Copyright © 2012 EMC Corporation. All Rights Reserved.
EMC believes the information in this publication is accurate as of its publication date. The
information is subject to change without notice.
THE INFORMATION IN THIS PUBLICATION IS PROVIDED "AS IS." EMC
CORPORATION MAKES NO REPRESENTATIONS OR WARRANTIES OF ANY KIND
WITH RESPECT TO THE INFORMATION IN THIS PUBLICATION, AND SPECIFICALLY
DISCLAIMS IMPLIED WARRANTIES OF MERCHANTABILITY OR FITNESS FOR A
PARTICULAR PURPOSE.
Use, copying, and d istribution of any EMC software described in this publication requires
an applicable software license.
For the most up-to-date listing of EMC product names, see EMC Corporation Trademarks
on EMC.com.
For the most up-to-date regulatory document for your product line, go to the Technical
Documentation and Advisories section on EMC Powerlink.
All other trademarks used herein are the property of their respective owners.