Deduplication and single instance storage

Deduplication and Single
Instance Storage

Practical Applications for Backups,
Archiving, and Primary Storage

Presented by:

Jacob Farmer
Cambridge Computer

© Copyright 2009-2010, Cambridge Computer Services, Inc. – All Rights Reserved
www.CambridgeComputer.com – 781-250-3000

About Your Lecturer

Jacob Farmer, CTO, Cambridge Computer
• Cambridge Computer, founded in 1991, provides training, integration,
sales, and consulting in the fields of storage management, data
protection, and digital archiving.
Been working in data protection and storage management for
almost 20 years.
• Lecturer on storage technologies for Usenix for the past 10 years.
Hybrid of industry analyst and consultant to end-users.
• Spend 25% of my time working in the industry, going to conferences,
meeting with vendors.
• 75% of my time customer-facing, helping the sales and services
departments design solutions for end users.
Email: jfarmer@CambridgeComputer.com

Deduplication and Single Instance Storage – Interop – Las Vegas – April 27, 2010
© Copyright 2009-2010, Cambridge Computer Services, Inc. All rights reserved.
www.CambridgeComputer.com 2

Follow Me on Twitter

My personal activities:
•@JacobAFarmer
–Note the “A” – my middle initial
My educational activities
•@Cambridge_EDU

Usenix-On-The-Road: The Latest Trends in Storage Networking
© Copyright 2009-2010-2010, Cambridge Computer Services, Inc. All rights reserved.

Agenda / Topics

Dedupe basics
• What is it, how does it work, and what is all the fuss about?
• Hashing, segmenting, indexing, etc.
Dedupe for backup systems
• Basic benefits
• Different approaches for scaling backups and how they relate back to dedupe
– Front end bottlenecks
– Backup data-movers
– Back-end bottlenecks and scalable deduping
Dedupe for primary storage
• Virtual servers, physical servers, VDI
• Rich media dedupe
WAN Accelerators
Questions as time permits


What is Deduplication?

A term that refers to a number of different methods
and techniques for reducing multiple instances of
identical data down to a single (or at least fewer)
instances.
• Common data is replaced with pointers or tokens that refer
back to the actual data.
Other terms for deduplication
• Data Reduction
• Commonality Factoring
• Capacity Optimization
• Single Instancing or Single-Instance Storage (SIS)


Is Deduplication a Form of
Compression?
Yes, and No.

YES – Deduplication results in data taking up less
storage space or consuming less bandwidth on a
network circuit.
• Note that dedupe is often used in conjunction with
conventional compression.
NO – Deduplication could work on data types that
are not compressible.
• If you have 10 identical JPEG files stored in an
uncompressible format, they could be reduced to a single
instance, thus freeing up 90% of your capacity.


Where Do You Find Dedupe
Solutions?
Deduplication solutions come to market whenever costs or
efficiencies can be achieved by eliminating redundancy.
• Backups
– Conventional backup systems generate tons of redundant data
• Email systems (at rest and in flight)
– I send an email with the same attachment to everyone in the company.
– Then everyone stores it in his/her personal home directory
– Everyone in the branch offices pulls it over the WAN
• File traffic over a WAN
• Application and O.S. binaries across multiple systems
– Virtual Servers and Virtual Desktops
– Backups over a WAN
• Very large collections of rich media files


Hashing / Fingerprinting

Hashing (aka fingerprints, digests, signatures)
• Generates a unique number (160+ bits) based on content
• Hash acts as a proxy for content
• Given a hash, not computationally feasible to generate
content
Common Hashing Algorithms
• MD5
• SHA-1
• SHA-256
• AES
Hash Size and the Birthday Paradox
• The size of the hash needs to be suitable to the task at
hand


Hash Collisions - Are they real?

Fibre Channel
Bit Error Rate
10-10 10-20 10-30
Probability

Hit by
lightning Simultaneous
triple disk fault Cryptographic
on RAID-6 hash collision

Win the Cretaceous extinction meteor
lottery hitting in the next second


What Makes Deduplication and SIS
Technology Difficult to Engineer?
Hash Processing
• Modern CPUs make this much easier
– 100+ MB/sec/core
• Hardware co-processor cards can hash at rates north of
1.5GB/Sec.
Disk performance
• Deduped data often ends up getting fragmented on disk
– This can hurt performance especially for backup systems

Alignment of de-dupe segments
Indexing


Indexing Can Be Hard
# lookups/sec

106
Router Fine grained
105 Purpose built hardware content
tracking
104

103 large
Software Database database
102
Technology
101
iPod NYC
100 Human Lookup Rates phonebook

101 102 103 104 105 106 107 108 109
# records

Parsing / Segmenting / Chunking

Data needs to be “chopped up” in a consistent way in order to get optimal
dedupe ratios
Without any kind of special segmenting strategies backup streams and
complex file types do not dedupe effectively

Large files are almost always changed with overstrike semantics
• Databases, structured data, .vmdk, .pst files
Small files are almost changed with insert semantics
• Office apps, editors etc
If there are large files (e.g. database tables, virtual machine images) in the
backup mix, their treatment usually will dominate any data reduction
strategy.
• Don’t sweat the small stuff!
Different vendors may have strengths with one type or another


Change Types: Insert v. Overstrike

Insert:
The quick brown fox jumped over the lazy dog.
The quick brown horse jumped over the lazy dog.
Identical data (may be)
misaligned
Overstrike:
“Fred” added to
Joe Sue employee database
Joe Fred Sue Identical data doesn’t
move


NetBackup OST (open storage option)

API and framework that makes it very easy for a dedupe
target device vendor to parse the data stream.
• Pre segments content
• Enables more efficient dedup solutions
• Allows for smart copy between systems of only changed
data

PQZ R PQR
Z


Deduplication for Backups


Backup Systems Have a Lot of
Redundant Data
Conventional backup solutions generate a ton of redundant
data
• Assuming weekly full backups, a file that has not changed in 5 years,
still gets backed up 260 times!
• Assuming daily full backups of email, a message you received 5 years
ago gets backed up 1825 times.
– Similarly, a record in a database from 5 years ago might be backed up
1825 times!
There are really two problems to solve:
• Minimizing the amount of redundant data that gets repeatedly
transferred
• Minimizing the amount of redundant data that gets stored.


Most of the Buzz on Dedupe is from
Backup Target Vendors
A “target” is a backup storage device
Dedupe disk targets generally come packaged as
• NAS
– File server (NFS or CIFS) interface
• Virtual tape library
– A disk device that emulates a tape library
– Fibre Channel or iSCSI interface
– NAS vs VTL outside the scope of this talk


When Do Dedupe Disk Targets Shine?

When you are backing up a lot of redundant data
• Files that never or seldom change between backups
• Duplicated files
• Databases and email repositories that are receptive to
commonality factoring
When you are retaining backup data for a decent
amount of time
• Ideally you are keeping several weeks of backups
When you seek to replicate a conventional backup
system over a WAN.


Example: NYC Law Firm with
NetBackup
6+ Terabytes
Full backups every day!
• Why? Because someone had a bad experience in the past
with incremental backups and has trust issues
90 day retention period
Most files seldom change
• Many files are scanned images that never change
Several large databases
Several TB of MS Exchange
Average result – 102x capacity optimization !

Try to Visualize 100x Capacity
Optimization

OR

One 3U cabinet v. 7 full racks full of gear!


Backup Vaulting – Another Use Case
for Dedupe


Why Replicate the Backup System?

Relatively easy DR solution
• Does not require additional software for the hosts
• Does not require storage devices with replication
capabilities
• One system that replicate all of your hosts
– Platform-independent

Eliminate the need to ship tapes off site
• Eliminate the need to encrypt tapes


Example: Defense Contractor
Replicating ERP System
Problem: CIO does not want employee data being
sent off site without encrypting the tapes.
• IT staff wants to avoid tape encryption.

Solution:
• Full backup of 800GB+ Oracle database to deduplicating disk target
every day.
• Retain backups for 60 days on disk.
– 60 x 800 = 48TB
• Vault backups to remote site over T1

Outcome
• Dedupe ration of about 70:1
• 800GB backup job traverses the T1 in a few hours


But, Before We Get all Hot and
Bothered . . .

Let’s review how backup systems
actually work!


Common Backup Bottlenecks

Backup Clients
You have to get data off the
host and transfer it
Network
Network Seldom the real bottleneck,
except over a WAN

Backup Servers
I/O processing is the most
common bottleneck

Storage Devices
Storage devices can be a
bottleneck, but are seldom
the whole problem.


Front-end and Network – Minimize
Duplication in the First Place
Backups generate a lot of redundant data, so what if
we had smarter client software that did not generate
redundant data?
• Incremental Forever
– After the first full backup, only do incremental backups
– This is what IBM TSM does, for instance
• Synthetic Full Backup
– Last weeks full backup is merged with this weeks incremental
backups to “synthesize” this week’s full backup.
– No need to transfer redundant data


Example: Energy Firm using IBM
Tivoli Storage Manager
TSM only backs up files that have changed.
• It does not generate a lot of duplicate files
Most of the 15TB of capacity are documentation and images
that do not change – ever.
• Relatively little of it is database.
• Images don’t compress
• Utilizing compression on TSM client for compressible files
Over all deduplication ratio: about 2:1
• Can’t justify the cost of dedupe across the board
• Resolution: Set up dedupe tier for database and email
– Do the file backups to conventional disk and tape


Synthetic Full Backups – An Approach
that Creates a Need for Dedupe
Synthetic Full Backups
• “Poor man’s incremental forever”
• Combine subsequent incremental backups with the
previous full backup to “synthesize” the next week’s full
backup.
• Great technique for minimizing networking traffic from
backups.
Synthetic full backups require that at least two weeks
of backups be available on disk.
• Dedupe disk targets tend to be a big win for synthetic full
backups


Example: Research Firm with 6 Week
Retention and Synthetic Fulls
60TB+
• Mix of large file systems, content management systems,
email, and database
Using Commvault with heavy use of synthetic full
backups
6 week retention on disk
Dedupe ratios between 8x and 16x
• NOTE: Their backup data could not fit in one dedupe box,
so they are managing 4 separate dedupe appliances in
each of their locations.


Theoretical v. Actual Capacity
Your Mileage May Vary
YMMV – one customer’s mileage
• 48 TB raw disk
• 36 TB with RAID-6
• 35 +/ TB for unique capacity
• 3-5 TB deliberately left empty for headroom
Might hold
• As much as 500 TB of backups
• Or as little as 50 TB.


Dedupe on the Backup Client

Host-side dedupe is a form of sub-file-level incremental
• Instead of catching block-level changes, the file system changes are
hashed and compared with the back-end storage repository.
• Alternative to block-based CDP
• Unique data segments are then transferred to the backup service.
Host-side deduping is very valuable over the WAN.
• Minimizes data that needs to be transferred
• Typically it will dedupe across hosts, reducing files that are common to
multiple hosts
– Such as application and operating system binaries


WAN Backup Software with Dedupe in
the Client
Dedupe Client

London
LAN

New York
Shared Client & Local Recovery
Local USB WAN

Hong Kong Backup Server(s) Jersey City

Backup System Network I/O
Processing Bottlenecks

Moving Backup Data Through the
Network


Backup Server I/O Processing is a
Major Bottleneck
In most enterprise backup systems a single backup
server would be a major performance bottleneck
• Unless you were doing incremental forever or sub-file-level
backups
• Add a dedup process to that and it becomes that much
harder
A common practice for scaling out backup server
performance is to add network “data movers”
• Also known as: storage nodes, media servers, media
agents, etc.


Interesting Idea – Add Deduplication
to the Network Data Movers

Network

Dedicated Storage Network


I/O Processing Bottlenecks
Network Data Movers and “LAN-Free”

Network data
movers “LAN-Free”
Clients

Dedicated Storage Network


LAN-Free Backup Clients and NDMP
Backups
In larger enterprise-class backup systems it is
common to have larger servers move data directly to
storage devices over Fibre Channel.
The fastest way to backup large NAS server is to do
NDMP dumps over Fibre Channel.


“LAN-FREE” – End-run Around the
Backup Server

SAN Clients work
G ig E like slave servers.
They back up
directly to the
storage media, while
reporting metadata
over the LAN to the
Storage Area Network backup server.

Tape Presumably all of these
Robot tape drives are part of a
Arm tape library.


Dedupe with LAN-Free Backup Clients

With LAN-Free backup you get no benefit from
dedupe processing residing on the data movers.
• The dedupe logic needs to sit on the target storage device
This is where VTLs shine
• VTL works just like tape
– Network data movers work fine
– LAN-Free clients work fine
• VTLs offer higher throughput than CIFS or NFS
– Common to see total throughput in excess of 1GB/Sec
• VTLs might offer tighter integration with tape
Many VTLs do dedupe as a post-process


Back End Bottlenecks

Can your dedupe appliance keep
pace with the backup system?


Back-End Bottlenecks: Can the
Dedupe Storage Devices Hack It?
If you open up the flood gates, you might find that a
single dedupe box on the LAN cannot hack it.
Some solutions:
• Buy lots of individual dedupe devices
• Maybe use a VTL implementation of dedupe
– Sorry out of the scope of this lecture
• Post-process deduping instead of deduping on-the-fly
– Less efficient from a capacity standpoint, but should be able to
achieve considerably better performance
• New grid-based architectures that offer parallel processing
for deduplication
• Newer dedupe devices that are up to the task

Stand-alone Deduplication Servers

Single server dedupe solutions are often constrained
by:
• RAM and processing power
• The size of the index they can manage
• Disk performance
When you max out the box, you need to buy another
one
• Very painful incremental upgrade
• No dedupe across multiple boxes
• Make sure that you but a big enough box!


Object-Based File System with Grid
Architecture and Global Dedupe
CIFS/NFS Clients
Backup System Data Movers
Conventional File System Consumers

Front-End Nodes
Export File Systems
Scale-out performance into GBs/Sec

Back-End Nodes
Manage disk, dedupe, and redundancy
Scale-deep to Petabytes of capacity


VTL with Scalable Deduplication

G ig E

Storage Network

De-Duplication Processors Single Instance
VTL Repository


Summary: Alternative Technologies
to Dedupe Disk Targets
Don’t Duplicate in the First Place
• Incremental Forever Backups
• WAN-enabled backups, perhaps with dedupe on the client

Throw disk at it
• Bulk SATA arrays cost typically less than $1K per TB
– Capacities up to 2PB
– Densities on the order of 1PB / rack
– MAID – power management to spin down inactive drives

Replicate Your SAN or NAS
• Use optimized file backup or archive solution to provide file recovery
and to meet retention requirements


Other Examples of Dedupe
Technology

Primary Storage, VDI, Rich Media
Archiving


Block-Level Dedupe for Primary
Storage
Most dedupe solutions are designed specifically for backup
and archival data.
A limited number of products can dedupe on live data.
• One day perhaps, dedupe for primary storage will be a way of life
Great applications – those with redundant data!
• Desktop virtualization (VDI)
– A number of very interesting solutions are coming to market
• VMDK backup, dedupe, and fail-over on one platform
• Boot image servers
Reclamation of empty disk space
• Blank space deduplicates very nicely!


Single-Instance Storage for Virtual
Desktops
Storage is a big deal-breaker for many VDI use cases
• Replaces desktop storage and desktop personnel with SAN storage and
highly specialized storage managers
New techniques for VDI storage break the desktop down into
elements and find commonality across all desktops
Virtual desktop file systems are “stitched together” from common
elements:
• Operating system
• Applications or sets of applications
• Variable elements
– For example: anti-virus signatures
• Personal elements
– Screen savers and background images
– Google toolbar
– Personal applications
– Personal files


Dedupe Across Large Collections of
Rich Media Files
Many types of files have content-level commonality across a
large collection of files.
• TIFF
• JPG
• PNG
• OpenEXR
• DICOM
• MS Office Documents
• PDFs
A high level of commonality can be detected and de-
duplicated, assuming a large enough sample set of data.
• Capacity optimization (depending on file type) on the order of 2x to 10x and
beyond.


Dedupe in WAN Accelerators


MS Exchange Branch Office: Example
of the Need for Dedupe over the WAN

Chicago New York

Atlanta
WAN

MS Exchange Server
Message with attachment sent to all staff. San Fran
Single instance message storage, but
the same message crosses the WAN multiple times


WAN Accelerator with Inline Dedupe

Site B

Site A

WAN Accelerators / WAFS Gateways
F ile S er ver s o r
NAS Ap p lian ce


Questions – If Time Permits


Deduplication and single instance storage

More Related Content

What's hot (18)

Viewers also liked (7)

Similar to Deduplication and single instance storage (20)

More from Interop (20)

Recently uploaded (20)

Deduplication and single instance storage