Вы находитесь на странице: 1из 29

89 Fifth Avenue, 7th Floor

New York, NY 10003


www.TheEdison.com
212.367.7400

White Paper

HP Thin Technologies
A Competitive Comparison
Printed in the United States of America
Copyright 2012 Edison Group, Inc. New York. Edison Group offers no warranty either expressed
or implied on the information contained herein and shall be held harmless for errors resulting
from its use.

All products are trademarks of their respective owners.


First Publication: September 2012
Produced by: Chris M Evans, Sr. Analyst; John Nicholson, Sr. Analyst; Barry Cohen, Editor-in-
Chief; Manny Frishberg, Editor
Table of Contents

Executive Summary ____________________________________________________ 1


Introduction __________________________________________________________ 2
Objective _________________________________________________________________ 2
Audience _________________________________________________________________ 2
Terminology ______________________________________________________________ 2
Overview _____________________________________________________________ 3
Thin Provisioning Overview ________________________________________________ 3
Thin Provisioning Drawbacks_______________________________________________ 4
HP 3PAR StoreServ Thin Technology ____________________________________ 5
Competitive Analysis __________________________________________________ 6
EMC VMAX ______________________________________________________________ 6
Background _____________________________________________________________________ 6
HP 3PAR StoreServ ComparisonStart Thin ________________________________________ 6
HP 3PAR StoreServ ComparisonGet Thin _________________________________________ 7
HP 3PAR StoreServ ComparisonStay Thin _________________________________________ 7
VMAX Restrictions _______________________________________________________________ 7

NetApp ___________________________________________________________________ 8
Background _____________________________________________________________________ 8
HP 3PAR StoreServ ComparisonStart Thin ________________________________________ 8
HP 3PAR StoreServ ComparisonGet Thin _________________________________________ 9
HP 3PAR StoreServ ComparisonStay Thin _________________________________________ 9

Hitachi VSP _______________________________________________________________ 9


Background _____________________________________________________________________ 9
HP 3PAR StoreServ ComparisonStart Thin _______________________________________ 10
HP 3PAR StoreServ ComparisonStay Thin ________________________________________ 10

EMC VNX _______________________________________________________________ 11


HP 3PAR StoreServ ComparisonStart Thin _______________________________________ 11
HP 3PAR StoreServ ComparisonGet Thin ________________________________________ 11
HP 3PAR StoreServ ComparisonStay Thin ________________________________________ 11
Performance Considerations ______________________________________________________ 12

IBM XIV _________________________________________________________________ 12


HP 3PAR StoreServ ComparisonStart Thin _______________________________________ 12
HP 3PAR StoreServ ComparisonGet Thin ________________________________________ 12
HP 3PAR StoreServ ComparisonStay Thin ________________________________________ 13
Dell Compellent __________________________________________________________ 13
HP 3PAR StoreServ ComparisonStart Thin _______________________________________ 13
HP 3PAR StoreServ ComparisonGet Thin ________________________________________ 14
HP 3PAR StoreServ ComparisonStay Thin ________________________________________ 14

Testing Overview and Methodology ____________________________________ 15


Test 1Zero-Page-Reclaim Performance ____________________________________ 15
Test 2Large Pre-Allocation _______________________________________________ 16

Test Results __________________________________________________________ 17


Test 1Zero-Page-Reclaim Performance ____________________________________ 17
Test 2Large Pre-Allocation _______________________________________________ 20
Conclusions and Recommendations _____________________________________ 22
Best Practices _____________________________________________________________ 22

Appendix ADocument References _____________________________________ 24


Appendix BTest Equipment Specification ______________________________ 25
Arrays ___________________________________________________________________ 25
Servers __________________________________________________________________ 25
Executive Summary

As the drive to "do more with less" becomes a mantra for many organizations,
optimizing space utilization is a key goal of many IT departments. Storage continues to
be one of the major cost components of today's infrastructure deployments. Thin
technology, including thin provisioning, offers efficiency benefits that can significantly
reduce both capital and operational costs. However implementations of thin
technologies differ with the storage vendors.

HP 3PAR StoreServ is seen as a leader in thin technology, with three key aims:
1. Start Thinensure thin provisioned storage occurs with minimum overhead.
2. Get Thinensure data moved to HP 3PAR StoreServ remains thin on migration.
3. Stay Thinensure data is kept at optimal efficiency over time.

To validate this statement, a literature review, extensive customer interviews, and two
tests were performed:
1. Zero-Page-Reclaim Performancevalidation of the ability to reclaim freed resources
as part of normal operations.
2. Large Pre-allocationtest of the ability to create new storage volumes with minimal
overhead.

Overall, HP 3PAR StoreServ was the best performer in achieving the goals of "start
thin," "get thin," and "stay thin."

Edison: HP Thin Technologies Comparison Page 1


Introduction

Objective

This report looks at thin provisioning technology from the major storage vendors in
today's marketplace. It compares the thin implementations from seven storage array
platforms, including Hewlett Packard's 3PAR storage arrays. In particular this white
paper highlights three important differentiating aspects of HP 3PAR StoreServ's thin
technology:
1. The ability to "start thin"provisioned storage is thin at deployment time.
2. Getting thinthe ability to move data from thick to thin.
3. Staying thinmaintaining thin LUNs.

Audience

Decision makers in organizations that are considering implementing a thin technology


strategy will find this paper provides high level information on vendor offerings.
Technical professionals looking to understand more about the implementation of vendor
thin technology solutions will also find the content of this paper useful.

Terminology

This white paper makes reference to the following common terminology:


"Thick" LUNa storage volume presented from an array in which all of the space
representing the logical size of the LUN (logical unit number) is reserved on the
array for exclusive use by that volume.
"Thin" LUNa storage volume presented from an array that is not tied to any
physical storage allocation and in which only the physically written space is
consumed on the array.
Thin technologiesa suite of features, including thin provisioning, that optimize the
use of a storage array.

Edison: HP Thin Technologies Comparison Page 2


Overview

In recent years storage has become one of the major cost components within the data
center. Although the price of storage continues to fall, the rate of data growth in many
organizations continues to rise steeply, resulting in increasing costs for managing the
storage systems. Every year there is a requirement to "do more with less," using storage
more efficiently without increasing the operational budget. There are a number of key
initiatives being undertaken by organizations in order to reduce their storage
consumption. These relate directly to the use of thin technology.

Reducing WasteStorage utilization never reaches 100 percent of the physical space
provisioned from an array, as each level of configurationfrom the array to the
hostintroduces some inefficiency. Reducing waste increases utilization and allows
the deferral of additional capital expenditure.
Reducing OverheadDeploying storage isnt a quick task; from purchase order to
deployment on the data center floor, the process can take months to achieve. Storage
administrators usually keep storage in reserve in order to manage the purchase
process.
FlexibilityEnd users want the minimum disruption to their applications and as a
result, many over-order storage resources, in many cases up to 36 months ahead of
when the space is actually needed. Ideally, end users should be able to lay out their
storage needs based on growth plans and then allocate that storage on-demand.
Improving Cost EfficiencyStorage Tiering (placing data on the most cost-effective
media for the I/O profile required) is a key technology in reducing storage costs.
Dynamic tiering can be used to automate the process of data placement, based on the
use of storage pools for LUN creation. Thin provisioned LUNs directly aid the
deployment of a tiered storage model. A thin LUN is built from blocks of physical
disk capacity from within a pool of storage with metadata to associate the logical
LUN to the physical space. The physical blocks can therefore be taken from multiple
pools, where each pool represents a different tier.

Thin Provisioning Overview

Thin provisioning is a space reduction technology implemented by all of the major


enterprise storage vendors. It enables the utilization of storage within an array to be
increased over traditional "thick" storage deployments.

Edison: HP Thin Technologies Comparison Page 3


In traditional "thick" storage deployments, physical space on disk is immediately
reserved for the entire size of a volume (or "LUN") at creation time, regardless of how
much space will subsequently be used by the host. In thin storage deployments, no
space is reserved in advance for the LUN. As the host writes data to the LUN, physical
space is assigned on-demand from the array, usually in blocks that vary from 4KB to
42MB, according to the vendor. Thin provisioned LUNs are therefore much more
efficient and more closely track the actual space in use on the host.

For many reasons, storage utilization on hosts never reaches 100 percent utilization.
However with "thick" LUNs, physical space is reserved out on an array for the entire
size of a volume. Thin provisioned deployments can take advantage of all physical
storage available by creating more logical storage capacity than is physically available.
So called "over-provisioning" enables the utilization of physical space to be pushed to
levels higher than can be achieved in normal deployments.

Thin Provisioning Drawbacks

The ability to over-provision storage does come with a few drawbacks. Should physical
space be completely exhausted, hosts will receive write errors, indicating a physical
problem on the array. Write failures are not usually handled gracefully by the host
operating system and can lead to system crashes. Therefore, physical versus logical
space capacity needs to be carefully managed.

Over time as files are created and deleted, thin LUNs become less efficient. This is due to
the way in which the file system on the LUN manages file allocations, free space and
metadata. Some file system implementations are more "thin friendly" than others and
are designed to re-use released space. However, housekeeping of thin provisioned
storage is required in order to maintain optimal levels of efficiency. Storage vendors
have introduced features that enable the array to recover unused storage resources:
Zero-Page-Reclaim (ZPR)A storage array identifies an entire block of storage
consisting of binary zeros, the block will be assumed to be unused and is released
back to the free pool. The ability to find unused blocks depends on a number of
factors, including the file system and array block-size and the level of file
fragmentation. Smaller array block-sizes are better for ZPR operations.
SCSI UNMAPThe UNMAP command is a low-level I/O operation that can be
used by the host to signal to the storage array that a block of storage is no longer in
use and can be released to the free pool. Unfortunately very few operating systems
currently support this feature.

Edison: HP Thin Technologies Comparison Page 4


HP 3PAR StoreServ Thin Technology

The HP 3PAR StoreServ storage platform has a "thin by design" architecture that places
no restrictions on the use of either thin or thick LUNs. This applies to LUN performance,
and capacity, removing the need for the storage administrator to design the layout of the
array to cater for thin technology. HP 3PAR StoreServ is specifically optimized for thin
provisioning and contains many unique design features that enable "starting thin,"
"getting thin," and "staying thin."
RAIDHP 3PAR StoreServ arrays offer a unique RAID technology that provides
chassis high availability, and divides physical disks into "chunklets" of either 256MB
or 1GB in size. Chunklets are combined to form Logical Volumes (LVs) and
Common Provisioning Groups (CPGs) from which Virtual Volumes are created.
Thin Provisioning Virtual Volumes use a block size increment of 16KB, which is the
minimum reclaimable unit of storage within the array.
Hardware ASICNow at Generation Four, HP 3PAR StoreServ uses dedicated
custom ASIC (Application Specific Integrated Circuit) processors to perform the
identification and recovery of unused resources that can be reclaimed from thin
provisioned virtual volumes. An ASIC enables processor-intensive tasks to be
offloaded to dedicated hardware, removing the performance impact of features such
as space reclamation from the array and ensures consistent host I/O response times.
The HP 3PAR StoreServ ASIC provides a range of functions, including inline ZPR.
Thin PersistenceAn operating system task that identifies and recovers freed
resources.
Thin ConversionPerforms the migration of thick to thin volumes through a
process of inline zero detection. As data is copied to the array, zeroed blocks of data
are identified and logically mapped rather than physically written to disk.
Thin Copy ReclamationPerforms space recovery on volume copies within the
array.
Thin Reclamation APIHP 3PAR StoreServ developed the Thin Reclamation API
in partnership with Symantec. This feature allows the file system to signal when
freed resources can be released on the array. It is supported by Symantec Veritas
Storage Foundation from Version 5 onwards.
Virtualization SupportHP 3PAR StoreServ supports the VMware VAAI API,
including the "block zeroing" command.
ManagementHP 3PAR StoreServ arrays provide alerts for specific thin
provisioning space issues. Alerts are issued based on pre-defined thresholds and
enable efficient monitoring of capacity in thin environments.

Edison: HP Thin Technologies Comparison Page 5


Competitive Analysis

EMC VMAX

Symmetrix VMAX, EMC's flagship enterprise storage platform, is the first enterprise-
class storage product to move away from custom hardware design. It uses commodity-
based Intel processors with customized hardware managing the interconnect between
storage modules. The VMAX operating systemEnginuityis an evolution of the code
developed for the first Symmetrix ICDA (Integrated Cache Disk Arrays) in 1991, and it
still retains many of the original architectural design features and constraints. The
discussion of VMAX in this section covers the latest 10K, 20K and 40K models.

Background
Thin provisioning in VMAX is implemented as a feature called Virtual Provisioning
EMC's brand name for their thin provisioning technology. Thin provisioned LUNs are
known as thin devices and take physical storage from thin pools. A thin pool is created
using standard "thick" LUNs (termed data devices), which are subdivided into allocation
units called thin device extents. Thin pools must use the same emulation and RAID
protection type and EMC recommend building them from disks of the same rotational
speed and data device size.

A thin extent is 12 tracks or 768KB in size and represents both the initial minimum
assigned to all thin devices when they are bound to a thin pool and also the lowest
increment of granularity when the capacity of a thin device is extended. HP 3PAR
StoreServ thin technology uses the much smaller increment of 16KB, which results in
less wastage, particularly with fragmented and thin-hostile file systems. Thin devices are
effectively cache-based objects that simply reference the underlying physical pool of
standard LUNs in array. These LUNs in turn, map to physical disks. A single VMAX
system supports up to 512 pools and 64,000 thin devices.

HP 3PAR StoreServ ComparisonStart Thin


VMAX Virtual Provisioning requires a significant amount of initial planning. EMC
recommends the use of large data devices within pools. As data devices are effectively
standard LUNs, typical configurations create LUNs to be used purely for thin pools and
LUNs to be used for non-thin provisioned usage. Meta devices (a linked series of
multiple standard devices) cannot be used as data devices. This practice leads to waste
and a shortage of the right type of storage. It is possible to dissolve and resize standard
LUNs, however the process is time consuming and can lead to unbalanced performance.

Edison: HP Thin Technologies Comparison Page 6


With HP 3PAR StoreServ thin technology, physical disks are simply assigned to a pool
from which either thin or thick LUNs can be provisioned.

When VMAX thin devices are bound to a thin storage pool, a minimum allocation of one
thin extent (768KB) is reserved. As thin devices are effectively cache objects, each device
consumed an additional 148KB of cache, plus 8KB per 1GB based on the size of the thin
device. With HP 3PAR StoreServ thin technology, no initial space reservations are made.

HP 3PAR StoreServ ComparisonGet Thin


There is no functionality within the VMAX array to optimize storage when converting
thick LUNs to thin devices. As thick LUNs are copied to a thin device, the space
occupied by the thin device is 100 percent of the logical allocation. Thin devices must be
optimized, using a process called Space Reclamation. This means data migrations
moving thick to thin LUNs require additional physical capacity to be available for the
migration process. HP 3PAR StoreServ thin technology uses the Thin Conversion feature
to optimize the migration of thick to thin LUNs in real time at line speeds.

HP 3PAR StoreServ ComparisonStay Thin


EMC VMAX arrays are able to reclaim "empty" or zero pages of a thin device using
Space Reclamation. This runs as a background task on the Disk Adapter associated with
the LUN. An entire thin extent is read into cache and examined, checking the T10-DIF
values for each block of data against a known T10-DIF value for an all-zeros block. If the
entire thin extent contains only zeroed data, then it is released from use. Until the space
reclamation process is run, any zeroed areas of a thin device still consume physical
space. VMAX space reclamation cannot be used with thin devices that are in an active
SRDF (Symmetrix Remote Data Facility) pair or are using local replication. With 3PAR
StoreServ thin technology, the process of zero-detecting is done inline using a custom
ASIC. This has no impact on the controller processor or cache utilization levels and
occurs in real time at line-speed. There are no restrictions on replicated LUNs.

VMAX Restrictions
EMC recommends a utilization level of between 60-80 percent per thin pool in order to
prevent "out of space" issues. With multiple pools (which are required for different
RAID data protection types) this can result in significant waste. HP 3PAR StoreServ thin
technology does not require separate pools for multiple protection types. When using
Synchronous SRDF with VMAX, only one active write is permitted per thin device.
Where thin devices are created into meta-devices, this can result in a performance
impact. There is also a limit of eight read-requests-per-paths for each thin device, which
can result in slow performance with high read miss rates.

Edison: HP Thin Technologies Comparison Page 7


NetApp

NetApp storage appliances were originally developed to deliver network-attached


storage using either the CIFS or NFS protocols. Over time, NetApp have developed their
platform to cater for block storage, using either iSCSI or Fibre Channel. The current
versions of NetApp filers can be configured in either 7-mode or cluster-mode and
represent two distinct product lines based on the original Data ONTAP operating
system, and the codebase from the acquisition of Spinnaker, Inc., respectively.

Background
NetApp filers implement block-based storage within Data ONTAP by emulating LUNs
within volumes known as FlexVols. FlexVols are then created on aggregates (pools of
physical storage) and physical disk RAID groups. The underlying architecture uses a
data layout called WAFL (Write Anywhere File Layout) that operates a "write-new"
policy for both new data and updates; no block or file data is ever updated in place.
WAFL uses a page size of 4KB, storing updates in non-volatile RAM before writing an
entire "stripe" of data to disk. In this way, writes are optimized on commit-to-disk using
a RAID-4 physical disk configuration. NetApp LUNs are emulated through files on
volumes; therefore, both block and file data can be mixed within the same storage pool.
LUN creation is a simple process to achieve, however the use of block-based LUNs
involves significant complexity.

HP 3PAR StoreServ ComparisonStart Thin


All NetApp volumes by default are "thick" provisioned with thin volumes simply
having a no space guarantees. In turn, LUNs within a volume are thin provisioned if
they have space reservation disabled. By default, all LUNs are thick provisioned and
have space reservation enabled. The administrator must turn off space reservation after
the LUN is created to make it thin-provisioned. However, as blocks of data are not
overwritten in place, an additional amount of space (called Fractional Reserve or
Overwrite Reserved Space) must be reserved to manage data updates where snapshots
are used on the LUN. By default Fractional Reserve is set at 100 percent when the space
guarantee is set to "file" for a volume. This is the only way to guarantee enough space is
available within the volume to hold updates to the entire contents of the LUN. Space
guarantees are complex and if used incorrectly can result in LUNs going offline in order
to protect data. The system creates a propensity to over-configure storage in order to
reduce the risk of data access issues. HP 3PAR StoreServ thin technology is implemented
in a much simpler way and does not have the management complexity seen in the
NetApp platform.

Edison: HP Thin Technologies Comparison Page 8


HP 3PAR StoreServ ComparisonGet Thin
NetApp has no native features for importing LUNs from other storage platforms. LUNs
migrated into the NetApp platform using host-based tools allocate physical space
matching the entire logical size of the LUN. There are no native features within Data
ONTAP to identify and reclaim zero or empty pages of data. Data ONTAP does
implement data deduplication (called dedupe) at the block level, and it is possible to use
this feature to deduplicate zeroed pages of data. However there are restrictions on the
size of LUNs that have dedupe enabled. In addition, deduping can have negative
performance impacts on highly utilized LUNs. NetApp quote tests that show the
performance impact for writes on deduplicated volumes becomes worse with larger
systems; for example the FAS6080 can have a performance degradation of up to 35
percent. HP 3PAR StoreServ thin technology has no performance impact.

HP 3PAR StoreServ ComparisonStay Thin


Over time, as data is written to a NetApp block-LUN, the physical space used trends
towards the logical LUN size. There are no in-built features for enabling zero-block
identification and reclaiming. Instead, NetApp requires the deployment of a host-based
agent called SnapDrive to track updates to the file system. The agent must be deployed
on every server to which NetApp LUNs are presented; otherwise the tracking of file
deletions cannot occur. Platform support for SnapDrive is limited and does not include
common operating systems such as RHEL6. HP 3PAR StoreServ thin technology
automatically detects zero-block data inline with no performance impact and does not
require the implementation of host agents.

Hitachi VSP

The VSP is Hitachi's current enterprise-level storage array and is the evolution of
previous Lightning and USP-V models. The VSP retains the use of custom ASIC
technology, in which the management of storage processes is handled by Virtual Storage
Directors connected to the back-end switch matrix. Custom ASIC usage has been a
feature of all of the Hitachi storage platforms; however it isn't used directly in the thin
provisioning approach or in managing the efficiency of thin provisioned storage.

Background
VSP thin provisioning technology is known as HDP--Hitachi Dynamic Provisioning.
HDP thin LUNs (called LDEVs or logical devices) are created from a HDP pool that
comprises standard LDEV devices. In turn, LDEVs are created from RAID groups, built
from up to 16 disks in one of seven RAID-5 or RAID-6 variations. At the physical level,
data is written in tracks of 256KB per physical disk, which results in a standard logical

Edison: HP Thin Technologies Comparison Page 9


page size of 42MB, in order to accommodate all possible RAID levels. This means initial
volume allocations and volume expansion of thin LUNs is in 42MB-page increments that
can result in inefficient use of space with small file block size and thin-unfriendly file
systems. With HP 3PAR StoreServ thin technology, space allocations are made in 16KB
increments, which results in much less wastage in thin-unfriendly environments.

HP 3PAR StoreServ ComparisonStart Thin


The creation of VSP HDP pools requires considerable planning. Hitachi recommends
that pools be created using large LUNs built from traditional RAID groups. RAID group
creation is typically performed at array installation time and so is a one-off task. The
RAID group size depends on how many physical disks and back-end directors have
been installed within the VSP. However it is normal to build HDP pools from many
RAID groups across all back-end directors to ensure maximum I/O performanceso
called wide striping. Although standard "thick" LDEVs can be created from the same
RAID groups used to create HDP pools, for performance reasons, the practice isn't
recommended. The implementation of HDP can result in wasted resources and always
requires the reservation of many RAID groups to thin provisioning. By contrast, with
HP 3PAR StoreServ thin technology physical disks are simply assigned to a pool from
which either thin or thick LUNs can be provisioned.

On VSP, thin LDEVs allocate a minimum of one 42MB page on assignment to an HDP
pool. With HP 3PAR StoreServ thin technology, no initial space reservations are made.

HP 3PAR StoreServ ComparisonStay Thin


The Hitachi VSP platform enables traditional "thick" LUNs to be imported into the
system using the external virtualization feature of the array, known as Universal
Volume Manager. In addition, "thick" LUNs can be imported from other VSP systems
using TrueCopy replication. Any imported LUNs remain fully allocated at their original
logical size until zero-block reclaim is performed using the zero-page-reclaim feature.
ZPR is a post-processing background task that examines individual LDEVs and releases
42MB pages back to the HDP pool. Hitachi recommends ZPR be executed during
periods of inactivity, as the task is performed by the back-end directors and can have a
performance impact on production I/O. As ZPR is not performed in-line, thin LDEVs
will "grow" over time as data is written to the file system on the LUN. This means thin
pools need to be provisioned with additional capacity to cater for this growth between
ZPR reclaim tasks. With 3PAR StoreServ thin technology, the process of zero-detection is
done inline using a custom ASIC. This has no impact on the controller processor and
occurs in real time at line-speed.

Edison: HP Thin Technologies Comparison Page 10


EMC VNX

EMC's VNX platform is an evolution of the previous CLARiiON and Celerra products
(serving block and file protocols respectively). The two platforms were brought together
and marketed as a single platform, using one management tool, called Unisphere. Block-
based storage LUNs are presented from the base hardware unit, with file access
implemented on x-blade modules. Thin provisioning technology is implemented using
the Virtual Provisioning (VP) feature. VP extends the capabilities of LUN configuration
to include both thick and thin LUNs on the same disk pool. VP disk pools can be
comprised of large numbers of disks (greater than the standard disk pool which is
limited to 16 devices), but still configures disks in RAID groups for resiliency.

HP 3PAR StoreServ ComparisonStart Thin


VNX thin LUNs can be logically defined in sizes from 1MB to 16TB. However each LUN
reserves a minimum of 3GB. This means allocation of a large number of LUNs has a
significant reservation on physical space, in contrast to HP 3PAR StoreServ thin
technologys no minimum physical spare reservation.

HP 3PAR StoreServ ComparisonGet Thin


VNX does not support any native inline zero-page reclaim functionality. Instead,
"empty" pages must be reclaimed by using either the LUN Migration (within the array)
or SAN Copy (from external arrays) functions. This means that thin LUN physical
capacity will trend towards their logical size over time. EMC recommends using the
sdelete host command and LUN migration as the method of reclaiming unused space in
the VNX array. By comparison, HP 3PAR StoreServ thin technology detects and
eliminates zero-page data inline with no need to perform additional manual data
migrations. VNX supports the Symantec Thin Reclamation API, however this requires
the deployment of the Veritas File System on every server for which reclaim is required.

HP 3PAR StoreServ ComparisonStay Thin


VNX thin LUNs grow in increments of 1GB of physical space, with 8KB blocks used
as the minimum level of granularity. The 8KB block refers to the space within which
zero page data can be reclaimed, however each increment of space assigned to a LUN
works in steps of 1GB, so any 8KB holes "punched" out of a 1GB slice can only be reused
by that volume, rather than as free space for all available volumes. HP 3PAR StoreServ
thin technology implements LUN mapping to internal logical disks at 32MB pages, with
16KB representing the minimum level of page granularity for thin provisioning. This
makes the 3PAR system much more efficient with thin LUNs. VNX has no ASIC
technology to perform inline zero space reclamation, unlike HP 3PAR StoreServ.

Edison: HP Thin Technologies Comparison Page 11


Performance Considerations
EMC highlight that Virtual Provisioning Thin LUNs do provide more flexibility, but
offer lower performance than traditional thick LUNs and so recommend their use only
for applications requiring "moderate" performance. HP 3PAR StoreServ thin technology
has no performance restrictions.

IBM XIV

The IBM XIV storage array platform was acquired from an Israeli startup, founded by
the inventor of the EMC Symmetrix, Moshe Yanai. The platform takes a radical
departure from traditional arrays and uses only high-capacity SATA or SAS hard drives,
although the configuration has recently been expanded to accelerate I/O using an SSD
cache layer. XIV is now at the third generation of hardware, utilizing either 2TB or 3TB
drives, with 6TB of SSD cache. Each array is comprised of between six and 15 server
nodes, which hold 12 hard drives each, resulting in a maximum configuration of 180
drives. Each node subdivides disks into 1MB chunks, which are then distributed across
all disks as a single large pool of mirrored data. XIV uses the terms "soft size" and "hard
size" to refer to the logical and physical size of a LUN respectively. These terms also
apply equally to pools that can be allocated physical capacity. It is possible for a pool to
deplete hard (physical capacity) and lock access to a volume, despite there being free
physical space in other pools. The overall capacity of an XIV array is referred to as the
"system hard size." A "system soft size"the degree of over-provisioning permitted at
the array levelis also defined, but can't be modified by the system administrator. This
value has to be set by an IBM engineer and requires the customer to indemnify IBM
against any issues that occur as a result of the change.

HP 3PAR StoreServ ComparisonStart Thin


By default, all XIV LUNs are thin provisioned and placed into storage pools. As data is
spread equally across all drives in the system, storage pools provide no more than a
logical administrative benefit for thin provisioning; no workload segregation is possible.
Due to the architectural design of the system, all LUNs reserve an initial 17GB at
creation time, which can result in a significant initial waste of space. In addition, all
LUNs are incremented in 17GB chunks. HP 3PAR StoreServ thin technology reserves no
space on initial LUN creation, and uses increments in 16KB, making it highly efficient.

HP 3PAR StoreServ ComparisonGet Thin


XIV can support the thick-to-thin conversion of volumes as part of the Data Migration
feature. Volumes imported from other systems do have zero-pages identified and
removed during the process. However, Data Migration requires configuration changes

Edison: HP Thin Technologies Comparison Page 12


and the placement of the XIV "inline" with the host and the original volume. This task
requires an outage to achieve. With HP 3PAR StoreServ thin technology, data can be
migrated into the array via the host, identifying zero-pages in line and without requiring
a host outage.

HP 3PAR StoreServ ComparisonStay Thin


XIV implements a zero-page-reclaim feature to identify and eliminate "empty" blocks of
data. However, the feature is implemented as part of the background data "scrubbing"
routine that trawls the array and looks for data integrity and parity issues. Reclamation
of ZPR data can take many days (or as long as three weeks) to complete, so there is an
over-allocation of physical space until zero pages can be identified and recovered. HP
3PAR StoreServ thin technology implements active ZPR detection inline at the time of
data write, ensuring "empty" pages are immediately identified and eliminated before
data is written to physical disk.

XIV supports the Symantec Thin Reclamation API for instant space reclamation,
enabling hosts running Symantec Storage Foundation, version 5 and above, to directly
signal to the array when storage is released. This requires the deployment of the Veritas
File System on each host connected to the array. Instant space reclamation does not
support mirrored volumes, volumes that have snapshots or snapshots themselves,
making the recovery process limited.

Dell Compellent

Compellent Technologies, Inc. was founded in 2002 and subsequently acquired by Dell
in 2011, from which time it was marketed under the Dell Compellent brand name. Based
on commodity components, Compellents unique offering is called Data Progression, an
automated tiered storage feature enabling migration of data between storage tiers at the
block level.

Compellents thin provisioning technology is known as Dynamic Capacity. A thin LUN


is allocated using 2MB blocks that can be assigned from any of the physical capacity
within the array. Although block size defaults to 2MB, the administrator may override
this value to either 512KB or 4MB.

HP 3PAR StoreServ ComparisonStart Thin


All Dell Compellent LUNs are allocated as thin provisioned LUNs with a minimum
allocation of at least 2MB, depending on the protection methodology (for example,
RAID-10 would allocate two 2MB blocks). By comparison, HP 3PAR StoreServ reserves

Edison: HP Thin Technologies Comparison Page 13


no minimum allocation. Dell Compellent can use 512KB or 4MB block sizes; however,
Dell recommends not mixing block sizes in a single system, as this can result in waste of
physical space.

HP 3PAR StoreServ ComparisonGet Thin


Dell Compellent supports the ability to migrate data into the array and remove unused
space using the Thin Import feature. HP 3PAR StoreServ also supports thin import;
however, unlike Compellent, the process is performed inline using a dedicated ASIC
rather than in software; so it has no impact on array performance.

HP 3PAR StoreServ ComparisonStay Thin


Dell Compellent supports space recovery on existing volumes using their Free Space
Recovery tool that uses the SCSI UNMAP command; however, the feature requires the
deployment of the tool on each server on which recovery is to be performed. By contrast,
HP 3PAR StoreServ supports inline recovery of zero-page data, inline dynamically with
no host agents.

Edison: HP Thin Technologies Comparison Page 14


Testing Overview and Methodology

The aim of performing vendor comparison tests is to show how HP 3PAR StoreServ
compares to other vendors in terms of performance and efficiency. Although the
implementations from each vendor appear to offer similar features, the implementations
differ greatly in their performance and efficiency. The following tests were performed in
the competitive summary list.

The storage systems tested were not directly comparable for performance targets or
specifications. They varied in the number, size and types of drives, the number of
controllers and other physical specifications. Therefore, the data generated in Edison's
zero-page-reclaim performance test should only be compared for the differences for each
tested array from the test baseline. The exception is the Large Pre-Allocation results,
which demonstrate the effects of the different thin provisioning and storage
architectures on capacity utilization, rather than a change in performance for the
systems.

Details of the hardware tested can be found in the Appendix, at the end of this
document.

Test 1Zero-Page-Reclaim Performance

This test aims to show the impact of zero-page-reclaim functionality on each array. The
reclaim function is an essential property of "stay thin," ensuring that ongoing allocations
don't turn thin volumes into thick ones over time. Ideally this test should not impact I/O
performance. The test process performed the following steps:
1. Create a single large "thick" 200GB LUN and assign to a Windows host.
2. Quick format the LUN with the NTFS file system.
3. Perform load test with IOMETER, writing binary zeros to the LUN, recording IOPS
and latency figures.
4. Repeat the test with a 200GB thin LUN.

Prior to the test, the zero-page-reclaim task was enabled on the VSP system. For the
EMC platforms, the zero-page-reclaim feature was enabled by performing a LUN
migration, the method recommended by EMC.

Edison: HP Thin Technologies Comparison Page 15


Test 2Large Pre-Allocation

This test aims to show the overhead at the initial creation of thin LUNs and addresses
the requirement to "start thin." Ideally the creation of thin LUNs should reserve the
minimum amount of storage possible on the array. The test process performed the
following steps:
1. Create five 200GB thin LUNs and assign to a Windows host.
2. Quick format the LUNs with the NTFS file system.
3. Measure the amount of space consumed as indicated by the array.

Edison: HP Thin Technologies Comparison Page 16


Test Results

Test 1Zero-Page-Reclaim Performance

The data in these tables represents the performance for each array capable of zero-page-
reclaim within the test parameters. NetApp FAS was not included in this test, as the
system has no native support for zero-page-reclaim. Data for IBM XIV was not included,
because that system performs reclaim over a very long period of time that was outside
the test parameters.1

Platform Degradation from Baseline (%)

EMC VMAX 48.19%

Dell Compellent 42.33%

EMC VNX 29.99%

Hitachi VSP 23.34%

HP 3PAR 0.00%

Table 1 - Test 1 - IOPS Performance during ZPR

1According to an IBM Redbook, IBM XIV Storage System: Copy Services and Migration, it,
"could take up to three weeks for used space value to decrease This is because recovery of
empty space runs as a background task."(Page 264). Not only is the time required for ZPR outside
the parameters of our research, enabling over-provisioning is, "not within the scope of the
administrator role."(Page 30) This suggests that an IBM engineer must perform an
overprovisioned configuration.

Edison: HP Thin Technologies Comparison Page 17


Figure 1 - Test 1 - IOPS Performance during ZPR

Platform Degradation from Baseline (%)

EMC VNX 42.61%

EMC VMAX 40.58%

Dell Compellent 74.22%

Hitachi VSP 30.33%

HP 3PAR 1.23%

Table 1 - Test 1 - I/O Latency during ZPR

Edison: HP Thin Technologies Comparison Page 18


Figure 2 - Test 1 - I/O Latency during ZPR

The results of this test show that ZPR activity has an impact on both the latency and
throughput of each platform except HP 3PAR StoreServ. The greatest effect was seen on
EMC VMAX performance and the latency increase with Dell Compellent.

Edison was able to determine that performance impact shown on EMC VMAX was
because the platform needs to read each thin device extent into cache in order to
perform ZPR processing. This cache load clearly has a direct impact on array
performance.

Edison was unable to diagnose the causes of the increase latency on the Dell Compellent
system.

The HP 3PAR StoreServ array has dedicated ASICs to handle the ZPR workload without
impacting on delivering I/O to hosts.

Edison: HP Thin Technologies Comparison Page 19


Test 2Large Pre-Allocation

The results from this test are shown in the following table and graph.

Platform Space Allocated (MB)

EMC VMAX 17

NetApp FAS 368

Dell Compellent 800

Hitachi VSP 1,230

HP 3PAR 3,125

EMC VNX 20,039

IBM XIV 86,000

Table 2 - Test 2 - Large Pre-Allocation

Figure 3 - Test 2 - Large Pre-Allocation

Edison: HP Thin Technologies Comparison Page 20


The results of this test show EMC VNX and IBM XIV performed poorly in pre-
allocations. EMC VNX reserves a minimum of 3GB per LUN; IBM XIV reserves a
minimum of 17GB per LUN. The other platforms performed well. Clearly when systems
have large volumes of LUNs, the minimum reserve can have a detrimental impact on
the aims of "starting thin," resulting in large amounts of unusable storage.

Edison: HP Thin Technologies Comparison Page 21


Conclusions and Recommendations

Thin provisioning is a great space optimization feature that can be used to increase
levels of utilization on storage arrays. The direct benefit is reducing capital expenditure
on hardware and operational expenditure on management. However as we have seen
from the tests, not all thin provisioning implementations are equal in terms of their
ability to optimize space with minimal impact on performance.

HP 3PAR StoreServ systems adhere to three basic principles:


1. Start Thinthe creation of new LUNs requires minimal overhead. In Test 2 we saw
all arrays perform efficiently at this, except for the EMC VNX and IBM XIV
platforms. With low overhead on LUN creation, the efficient platforms can scale to
far greater numbers of LUNs and so can deliver storage resources more efficiently.
2. Get Thinthe ability to move data from thick to thin deployments. The import of
existing data into an array requires features that enable data to be optimized as it is
written to disk. Only HP 3PAR StoreServ is able to perform inline zero detection at
write time. EMC VNX, Dell Compellent and IBM XIV are able to zero detect when
data is imported under certain circumstances, for example as part of replication, but
these are not completely flexible solutions.
3. Stay Thinthe ability to detect and free unused space over time. As data is written
to thin volumes, the trend is for LUNs to grow in size to equal the logically allocated
capacity. This can happen because of defragmentation or with "thin unfriendly" file
systems that embed metadata with content, or are inefficient at reusing released
resources. Most vendors, with the exception of NetApp now support some form of
zero-page-reclaim or UNMAP feature, where space is returned to the array when
released by the host. However these background tasks can have a significant impact
on host I/O performance as was demonstrated in Test 1.

Only the HP 3PAR StoreServ platform provides a thin provisioning implementation


that delivers the most efficient storage utilization.

Best Practices

The testing and research in this white paper highlights a number of best practice
considerations:
1. Implement Zero-Page-ReclaimThis feature should be used to ensure LUNs stay
thin, however on most platforms (except 3PAR because of its custom ASIC and XIV

Edison: HP Thin Technologies Comparison Page 22


because it runs so slowly) needs to be scheduled out of normal production hours to
minimize performance impact.
2. Be aware of minimum LUN sizesWhen setting a standard for thin provisioned
LUNs, ensure that the minimum configured LUN size is not likely to waste capacity.

Edison: HP Thin Technologies Comparison Page 23


Appendix ADocument References

The following documents were referenced during the production of this white paper.

TR-3505 - NetApp Deduplication for FAS and V-Series Deployment and


Implementation Guide
TR-3563 NetApp Thin Provisioning Increases Storage Utilization with On Demand
Allocation
TR-3483 Thin Provisioning in a NetApp SAN or IP SAN Enterprise Environment
GC27-3913-03 - IBM XIV Storage System Planning Guide
GC27-3912-02 IBM XIV Storage System Product Overview
4AA3-3516ENW HP 3PAR Architecture
300-006-718 Best Practices for Fast, Simple Capacity Application with EMC
Symmetrix
H2222.3 EMC VNX Virtual Provisioning White Paper
300-011-798 EMC VNX Series Release 7.0 VNX System Operations
Dell Compellent Data Progression Data Sheet

Edison: HP Thin Technologies Comparison Page 24


Appendix BTest Equipment Specification

The following equipment was used to perform the testing documented in this white
paper.

Arrays

Hitachi VSP
NetApp FAS3140, running Data ONTAP 8.0.2, RAID-DP across 28 drives.
Dell CompellentRAID-5 across 72x 600GB SAS drives
EMC VNX5700 running microcode 5.31, RAID-5 across 24x 300GB 15K SAS drives.
HP 3PAR F400 InForm OS 3.1.1 (MU1)
EMC VMAX-20K
IBM XIV Gen2, 72x 1TB SATA drives

Servers

HP BL Blade Servers, 2x Intel X5650 CPU, 16GB RAM, HP Flex10 I/O


Windows 2008R2 SP1 & CentOS 6.2
IOMeter v2006.07.27

4AA4-4079ENW

Edison: HP Thin Technologies Comparison Page 25

Вам также может понравиться