2. Introduction
⢠Every day,
approximately 15
petabytes of new
information is
generated worldwide
⢠The total amount of
digital data doubles
approximately every
two years
2
3. History
⢠Early computers used a very basic
persistent storage system, based on
punched cards or paper tape
⢠Drum memory was one of the first
magnetic read/write storage systems
â It was widely used in the 1950s and
into the 1960s
â Consisted of a large rotating metal
cylinder that was coated on the
outside with magnetic recording
material
â Multiple rows of fixed read-write
heads were placed along the drum,
each head reading or writing to one
track
â The drum could store 62 kB of data
3
4. History â Hard disks
⢠The first commercial digital disk
storage device was part of the IBM
RAMAC 350 system, shipped in
1956
â Approximately 5 MB of data
â Fifty 61 cm diameter disks
â Weighed over a ton
⢠Over the years:
â Physical size of hard disks shrunk
â Magnetic density increased
â Rotation speed increased from 3,600
rpm to 15,000 rpm
â Seek times lowered as a result of
using servo controlled read/write
heads instead of stepper motors
4
5. History â Tapes
⢠The IBM 726, introduced in 1952, was
one of the first magnetic tape systems
â 2 MB per 20-centimeter-diameter reel
of tape
⢠Reel tapes were used until the late
1980s, mostly in mainframes
⢠In 1984, DEC introduced the Digital
Linear Tape (DLT)
â Super DLT (SDLT) tape cartridges can
store up to 300 GB of data
⢠Linear Tape Open (LTO) was originally
developed in the late 1990s
â LTO version 7 was released in 2015 and
can hold up to 6 TB of data
5
7. Storage model
⢠Most servers use
external storage,
sometimes
combined with
internal storage
⢠A model of
storage building
blocks is shown
on the right
7
8. Disks â command sets
⢠Disks are connected to disk controllers using a
command set, based on either ATA or SCSI
â Advanced Technology Attachment (ATA), also known as
IDE, uses a relatively simple hardware and
communication protocol to connect disks to computers
(mostly PCs)
â Small Computer System Interface (SCSI) is a set of
standards for physically connecting and transferring
data between computers (mostly servers) and
peripheral devices, like disks and tapes
⢠The SCSI command set is complex - there are about 60
different SCSI commands in total
⢠Serial interfaces replaced the parallel interfaces,
but the disk commands are still the same 8
9. Mechanical hard disks
⢠Mechanical disks consist of:
â A vacuum sealed case
â One or more spinning
magnetic disks on one
spindle
â A number of read/write
heads that can move to
reach each part of the
spinning disks
9
10. Mechanical hard disks
⢠Serial ATA (SATA) disks
â Low-end high-capacity disks
â Ideal for bulk storage applications (like archiving or backup)
â Have a low cost per gigabyte
â Often used in PCs and laptops
â Use the SMART command set to control the disk
10
11. Mechanical hard disks
⢠Serial Attached SCSI (SAS) disks
â Relatively expensive
â High end disks
â Spinning disk platters with a rotational speed of 10,000 or 15,000
rpm
â Typically have 25% of the capacity of SATA or NL-SAS disks
â Uses the SCSI command set that includes error-recovery and error-
reporting and more functionality than the SMART commands used by
SATA disks
11
12. Mechanical hard disks
⢠Near-Line SAS (NL-SAS) disks
â Have a SAS interface, but the mechanics of SATA disks
â Can be combined with faster SAS disks in one storage array
12
13. Solid State Drives (SSDs)
⢠SSD disks donât have moving parts
⢠Based on flash technology
â Flash technology is semiconductor-
based memory that preserves its
information when powered off
⢠Connected using a standard SAS
disk interface
⢠Data can be accessed much faster
than using mechanical disks
â Microseconds vs. milliseconds
⢠Most storage vendors now offer
all-flash arrays â storage systems
using only SSD disks
13
14. Solid State Drives (SSDs)
⢠SSDs consume less power, and therefore
generate less heat, than mechanical disks
⢠They have no moving parts
⢠They generate no vibrations that could
influence or harm other components, or
shorten their lifetime
⢠The main disadvantage of SSDs is their price
per gigabyte
â Considerably higher than mechanical disks
â Price per GB is dropping fast
14
15. Solid State Drives (SSDs)
⢠Flash memory can only be rewritten a limited
number of times
â SSD disks âwear outâ more rapidly than mechanical
disks
â SSDs keep track of the number of times a sector is
rewritten, and map much used sectors to spare
sectors if they are about to wear out
â It is important to monitor the wear level of heavily
used SSDs
⢠Replace them before they break
15
16. Disk capacity - Kryder's law
⢠The density of
information
on hard drives
doubles every
13 months
⢠An average
single disk
drive in 2025
will hold more
than 20,000
TB (20 PB) of
data
Please note that the vertical scale is logarithmic instead of linear 16
17. Disk capacity - Kryder's law
⢠The picture on the right shows
8 bytes core memory and 8 GB
SD flash memory
â An increase of 1,000,000,000
times in 50 years
⢠To have full benefits of
Kryder's law, the storage
infrastructure should be
designed to handle just in
time expansion
â Buy disks as late as possible!
17
18. Tapes
⢠When storing large amounts of data, tape is
the most inexpensive option
⢠Tapes are suitable for archiving
â Tape manufacturers guarantee a long life
expectancy
â DLT, SDLT, and LTO Ultrium cartridges are
guaranteed to be readable after 30 years on the
shelf
18
19. Tapes
⢠Disadvantages:
â Tapes are fragile
⢠Manual handling can lead to mechanical defects:
â Tapes dropping on the floor
â Bumping
â Bad insertions of tapes in tape drives
â Tape cartridges contain mechanical parts
⢠Manually changed tapes get damaged easily
â Frequent rewinding causes stress to the tape
substrate
⢠Leads to lower reliability of data reads
19
20. Tapes
â Tapes are extremely slow
⢠They only write and read data sequentially
⢠When a particular piece of data is required, it must be
searched by reading all data on tape until the required
data is found
⢠Together with rewinding of the tape (needed for
ejecting the tapes) handling tapes is expressed in
minutes instead of in milliseconds or microseconds
20
21. Tapes
⢠(S)DLT and LTO are the most popular tape
cartridge formats in use today
â LTO has a market share of more than 80%
â LTO-7 tape cartridges can store 6 TB of uncompressed
data
⢠Tape throughput is in the 100 to 150 MB/s range
â The tape drive interface is capable of even higher
speeds
â Most tape drives use 4 Gbit/s Fibre Channel interfaces
⢠A sustained throughput of between 350 and 400 MB/s
21
22. Tape library
⢠Tape libraries can be used to
automate tape handling
⢠A tape library is a storage
device that contains:
â One or more tape drives
â A number of slots to hold tape
cartridges
â A barcode or RFID tag reader to
identify tape cartridges
â An automated method for
loading tapes
22
23. Virtual tape library
⢠A Virtual Tape Library (VTL) uses disks for storing
backups
⢠A VTL consists of:
â An appliance or server
â Software that emulates traditional tape devices and
formats
⢠VTLs combine high performance disk based backup and
restore with well-known backup applications,
standards, processes, and policies
⢠Most of the current VTL solutions use NL-SAS or SATA
disk arrays because of their relatively low cost
⢠They provide multiple virtual tape drives for handling
multiple tapes in parallel
23
24. Controllers
⢠Controllers connect disks and/or tapes to a server, in one of
two ways:
â Implemented as a PCI expansion boards in the server
â Part of a NAS or SAN deployment, where they connect all available
disks and tapes to redundant Fibre Channel, iSCSI, or FCoE
connections
24
25. Controllers
⢠A controller can implement:
â High performance
â High availability
â Virtualized storage
â Cloning
â Data deduplication
â Thin provisioning
25
26. Controllers
⢠The controller splits
up all disks in small
pieces called
physical extents
⢠From these physical
extents, new virtual
disks (Logical Unit
Numbers - LUNs)
are composed and
presented to the
operating system
26
27. RAID
⢠Redundant Array of Independent Disks (RAID)
solutions provide:
â High availability of data
â Improvements of performance
⢠RAID uses multiple redundant disks
⢠RAID can be implemented:
â In the disk controllerâs hardware
â As software running in a serverâs operating system
27
28. RAID
⢠RAID can be implemented in several
configurations, called RAID levels
⢠In practice, five RAID levels are implemented
most often:
â RAID 0 - Striping
â RAID 1 - Mirroring
â RAID 10 - Striping and Mirroring
â RAID 5 - Striping with distributed parity
â RAID 6 - Striping with distributed double parity
28
29. RAID 0 - Striping
⢠RAID 0 is also known as striping
⢠Provides an easy and cheap way
to increase performance
⢠Uses multiple disks, each with a
part of the data on it
⢠RAID 0 actually lowers availability
â If one of the disks in a RAID 0 set
fails, all data is lost
⢠Only acceptable if losing all data
on the RAID set is no problem
(for instance for temporary data) 29
30. RAID 1 - Mirroring
⢠RAID 1 is also known as
mirroring
⢠A high availability solution that
uses two disks that contain the
same data
⢠If one disk fails, data is not lost
as it is still available on the
mirror disk
⢠The most reliable RAID level
⢠High price
â 50% of the disks are used for
redundancy only
⢠A spare physical disk can be
configured to automatically take
over the task of a failed disk
30
31. RAID 10 - Striping and mirroring
⢠RAID 10 uses a combination of striping and mirroring
⢠Provides high performance and availability
⢠Only 50% of the available disk space is used
31
32. RAID 5 - Striping with distributed parity
⢠Data is written in disk blocks on all disks
⢠A parity block of the written disk blocks is stored as well
⢠This parity block is used to automatically reconstruct data in a RAID 5
set (using a spare disk) in case of a disk failure
32
33. RAID 6 - Striping with distributed double parity
⢠RAID 6 protects against double disk failures by using two
distributed parity blocks instead of one
⢠Important in case a second disk fails during reconstruction of
the first failing disk
33
34. Data deduplication
⢠Data deduplication searches the storage system
for duplicate data segments (disk blocks or
files) and removes these duplicates
⢠Data deduplication is used in archived as well
as in production data
34
35. Data deduplication
⢠The deduplication system keeps a table of hash
tags to quickly identify duplicate disk blocks
â The incoming data stream is segmented
â Hash tags are calculated of those segments
â The hashes are compared to hash tags of segments
already on disk
â If an incoming data segment is identified as a
duplicate, the segment is not stored again, but a
pointer to the matching segment is created for it
instead
35
37. Data deduplication
⢠Deduplication can be done inline or
periodically
â Inline deduplication checks for duplicate data
segments before data is written to disk
⢠Avoids duplicate data on disks at any time
⢠Introduces a relatively large performance penalty
37
38. Data deduplication
â Periodically: writing data to disk first, and
periodically check if duplicate data exists
⢠Duplicate data is deduplicated by changing the
duplicate data to a pointer to existing data on disk, and
freeing disk space of the original block
⢠This process can be done at times when performance
needs are low
⢠Duplicate data will be stored on the disks for some time
38
39. Cloning and snapshots
⢠With cloning and snapshotting, a copy of data is
made at a specific point in time that can be used
independently from the source data
⢠Usage:
â Create a backup at a specific point in time, when the
data is in a stable, consistent state
â Creating test sets of data and an easy way to revert to
older data without restoring data from a backup
⢠Cloning: the storage system creates a full copy of
a disk, much like a RAID 1 mirror disk
39
40. Cloning and snapshots
⢠Snapshot: represents a point in time of the
data on the disks
â No writing to those disks is permitted anymore, as
long as the snapshot is active
â All writing is done on a separate disk volume in
the storage system
â The original disks still provide read-access
40
41. Thin provisioning
⢠Thin provisioning enables the allocation of more
storage capacity to users than is physically installed
â About 50% of allocated storage is never used
⢠Thin provisioning still provides the applications with
the required storage
â Storage is not really available on physical disks
â Uses automated capacity management
â The application's real storage need is monitored closely
â Physical disk space is added when needed
⢠Typical use: Providing users with large sized home
directories or email storage
41
42. Direct Attached Storage (DAS)
⢠DAS â also known as local disks â is a storage
system where one or more dedicated disks
connect via the SAS or SATA protocol to a built-in
controller, connected to the rest of the computer
using the PCI bus
⢠The controller provides a set of disk blocks to the
computer, organized in LUNs (or partitions)
⢠The computerâs operating system uses these disk
blocks to create a file system to store files
42
43. Storage Area Network (SAN)
⢠A Storage Area Network (SAN) is a specialized storage
network that consists of SAN switches, controllers and
storage devices
⢠It connects a large pool of central storage to multiple
servers
⢠A SAN physically connects servers to disk controllers
using specialized networking technologies like Fibre
Channel or iSCSI
⢠Via the SAN, disk controllers offer virtual disks to
servers, also known as LUNs (Logical Unit Numbers)
⢠LUNs are only available to the server that has that
specific LUN mounted
43
44. Storage Area Network (SAN)
⢠The core of the SAN is
a set of SAN switches,
called the Fabric
â Comparable with a
LANâs switched network
segment
⢠Host bus adapters
(HBAs) are interface
cards implemented in
servers
â Comparable to NICs
used in networking
â Connected to SAN
switches, usually in a
redundant way
44
45. Storage Area Network (SAN)
⢠In SANs, a large number of
disks are installed in one
or more disk arrays
⢠The number of disks
varies between dozens of
disks and hundreds of
disks
⢠A disk array can easily
contain many hundreds of
terabytes (TB) of data or
more
45
47. Fibre Channel
⢠Fibre Channel (FC) is a dedicated level 2 network
protocol, specially designed for transportation of
storage data blocks
⢠Speeds: 2 Gbit/s, 4 Gbit/s, 8 Gbit/s, or 16 Gbit/s
⢠Runs on:
â Twisted pair copper wire (i.e. UTP and STP)
â Fiber optic cables
⢠The Fibre Channel protocol was specially
developed for the transport of disk blocks
⢠The protocol is very reliable, with guaranteed
zero data loss
47
48. Fibre Channel
⢠Three network topologies:
â Point-to-Point
⢠Two devices are connected directly to each other
â Arbitrated loop
⢠Also known as FC-AL
⢠All devices are in a loop
â Switched fabric
⢠All devices are connected to Fibre Channel switches
⢠A similar concept as in Ethernet implementations
⢠Most implementations today use a switched fabric
48
49. FCoE
⢠Fibre Channel over Ethernet (FCoE) encapsulates
Fibre Channel data in Ethernet packets
⢠Allows Fibre Channel traffic to be transported
over 10 Gbit or higher Ethernet networks
⢠FCoE eliminates the need for separate Ethernet
and Fibre Channel cabling and switching
technology
⢠PCoE needs at least 10 Gbit Ethernet with special
extensions, known as Data Center Bridging (DCB)
or Converged Enhanced Ethernet (CEE)
49
50. FCoE
⢠Ethernet extensions:
â Lossless Ethernet connections
⢠A FCoE implementation must guarantee that no
Ethernet packets are lost
â Quality of Service (QoS)
⢠Allows FCoE packets to have priority over other
Ethernet packets to avoid storage performance issues
â Large Maximum Transfer Unit (MTU) support
⢠Allows Ethernet packets of 2500 bytes in size, instead of
the standard 1500 bytes
⢠Also known as Jumbo frames
50
51. FCoE
⢠FCoE needs specialized
Converged Network
Adapters (CNAs)
⢠CNAs support the Ethernet
extensions
⢠They present themselves to
the operating system as two
adapters:
â Ethernet Network Interface
Controller (NIC)
â Fibre Channel Host Bus
Adapter (HBA)
51
52. iSCSI
⢠iSCSI allows the SCSI protocol to run over
Ethernet LANs using TCP/IP
⢠Uses the familiar TCP/IP protocols and well
known SCSI commands
⢠Performance is typically lower than that of
Fibre Channel, due to the TCP/IP overhead
⢠With 10 or 40 Gbit/s Ethernet and jumbo
frames, iSCSI is now rapidly conquering a big
part of the SAN market
52